Assessing modeling and simulation tools and methods for predicting metabolic-based DDI Update from the IQ Work Group Presented by Scott Obach on behalf of the Induction Working Group Heidi Einolf, Odette Fahmi, Chris Gibson, Mohamed Shebley, Jose Silva, Mike Sinz, Jash Unadkat, Lei Zhang, Liangfu Chen, Ping Zhao October 6, 2011 North Jersey Drug Metabolism Discussion Group
42
Embed
Assessing modeling and simulation tools and methods for ... · 6/10/2011 · Assessing modeling and simulation tools and ... North Jersey Drug Metabolism Discussion Group. Outline
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Assessing modeling and simulation tools and methods for predicting metabolic-based DDI Update from the IQ Work Group
Presented by Scott Obach on behalf of the Induction Working Group Heidi Einolf Odette Fahmi Chris Gibson Mohamed Shebley Jose Silva Mike Sinz Jash Unadkat Lei Zhang Liangfu Chen Ping Zhao October 6 2011 North Jersey Drug Metabolism Discussion Group
Outline
IQ Consortium DDI Work Groups
Sub-objectives of the Induction Working Group ndash General overview
Modelsapproaches for the prediction of clinical CYP3A induction
A look at the new FDA draft decision tree for induction (R3 value)
SummaryConclusions
2
IQ Consortium Drug Interaction Work Groups General overview
bull 2009 Meeting between individuals from PhRMA Drug Metabolism Technical Group and FDA Office of Clinical Pharmacology
bull Agreement that the area of modelling and simulation of DDI from in vitro data is a valuable endeavor
bull Established Industry-FDA-Academia work groups
bull Main Objective To test various models and approaches for predicting DDI using agreed upon input parameters
bull Two teams formed
bull Inhibition+Inactivation ndash Discussion for another time
bull Induction ndash Wersquoll share findings now
3
To identify method(s) that can reliably use in vitro data to provide
quantitative forecasts of clinical DDI across a broad range of
drugs and provide a recommendation as to what approach to use
to inform the need for clinical DDI studies
Scope Development not early research
Sub-objectives ndash Induction Working Group General overview
bull Define acceptable approaches for the prediction of the clinical induction of CYP3A (models used input parameters needed what type of data is needed)
bull Standardize input parameters that can be standardized (eg fmCYP kdeg)
bull Define criteria for quality of input parameters
bull Define the extent to which we need to qualify the models
bull Work together to qualify the models identified on the ability to predict clinical DDI
These sub-objectives are in support of the main objective
4
01 1 10 100 10000
5
10
15
20
25
[Ind] M
Effe
ct(e
g F
old-
chan
ge m
RN
A)
Predictions of Induction DDI ndash General Principle
5
Most cytochrome P450 (CYP) induction prediction models use a relationship which include the parameters bull EC50
bull Emax
bull [Inducer]
These prediction models are generally based upon the same principal that being the law of mass action for receptor binding (and activation since induction is an agonist property
[Ind]EC[Ind]EEffect
50
max
Emax = 20-fold
EC50 = 15 microM
Approaches for prediction of clinical CYP3A induction
What modelsapproaches are being used now general methods and examples of approach(es)
Empirical Approaches - of positive control Correlation Methods - Relative Induction Score - CmaxEC50
Mathematical Equations (Mechanistic Static Models) - Net effect model
Ripp et al 2006 DMD 341742 Fahmi et al 2008 DMD 361971
Chu et al 2009 DMD 371339 Persson et al 2006 Pharm Res 2356
Fahmi et al 2008 DMD 361698
FDA Draft Guidance ndash Drug Interaction Studies 2006 Bjornsson et al DMD 2003 31815
Almond et al 2009 Curr Drug Metab 10420
Qualification of the different modelsapproaches Identification of trials and available in vitro induction data
bull Initially identified trials to model from the literature (U Washington Drug Interaction Database) or from internal sources
- Literature searches were focused on DDI studies involving common CYP3A substrates (eg midazolam triazolam alprazolam simvastatin buspirone etc)
- Compiled a list of perpetrators that caused clinical induction or no inductioninhibition of these CYP3A substrates
bull Using the initial search four victim drugs were chosen to further investigate based upon the number of trials that were available and differences in fmCYP and FG (midazolam alprazolam nifedipine and simvastatin)
bull Identified the in vitro (internal) mRNA induction data (and activity) we had with respect to the inducersnon-inducers in the identified trials (EC50 and Emax data) and consolidated it for use in modeling (median values)
7
Trials chosen for the model comparison
8
perpetrator victim of trials
rifampin midazolam 10
rifampin alprazolam 1
rifampin nifedipine 1
carbamazepine midazolam 1
carbamazepine alprazolam 1
nafcillin nifedipine 1
phenobarbital nifedipine 1
pleconaril midazolam 1
rosiglitazone nifedipine 1
Merck MK-1 midazolam 2
Merck MK-2 midazolam 1
pioglitazone simvastatin 1
pioglitazone midazolam 1
troglitazone simvastatin 1
omeprazole nifedipine 1
phenytoin midazolam 1
ranitidine nifedipine 2
bull 28 trials (17 victimperpetrator pairs)
Empirical Approaches
9
eg of positive control
Intended to indicate an in vitro induction response
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Outline
IQ Consortium DDI Work Groups
Sub-objectives of the Induction Working Group ndash General overview
Modelsapproaches for the prediction of clinical CYP3A induction
A look at the new FDA draft decision tree for induction (R3 value)
