-
Toxicology and Applied Pharmacology 237 (2009) 317–330
Contents lists available at ScienceDirect
Toxicology and Applied Pharmacology
j ourna l homepage: www.e lsev ie r.com/ locate /ytaap
Synergistic drug–cytokine induction of hepatocellular death as
an in vitro approachfor the study of inflammation-associated
idiosyncratic drug hepatotoxicity☆
Benjamin D. Cosgrove a,c,d, Bracken M. King a, Maya A. Hasan
a,1, Leonidas G. Alexopoulos c,e,2,Paraskevi A. Farazi a,c,3, Bart
S. Hendriks f, Linda G. Griffith a,c,d, Peter K. Sorger c,e, Bruce
Tidor a,b,Jinghai J. Xu f,4, Douglas A. Lauffenburger a,c,d,⁎a
Department of Biological Engineering, Massachusetts Institute of
Technology, Cambridge, MA, USAb Department of Electrical
Engineering and Computer Science, Massachusetts Institute of
Technology, Cambridge, MA, USAc Cell Decision Processes Center,
Massachusetts Institute of Technology, Cambridge, MA, USAd
Biotechnology Process Engineering Center, Massachusetts Institute
of Technology, Cambridge, MA, USAe Department of Systems Biology,
Harvard Medical School, Boston, MA, USAf Pfizer Research Technology
Center, Cambridge, MA, USA
Abbreviations: APAP, acetaminophen; Cmax, averageacetyl ester;
TMRM, tetramethyl rhodamine ester; DILIglutathione; HGM, hepatocyte
growth medium; IFNγ,monochlorobimane; MtMP, mitochondrial
membranemitogen-activated protein kinase; ROS, reactive oxygen s☆
Animal experimentation statement: All animals recethe National
Academy of Sciences and published by the⁎ Corresponding author.
Building 16, Room 343, Mass
E-mail address: [email protected] (D.A. Lauffenburger1 Current
address: Yale School of Medicine, New Have2 Current address:
Department of Mechanical Enginee3 Current address: Department of
Life and Health Scie4 Current address: Department of Automated
Biotech
0041-008X/$ – see front matter © 2009 Elsevier Inc.
Adoi:10.1016/j.taap.2009.04.002
a b s t r a c t
a r t i c l e i n f o
Article history:Received 2 November 2008Revised 29 March
2009Accepted 2 April 2009Available online 9 April 2009
Keywords:Drug-induced liver injuryAdverse drug
reactionsPre-clinical assaysInformation theoryPartial least-squares
modeling
Idiosyncratic drug hepatotoxicity represents a major problem in
drug development due to inadequacy ofcurrent preclinical screening
assays, but recently established rodent models utilizing bacterial
LPS co-administration to induce an inflammatory background have
successfully reproduced idiosyncratichepatotoxicity signatures for
certain drugs. However, the low-throughput nature of these models
rendersthem problematic for employment as preclinical screening
assays. Here, we present an analogous, but high-throughput, in
vitro approach in which drugs are administered to a variety of cell
types (primary human andrat hepatocytes and the human HepG2 cell
line) across a landscape of inflammatory contexts containing LPSand
cytokines TNF, IFNγ, IL-1α, and IL-6. Using this assay, we observed
drug–cytokine hepatotoxicitysynergies for multiple idiosyncratic
hepatotoxicants (ranitidine, trovafloxacin, nefazodone,
nimesulide,clarithromycin, and telithromycin) but not for their
corresponding non-toxic control compounds(famotidine, levofloxacin,
buspirone, and aspirin). A larger compendium of drug–cytokine mix
hepatotoxicitydata demonstrated that hepatotoxicity synergies were
largely potentiated by TNF, IL-1α, and LPS within thecontext of
multi-cytokine mixes. Then, we screened 90 drugs for cytokine
synergy in human hepatocytes andfound that a significantly larger
fraction of the idiosyncratic hepatotoxicants (19%) synergized with
a singlecytokine mix than did the non-hepatotoxic drugs (3%).
Finally, we used an information theoretic approach toascertain
especially informative subsets of cytokine treatments for most
highly effective construction ofregression models for drug- and
cytokine mix-induced hepatotoxicities across these cell systems.
Our resultssuggest that this drug–cytokine co-treatment approach
could provide a useful preclinical tool forinvestigating
inflammation-associated idiosyncratic drug hepatotoxicity.
© 2009 Elsevier Inc. All rights reserved.
plasma maximum drug concentration; CM-H2DCFDA,
5-(and-6)-chloromethyl-2′7′-dichlorodihydrofluorescein diacetate,
drug-induced liver injury; DRAQ5,
1,5-bis{[2-(di-methylamino)ethyl]amino}-4,8-dihydroxyanthracene-9,10-dione;
GSH,interferon-γ; IKK, inhibitor of NF-κB kinase; IL, interleukin;
LDH, lactate dehydrogenase; LPS, lipopolysaccharide;
mBCl,potential; NF-κB, nuclear factor-κB; NSAID, non-steroidal
anti-inflammatory drug; PLS, partial least-squares; p38, p38pecies;
SEM, standard error of the mean; STAT, signal transducer and
activator of transcription; TNF, tumor necrosis factor-α.ived
humane care according to the criteria outlined in the “Guide for
the