SummaryConclusions
2
IQ Consortium Drug Interaction Work Groups General overview
bull 2009 Meeting between individuals from PhRMA Drug Metabolism Technical Group and FDA Office of Clinical Pharmacology
bull Agreement that the area of modelling and simulation of DDI from in vitro data is a valuable endeavor
bull Established Industry-FDA-Academia work groups
bull Main Objective To test various models and approaches for predicting DDI using agreed upon input parameters
bull Two teams formed
bull Inhibition+Inactivation ndash Discussion for another time
bull Induction ndash Wersquoll share findings now
3
To identify method(s) that can reliably use in vitro data to provide
quantitative forecasts of clinical DDI across a broad range of
drugs and provide a recommendation as to what approach to use
to inform the need for clinical DDI studies
Scope Development not early research
Sub-objectives ndash Induction Working Group General overview
bull Define acceptable approaches for the prediction of the clinical induction of CYP3A (models used input parameters needed what type of data is needed)
bull Standardize input parameters that can be standardized (eg fmCYP kdeg)
bull Define criteria for quality of input parameters
bull Define the extent to which we need to qualify the models
bull Work together to qualify the models identified on the ability to predict clinical DDI
These sub-objectives are in support of the main objective
4
01 1 10 100 10000
5
10
15
20
25
[Ind] M
Effe
ct(e
g F
old-
chan
ge m
RN
A)
Predictions of Induction DDI ndash General Principle
5
Most cytochrome P450 (CYP) induction prediction models use a relationship which include the parameters bull EC50
bull Emax
bull [Inducer]
These prediction models are generally based upon the same principal that being the law of mass action for receptor binding (and activation since induction is an agonist property
[Ind]EC[Ind]EEffect
50
max
Emax = 20-fold
EC50 = 15 microM
Approaches for prediction of clinical CYP3A induction
What modelsapproaches are being used now general methods and examples of approach(es)
Empirical Approaches - of positive control Correlation Methods - Relative Induction Score - CmaxEC50
Mathematical Equations (Mechanistic Static Models) - Net effect model
Ripp et al 2006 DMD 341742 Fahmi et al 2008 DMD 361971
Chu et al 2009 DMD 371339 Persson et al 2006 Pharm Res 2356
Fahmi et al 2008 DMD 361698
FDA Draft Guidance ndash Drug Interaction Studies 2006 Bjornsson et al DMD 2003 31815
Almond et al 2009 Curr Drug Metab 10420
Qualification of the different modelsapproaches Identification of trials and available in vitro induction data
bull Initially identified trials to model from the literature (U Washington Drug Interaction Database) or from internal sources
- Literature searches were focused on DDI studies involving common CYP3A substrates (eg midazolam triazolam alprazolam simvastatin buspirone etc)
- Compiled a list of perpetrators that caused clinical induction or no inductioninhibition of these CYP3A substrates
bull Using the initial search four victim drugs were chosen to further investigate based upon the number of trials that were available and differences in fmCYP and FG (midazolam alprazolam nifedipine and simvastatin)
bull Identified the in vitro (internal) mRNA induction data (and activity) we had with respect to the inducersnon-inducers in the identified trials (EC50 and Emax data) and consolidated it for use in modeling (median values)
7
Trials chosen for the model comparison
8
perpetrator victim of trials
rifampin midazolam 10
rifampin alprazolam 1
rifampin nifedipine 1
carbamazepine midazolam 1
carbamazepine alprazolam 1
nafcillin nifedipine 1
phenobarbital nifedipine 1
pleconaril midazolam 1
rosiglitazone nifedipine 1
Merck MK-1 midazolam 2
Merck MK-2 midazolam 1
pioglitazone simvastatin 1
pioglitazone midazolam 1
troglitazone simvastatin 1
omeprazole nifedipine 1
phenytoin midazolam 1
ranitidine nifedipine 2
bull 28 trials (17 victimperpetrator pairs)
Empirical Approaches
9
eg of positive control
Intended to indicate an in vitro induction response
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
IQ Consortium Drug Interaction Work Groups General overview
bull 2009 Meeting between individuals from PhRMA Drug Metabolism Technical Group and FDA Office of Clinical Pharmacology
bull Agreement that the area of modelling and simulation of DDI from in vitro data is a valuable endeavor
bull Established Industry-FDA-Academia work groups
bull Main Objective To test various models and approaches for predicting DDI using agreed upon input parameters
bull Two teams formed
bull Inhibition+Inactivation ndash Discussion for another time
bull Induction ndash Wersquoll share findings now
3
To identify method(s) that can reliably use in vitro data to provide
quantitative forecasts of clinical DDI across a broad range of
drugs and provide a recommendation as to what approach to use
to inform the need for clinical DDI studies
Scope Development not early research
Sub-objectives ndash Induction Working Group General overview
bull Define acceptable approaches for the prediction of the clinical induction of CYP3A (models used input parameters needed what type of data is needed)
bull Standardize input parameters that can be standardized (eg fmCYP kdeg)
bull Define criteria for quality of input parameters
bull Define the extent to which we need to qualify the models
bull Work together to qualify the models identified on the ability to predict clinical DDI
These sub-objectives are in support of the main objective
4
01 1 10 100 10000
5
10
15
20
25
[Ind] M
Effe
ct(e
g F
old-
chan
ge m
RN
A)
Predictions of Induction DDI ndash General Principle
5
Most cytochrome P450 (CYP) induction prediction models use a relationship which include the parameters bull EC50
bull Emax
bull [Inducer]
These prediction models are generally based upon the same principal that being the law of mass action for receptor binding (and activation since induction is an agonist property
[Ind]EC[Ind]EEffect
50
max
Emax = 20-fold
EC50 = 15 microM
Approaches for prediction of clinical CYP3A induction
What modelsapproaches are being used now general methods and examples of approach(es)
Empirical Approaches - of positive control Correlation Methods - Relative Induction Score - CmaxEC50
Mathematical Equations (Mechanistic Static Models) - Net effect model
Ripp et al 2006 DMD 341742 Fahmi et al 2008 DMD 361971
Chu et al 2009 DMD 371339 Persson et al 2006 Pharm Res 2356
Fahmi et al 2008 DMD 361698
FDA Draft Guidance ndash Drug Interaction Studies 2006 Bjornsson et al DMD 2003 31815
Almond et al 2009 Curr Drug Metab 10420
Qualification of the different modelsapproaches Identification of trials and available in vitro induction data
bull Initially identified trials to model from the literature (U Washington Drug Interaction Database) or from internal sources
- Literature searches were focused on DDI studies involving common CYP3A substrates (eg midazolam triazolam alprazolam simvastatin buspirone etc)
- Compiled a list of perpetrators that caused clinical induction or no inductioninhibition of these CYP3A substrates
bull Using the initial search four victim drugs were chosen to further investigate based upon the number of trials that were available and differences in fmCYP and FG (midazolam alprazolam nifedipine and simvastatin)
bull Identified the in vitro (internal) mRNA induction data (and activity) we had with respect to the inducersnon-inducers in the identified trials (EC50 and Emax data) and consolidated it for use in modeling (median values)
7
Trials chosen for the model comparison
8
perpetrator victim of trials
rifampin midazolam 10
rifampin alprazolam 1
rifampin nifedipine 1
carbamazepine midazolam 1
carbamazepine alprazolam 1
nafcillin nifedipine 1
phenobarbital nifedipine 1
pleconaril midazolam 1
rosiglitazone nifedipine 1
Merck MK-1 midazolam 2
Merck MK-2 midazolam 1
pioglitazone simvastatin 1
pioglitazone midazolam 1
troglitazone simvastatin 1
omeprazole nifedipine 1
phenytoin midazolam 1
ranitidine nifedipine 2
bull 28 trials (17 victimperpetrator pairs)
Empirical Approaches
9
eg of positive control
Intended to indicate an in vitro induction response
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Sub-objectives ndash Induction Working Group General overview
bull Define acceptable approaches for the prediction of the clinical induction of CYP3A (models used input parameters needed what type of data is needed)
bull Standardize input parameters that can be standardized (eg fmCYP kdeg)
bull Define criteria for quality of input parameters
bull Define the extent to which we need to qualify the models
bull Work together to qualify the models identified on the ability to predict clinical DDI
These sub-objectives are in support of the main objective
4
01 1 10 100 10000
5
10
15
20
25
[Ind] M
Effe
ct(e
g F
old-
chan
ge m
RN
A)
Predictions of Induction DDI ndash General Principle
5
Most cytochrome P450 (CYP) induction prediction models use a relationship which include the parameters bull EC50
bull Emax
bull [Inducer]
These prediction models are generally based upon the same principal that being the law of mass action for receptor binding (and activation since induction is an agonist property
[Ind]EC[Ind]EEffect
50
max
Emax = 20-fold
EC50 = 15 microM
Approaches for prediction of clinical CYP3A induction
What modelsapproaches are being used now general methods and examples of approach(es)
Empirical Approaches - of positive control Correlation Methods - Relative Induction Score - CmaxEC50
Mathematical Equations (Mechanistic Static Models) - Net effect model
Ripp et al 2006 DMD 341742 Fahmi et al 2008 DMD 361971
Chu et al 2009 DMD 371339 Persson et al 2006 Pharm Res 2356
Fahmi et al 2008 DMD 361698
FDA Draft Guidance ndash Drug Interaction Studies 2006 Bjornsson et al DMD 2003 31815
Almond et al 2009 Curr Drug Metab 10420
Qualification of the different modelsapproaches Identification of trials and available in vitro induction data
bull Initially identified trials to model from the literature (U Washington Drug Interaction Database) or from internal sources
- Literature searches were focused on DDI studies involving common CYP3A substrates (eg midazolam triazolam alprazolam simvastatin buspirone etc)
- Compiled a list of perpetrators that caused clinical induction or no inductioninhibition of these CYP3A substrates
bull Using the initial search four victim drugs were chosen to further investigate based upon the number of trials that were available and differences in fmCYP and FG (midazolam alprazolam nifedipine and simvastatin)
bull Identified the in vitro (internal) mRNA induction data (and activity) we had with respect to the inducersnon-inducers in the identified trials (EC50 and Emax data) and consolidated it for use in modeling (median values)
7
Trials chosen for the model comparison
8
perpetrator victim of trials
rifampin midazolam 10
rifampin alprazolam 1
rifampin nifedipine 1
carbamazepine midazolam 1
carbamazepine alprazolam 1
nafcillin nifedipine 1
phenobarbital nifedipine 1
pleconaril midazolam 1
rosiglitazone nifedipine 1
Merck MK-1 midazolam 2
Merck MK-2 midazolam 1
pioglitazone simvastatin 1
pioglitazone midazolam 1
troglitazone simvastatin 1
omeprazole nifedipine 1
phenytoin midazolam 1
ranitidine nifedipine 2
bull 28 trials (17 victimperpetrator pairs)
Empirical Approaches
9
eg of positive control
Intended to indicate an in vitro induction response
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
01 1 10 100 10000
5
10
15
20
25
[Ind] M
Effe
ct(e
g F
old-
chan
ge m
RN
A)
Predictions of Induction DDI ndash General Principle
5
Most cytochrome P450 (CYP) induction prediction models use a relationship which include the parameters bull EC50
bull Emax
bull [Inducer]
These prediction models are generally based upon the same principal that being the law of mass action for receptor binding (and activation since induction is an agonist property
[Ind]EC[Ind]EEffect
50
max
Emax = 20-fold
EC50 = 15 microM
Approaches for prediction of clinical CYP3A induction
What modelsapproaches are being used now general methods and examples of approach(es)
Empirical Approaches - of positive control Correlation Methods - Relative Induction Score - CmaxEC50
Mathematical Equations (Mechanistic Static Models) - Net effect model
Ripp et al 2006 DMD 341742 Fahmi et al 2008 DMD 361971
Chu et al 2009 DMD 371339 Persson et al 2006 Pharm Res 2356
Fahmi et al 2008 DMD 361698
FDA Draft Guidance ndash Drug Interaction Studies 2006 Bjornsson et al DMD 2003 31815
Almond et al 2009 Curr Drug Metab 10420
Qualification of the different modelsapproaches Identification of trials and available in vitro induction data
bull Initially identified trials to model from the literature (U Washington Drug Interaction Database) or from internal sources
- Literature searches were focused on DDI studies involving common CYP3A substrates (eg midazolam triazolam alprazolam simvastatin buspirone etc)
- Compiled a list of perpetrators that caused clinical induction or no inductioninhibition of these CYP3A substrates
bull Using the initial search four victim drugs were chosen to further investigate based upon the number of trials that were available and differences in fmCYP and FG (midazolam alprazolam nifedipine and simvastatin)
bull Identified the in vitro (internal) mRNA induction data (and activity) we had with respect to the inducersnon-inducers in the identified trials (EC50 and Emax data) and consolidated it for use in modeling (median values)
7
Trials chosen for the model comparison
8
perpetrator victim of trials
rifampin midazolam 10
rifampin alprazolam 1
rifampin nifedipine 1
carbamazepine midazolam 1
carbamazepine alprazolam 1
nafcillin nifedipine 1
phenobarbital nifedipine 1
pleconaril midazolam 1
rosiglitazone nifedipine 1
Merck MK-1 midazolam 2
Merck MK-2 midazolam 1
pioglitazone simvastatin 1
pioglitazone midazolam 1
troglitazone simvastatin 1
omeprazole nifedipine 1
phenytoin midazolam 1
ranitidine nifedipine 2
bull 28 trials (17 victimperpetrator pairs)
Empirical Approaches
9
eg of positive control
Intended to indicate an in vitro induction response
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Approaches for prediction of clinical CYP3A induction
What modelsapproaches are being used now general methods and examples of approach(es)
Empirical Approaches - of positive control Correlation Methods - Relative Induction Score - CmaxEC50
Mathematical Equations (Mechanistic Static Models) - Net effect model
Ripp et al 2006 DMD 341742 Fahmi et al 2008 DMD 361971
Chu et al 2009 DMD 371339 Persson et al 2006 Pharm Res 2356
Fahmi et al 2008 DMD 361698
FDA Draft Guidance ndash Drug Interaction Studies 2006 Bjornsson et al DMD 2003 31815
Almond et al 2009 Curr Drug Metab 10420
Qualification of the different modelsapproaches Identification of trials and available in vitro induction data
bull Initially identified trials to model from the literature (U Washington Drug Interaction Database) or from internal sources
- Literature searches were focused on DDI studies involving common CYP3A substrates (eg midazolam triazolam alprazolam simvastatin buspirone etc)
- Compiled a list of perpetrators that caused clinical induction or no inductioninhibition of these CYP3A substrates
bull Using the initial search four victim drugs were chosen to further investigate based upon the number of trials that were available and differences in fmCYP and FG (midazolam alprazolam nifedipine and simvastatin)
bull Identified the in vitro (internal) mRNA induction data (and activity) we had with respect to the inducersnon-inducers in the identified trials (EC50 and Emax data) and consolidated it for use in modeling (median values)
7
Trials chosen for the model comparison
8
perpetrator victim of trials
rifampin midazolam 10
rifampin alprazolam 1
rifampin nifedipine 1
carbamazepine midazolam 1
carbamazepine alprazolam 1
nafcillin nifedipine 1
phenobarbital nifedipine 1
pleconaril midazolam 1
rosiglitazone nifedipine 1
Merck MK-1 midazolam 2
Merck MK-2 midazolam 1
pioglitazone simvastatin 1
pioglitazone midazolam 1
troglitazone simvastatin 1
omeprazole nifedipine 1
phenytoin midazolam 1
ranitidine nifedipine 2
bull 28 trials (17 victimperpetrator pairs)
Empirical Approaches
9
eg of positive control
Intended to indicate an in vitro induction response
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Qualification of the different modelsapproaches Identification of trials and available in vitro induction data
bull Initially identified trials to model from the literature (U Washington Drug Interaction Database) or from internal sources
- Literature searches were focused on DDI studies involving common CYP3A substrates (eg midazolam triazolam alprazolam simvastatin buspirone etc)
- Compiled a list of perpetrators that caused clinical induction or no inductioninhibition of these CYP3A substrates
bull Using the initial search four victim drugs were chosen to further investigate based upon the number of trials that were available and differences in fmCYP and FG (midazolam alprazolam nifedipine and simvastatin)
bull Identified the in vitro (internal) mRNA induction data (and activity) we had with respect to the inducersnon-inducers in the identified trials (EC50 and Emax data) and consolidated it for use in modeling (median values)
7
Trials chosen for the model comparison
8
perpetrator victim of trials
rifampin midazolam 10
rifampin alprazolam 1
rifampin nifedipine 1
carbamazepine midazolam 1
carbamazepine alprazolam 1
nafcillin nifedipine 1
phenobarbital nifedipine 1
pleconaril midazolam 1
rosiglitazone nifedipine 1
Merck MK-1 midazolam 2
Merck MK-2 midazolam 1
pioglitazone simvastatin 1
pioglitazone midazolam 1
troglitazone simvastatin 1
omeprazole nifedipine 1
phenytoin midazolam 1
ranitidine nifedipine 2
bull 28 trials (17 victimperpetrator pairs)
Empirical Approaches
9
eg of positive control
Intended to indicate an in vitro induction response
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Trials chosen for the model comparison
8
perpetrator victim of trials
rifampin midazolam 10
rifampin alprazolam 1
rifampin nifedipine 1
carbamazepine midazolam 1
carbamazepine alprazolam 1
nafcillin nifedipine 1
phenobarbital nifedipine 1
pleconaril midazolam 1
rosiglitazone nifedipine 1
Merck MK-1 midazolam 2
Merck MK-2 midazolam 1
pioglitazone simvastatin 1
pioglitazone midazolam 1
troglitazone simvastatin 1
omeprazole nifedipine 1
phenytoin midazolam 1
ranitidine nifedipine 2
bull 28 trials (17 victimperpetrator pairs)
Empirical Approaches
9
eg of positive control
Intended to indicate an in vitro induction response
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Empirical Approaches
9
eg of positive control
Intended to indicate an in vitro induction response
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Empirical Approach- Pros and Cons
10
Pros
Can identify non-inducers
Cons
Not proven to be predictive of DDI ndash may lead to false negativesfalse positives
Does not account for EC50 value
What in vitro [I] would be relevant to decision making for the clinics
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
- Sigmoidal relationships observed between RIS and the percent decrease in the AUC values of a CYP3A substrate (eg midazolam) after co-administration of enzyme inducers
- Relationships used to predict in vivo induction from RIS for test compounds
Fahmi et al 2008 DMD 361971
][IndEC][IndE(RIS) Score Induction Relative
u50
umax
What value to use for [I]
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
The value of [I] Inducer concentration to use in correlation methods or in static models
12
Total Cmax
Total Portal Cmax
Portal Cmax = Cmax + kaDFa
QH
Free Cmax (Cmaxfu)
Free Portal Cmax (Portal Cmaxfu)
ka = absorption rate constant D = dose Fa = fraction of dose absorbed QH = hepatic blood flow Qent = enterocytic blood flow
Systemic
Hepatic
Gut [I]G [I]G= kaDFa
Qent
Fahmi et al (2009) DMD 371658 and references therein
All are surrogates however some have been preferred eg free Cmax for inactivation and induction and free portal Cmax for reversible inhibition
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Correlation Methods ndash Pros and Cons
13
Pros
Published reports of good correlation with clinical outcome for CYP3A inducers
Can use these models with relatively simple assumptions about human PK (eg Cmax) No need to simulate the entire conc-time profile
Cons
A calibration curve is needed for each laboratorylot of hepatocytes for calibrator compounds (~6-8+)
Does not account for CYP3A inducers that are also inhibitors inactivators
Limited to CYP enzymes in which there is sufficient clinical data available to set up a calibration curve (eg CYP3A)
Unlikely universal cut-off criteria (eg RIS value) can be established
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Correlation Methods - Universal cut-off criteria
14
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
Company 1
Relative Induction Score (RIS)
D
ecre
ase
in A
UC
of
CYP
3A v
ictim
dru
g
][IndEC][IndERIS
u50
umax
Universal cut-off values of RIS are not recommended as they can be very different between investigators
Similar conclusions were made for CmaxuEC50
19 trials 12 inducers R2 = 081
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
16 trials 8 inducers R2 = 086
Company 2
01 1 10 100 1000
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120
0001 001 01 1 10 100
0
20
40
60
80
100
120Pfizer Lot 2r2=081
Pfizer Lot 3r2=078
01 1 10 100 1000
0
20
40
60
80
100
120
Pfizer Lot 1r2=068
Novartisr2=086
Merck Lot 1 Merck Lot 2
Relative Induction Score (RIS)
11 trials 6 inducers
Company 3
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Correlation methods- Results Relative Induction Score (RIS) Model
15 True negatives or false positives with respect to induction
The RIS Model worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
Predictions of clinical DDI were made by use of the RIS calibration curve (RIS vs observed DDI) established for all the trials (n=28 17 victimperpetrator pairs)
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
16
Correlation methods- Results CmaxEC50 Model
True negatives or false positives with respect to induction
Predictions of clinical DDI were made by use of the linear regression line of CmaxEC50 vs observed DDI established for all the trials (n=28 17 victimperpetrator pairs)
The CmaxEC50 Model also worked very well in the prediction of the inductionnon-induction effect within the 08-fold (-20 ) boundary
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Mathematical Equations (Mechanistic Static Model)
17
GGGGGmmHHHpo
po
FFCBAffCBAAUCAUC
1
1
1
1
eg Net Effect Model
I
inactdeg
deg
KIkIk
kA inactivation
(TDI)
50
max
ECIIEB
d1 induction
iKIC
1
1
reversible inhibition
Incorporates - fm (fraction of victim drug
metabolized by the affected enzyme)
- FG for CYP3A (fraction of the victim drug escaping first pass metabolism in the gut)
- Inactivation induction and reversible inhibition equations
H= hepatic G = gut
18
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitorsinactivators of CYPs in gut and liver
As long as fmCYP3A is estimated properly can be used for different probe (victim) substrates
Cons
Requires more input parameters compared to Correlation Methods (eg fmCYP3A and FG for victim)
It requires an empirical calibratorscalar factor to bridge in vitro data to in vivo (eg d)
Mathematical Equations ndash Pros and Cons
19
Mathematical Equations ndash Results Net Effect Model
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
18
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitorsinactivators of CYPs in gut and liver
As long as fmCYP3A is estimated properly can be used for different probe (victim) substrates
Cons
Requires more input parameters compared to Correlation Methods (eg fmCYP3A and FG for victim)
It requires an empirical calibratorscalar factor to bridge in vitro data to in vivo (eg d)
Mathematical Equations ndash Pros and Cons
19
Mathematical Equations ndash Results Net Effect Model
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
19
Mathematical Equations ndash Results Net Effect Model
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
The Net Effect Model also proved to be a good model to predict the inductionnon-induction
effect within the 08-fold (-20) boundary
One trial (MK-2) is off-scale in the true negative quadrant (predicted DDI of 16-3-fold depending on [I])
PIOMID PIOMID
Total Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Unbound Portal Vein
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Mixed I
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
PBPK-based (Mechanistic Dynamic Model)
20
eg Simcyp
Furukori et al Neuropsychopharmacology 18364 (1998)
100 mg tid carbamazepine 10 days + 08 mg alprazolam on day 8 Actual = 58 increase in clearance
Predicted = 52 increase in clearance
0 24 48 72 96 120 144 168 192 2160
1000
2000
3000
4000
5000
6000
00
25
50
75
100
Time (hours)
Syst
emic
con
cent
ratio
n of
carb
amaz
epin
e (n
gm
L)
Systemic concentration of
alprazolam (ngm
L)
carbamazepine
alprazolam alone
alprazolam + carbamazepine
day 1 day 8
Incorporates - Various algorithms
incorporating both physiological and test compound properties
- Simulates time-concentration profiles of both perpetrator and victim
- Inactivation induction and reversible inhibition
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
PBPK-based Model ndash Pros and Cons
21
Pros
No calibration curves needed ndash savings in time effort cost
Can be used to predict DDI for compounds that are both inducers and inhibitors inactivators of CYPs in the gut and liver
Physiologically-based model incorporates drug and system dynamics and population aspects
Cons
Accurate predictions require extensive knowledge of drug (victim and perpetrator) parameters
Requires normalization of in vitro induction parameters with rifampin positive control (for IVIVE)
requires that concentration-time profile of the perpetrator to be accurately predicted
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
PBPK-based ndash Results Simcyp Model
22
Time-based model more predictive of actual DDI (AUC) than lsquoSteady-statersquo model Also proven to be a good model to predict the inductionnon-induction effect
within the 08-fold (-20) boundary
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
False positives
True negatives False negatives
True positives
True negatives or false positives with respect to induction
False positives
True negatives False negatives
True positives False positives
True negatives False negatives
True positives
MK-1MID
ROSINIF
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Fold Change in Cmax of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
SS Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
Observed DDI
Pred
icte
d D
DI
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Relative accuracy of the various models
23
Simcyp time-based model AUC ratio (steady-state model for AUC ratio GMFE = 227 GMedFE = 242) The scaling parameter for induction (ie d) in each of the four sets of predictions was estimated through linear regression to a value that minimized the GMFE of the prediction via linear weighted least-squares regression
In general all the models predict the actual DDI with similar accuracy particularly using free or free portal Cmax for [I] in the non-PBPK models
GMFE = geometric mean fold-error GmedFE = geometric median fold-error
Correlation Methods Mathematical Equations PBPK
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Predictions of DDI within 2-fold of actual
24
RIS CmaxEC50
Net Effect Simcyp (ldquoTime-basedrdquo)
08-fold (-20)
08-fold (-20)
within 2-fold of the actual DDI
within 2-fold of the actual DDI
[I] = Cmaxu for static models
Although all models can properly predict DDI within the 08-fold (-20) boundary importantly the least amount of under-predictions were observed with the Net Effect model
Least amount of under-predictions
Fold Change in AUC of Object Drug
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDIPr
edic
ted
DD
I
Unbound Cmax
001 01 1 10001
01
1
10
Observed DDI
Pred
icte
d D
DI
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
New FDA draft guidance decision tree
25
How well does R3 work with our dataset
Courtesy of Dr Shiew-Mei Huang
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Evaluation of the ldquoR3rdquo value to assess risk
26
R3 = 1[1+(Emax x [I]γ)(EC50γ +[I]γ)] where γ = 1 in this case
False positives R3 = 111 or 09
False negatives True negatives
True positives
R3 value using [I] as free Cmax or free portal Cmax appears to work well to categorize DDI risk within the R3 = 09 boundary Using total Cmax would
result in many over-predictions and some false positives
PIOMID ROSINIF MK-2MID
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 16 1800
02
04
06
08
10
12
14
16
18
[I] = free Cmax
[I] = free portal Cmax
[I] = total Cmax
Observed DDI (fold-change)
R3
valu
e
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
SummaryConclusions Highlights of the Working Group outcomes thus far
Defined and worked towards the qualification of approaches to predict clinical induction of CYP3A bull Results of the predictions
- All the models evaluated appeared to predict the DDI effect within the range of the 08-fold (-20) boundary
- Some differences in the models to predict within 2-fold of actual with the Net Effect Model having the least amount of under-predictions in this dataset
bull Interpretation across-Co - It is not recommended to establish universal cut-off values (RIS CmaxEC50 and even fold-
change) as these are likely only valid for a particular lot of hepatocytes and the laboratory
bull Ease of use - The lack of the need to set-up of calibration curves such as for the Net Effect Model or
PBPK models is a major advantage to these models - mRNA is the preferred data to use in these models (better dynamic range and
differentiation of induction vs inhibitioninactivation)
Proposed some good starting estimates for those input parameters that we believe can be standardized (with the inhibition team) and defined quality of some important input parameters (eg EC50 and Emax)
27
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Induction Working Group A global collaboration across Academia Pharma and the FDA
28
bull Heidi Einolf (Novartis) bull Scott Obach (Pfizer) bull Lei Zhang (FDA) bull Ping Zhao (FDA) ndash ad hoc bull Odette Fahmi (Pfizer) bull Mohamad Shebley (Abbott) bull Chris Gibson (Merck) bull Liangfu Chen (GSK) bull Jash Unadkat (University of Washington) bull Mike Sinz (BMS) bull Jose Silva (JampJ)
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Back-up slides
29
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Input parameters needed in the different models A review
Correlation Methods bull Of test perpetrator (and calibrator compounds) EC50 Emax [I] bull Known clinical DDI effect of calibrator compounds
Mathematical Equations (Mechanistic Static Models) bull Of test perpetrator EC50 Emax [I] bull Scaling factor ldquodrdquo bull If also a reversible andor TDI Ki andor KI and kinact (also kdeg CYP3A4) fumic
bull Depending on value used for [I] ka fa dose may be needed (eg hepintlet Cmax) fu
bull Of victim fmCYP FG
Physiologically-Based Pharmacokinetic (Mechanistic Dynamic Models) bull Of test perpetrator (and rifampin control as a calibrator) EC50 Emax
bull If also a reversible andor TDI Ki andor KI and kinact fumic
bull Of both victim and perpetrator physical chemical properties protein binding and blood cell distribution absorption parameters Vss renal and hepatic clearances (correct fmCYP being calculated for victim) etc
30
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Input parameters that can be standardized
fmCYP and FG for common CYP3A victim drugs (input values we chose to use for this evaluation)
- Midazolam bull fmCYP range 086-094 (used 090) bull FG point estimate 05
- Alprazolam bull fmCYP 08 bull FG 094
- Nifedipine bull fmCYP 071 bull FG 078
- Simvastatin bull fmCYP 092 bull FG 058
kdeg CYP3A - Hepatic
bull kdeg 002h (tfrac12 36h)
- Intestinal bull kdeg 003h
31
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Details of the modeling using Simcyp
Design of the trials were as described in the publication bull Demographics (number of subjects gender age range) bull Dose(s) dosing regimen route of administration
Victim drug input parameters were used as already provided in Simcyp (Version 1010 SP1)
Perpetrator drug input parameters were qualified (in most cases the compound file had to be built) bull Ensured in vitro induction parameters (EC50 and Emax) were identical as those used in
the other modelsapproaches (calibrationnormalization with rifampin was included) bull Other DDI parameters were included when available from the literature (Ki kinact and
KI) bull Qualified the perpetrator input parameters by comparing simulated time-concentration
curves and PK parameters with actual clinical trials
32
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Perpetrator compound model building Qualification of perpetrator inputs for PBPK modeling
Example Rosiglitazone 8 mg single dose (SD)
33 Simulated multiple dose (MD) PK is not substantially different than SD
0
100
200
300
400
500
600
700
800
900
1000
0 2 5 7 10 12 14 17 19 22 24
Sy
ste
mic
Co
nce
ntr
ati
on
(ng
mL)
Time (h)
Mean Values of Systemic concentration in plasma of Rosiglitazone over Time
CSys Upper CI for CSys Lower CI for CSys
Chapelsky 2003 JCP 43252 Park 2004 CPT 75157 Park 2004 BJCP 58397
Cox 2000 DMD 28772
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Evaluations with and without rifampin calibration PBPK-Simcyp model
34
Fold Change in AUC of Object Drug
00 02 04 06 08 10 12 14 1600
02
04
06
08
10
12
14
16
with RIF calibrationwithout RIF calibration
Observed DDI
Pred
icte
d D
DI
AUC ratio Cmax ratio ldquoSteady-staterdquo AUC ratio vs
Observed AUC ratio
With RIF Calibration
Without RIF Calibration
GMFE 240 245
GMedFE 135 154
With RIF Calibration
Without RIF Calibration
GMFE 348 370
GMedFE 179 221
With RIF Calibration
Without RIF Calibration
GMFE 244 264
GMedFE 150 209
Without RIF calibration a consistently greater DDI (induction effect) is predicted In general the calibration with RIF resulted in lower GMFE or GMedFE values
False positives
True negatives False negatives
True positives
Needs to be
updated
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Why the ldquo40rdquo rule may result in false negpos
35
Simulation Based Upon Fitted RIS Model
CmaxuEC50
0001 001 01 1 10 100 1000
D
ecre
ase
in O
ral M
idaz
olam
AU
C
0
20
40
60
80
100
120
Emax 100 of RifampinEmax 50 of RifampinEmax 25 RifampinEmax 125 RifampinEmax 63 RifampinEmax 38 Rifampin
bull Each curve represents a prediction of DDI from a RIS calibration curve for a test compound bull Each compound (curve) has an EC50 of 1 and a different Emax represented as a certain of RIF bull As the CmaxEC50 ratio is varied the projected DDI varies
Theoretically even compounds with a low of RIF (eg le25) could have the potential for a DDI if ldquo[I]EC50rdquo is high enough (gt01)
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Relative induction of CYP3A4 mRNA by several compounds in replicate experiments
36
0
20
40
60
80
100
Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8
R
elat
ive
to R
ifam
pin
(mR
NA
)
Modafinil Oxcarbazepine
Topiramate Nafcillin
Nevirapine
0
20
40
60
80
100
Run 1
Run 2
Run 3
Run 4
Run 5
Run 6
Run 7
Run 8
R
elat
ive
to R
ifam
pin
(Act
ivit
y)
Modafinil Oxcarbazepine
Topiramate nafcilin
Nevirapine
Human Cryopreserved hepatocytes Lot Hu4165 CYP3A4-mRNA Relative to Rifampin (based on mRNA)
Compound Name Conc (uM) Run 1 Run 2 Run 3 Run 4 Run 5 Run 6 Run 7 Run 8
Evidence that some compounds may not follow rifampin CYP3A4 mRNA induction to the same
extent in different experiments
Pink highlight indicates gt40 of RIF
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
EC50 and Emax values determined for Rifampin Within Co run-to-run variability using the same lot of hepatocytes
37
Emax based on mRNARun Name Celsis Lot NPV SD
1 1 11 162 2 14 0933 3 28 10
mean 18CV 51
Company 1
Company 2
Emax based on mRNARun Name Lot Hu4165 SDC100614 43 3C100607 12 1C090810 19 1C100628 24 2C100628M 24 3mean 24CV 48
EC50 based on mRNARun Name Lot Hu4165 SDC100614 043 013C100607 035 011C090810 070 020C100628 36 10C100628M 076 026mean 12CV 117
EC50 based on mRNARun Name Celsis Lot NPV SD1 430 2002 072 0073 060 013mean 19CV 112
Company 3
Emax based on mRNARun Name SD
1 392 273 224 13
mean 25CV 43
EC50 based on mRNARun Name SD1 1022 1913 1644 161
mean 15CV 24
Company 4
Emax based on mRNARun Name SD
1 52 732 30 2023 51 3064 40 159
mean 43CV 24
EC50 based on mRNARun Name SD1 188 0522 052 0143 082 0124 094 010
mean 10CV 56
Emax EC50
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Common fitting algorithms for in vitro induction data Calculation of Emax and EC50 values
Simple Emax model
Sigmoidal Emax model
Sigmoid 3-parameter
38
[I]EC[I]EEffect
50
max
[I]EC[I]EEffect
50
max
50max ECIEEffect exp1
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
EC50 and Emax values Collection of data across Co
In vitro induction data compiled across several Pharma Co (Pfizer Novartis Abbott Merck BMS JampJ) bull rifampin (n=20) carbamazepine (n=7) nifedipine (n=7) phenobarbital (n=8)
phenytoin (n=8) pioglitazone (n=7) rosiglitazone (n=5) troglitazone (n=4) bull Additional data provided by Pfizer (nafcillin pleconaril omeprazole ranitidine) and
Merck (MK-1 MK-2)
Several Co evaluated each of the 3 fitting algorithms to determine if one algorithm should be favored over the others bull Akaike Information Criterion (AIC) AIC = Nln (sum of squared residuals)+2P where N = of observations and P = of
parameters fit in the model A minimum AIC would be regarded as the best representation of the data Used for rank ordering
models by goodness of fit (with penalty for increasing number of estimated parameters)
39
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Calculation of EC50 and Emax values Use of different fitting algorithms - Example Nifedipine
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Calculation of EC50 and Emax values Comparisons of different algorithms
bull Overall the team felt that there was no consistent trend and substantial difference to give universal preference of one fitting algorithm over another bull The recommendation would be to monitor the standard error given and if comparing data use the same model
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to
work relatively well in the models
Compiled Co EC50 and Emax data Consistently used the Sigmoid 3-parameter model
The combined data set showed variability in the in vitro induction parameters in the range of 35-127 Based on the results of the predictions these median values appeared to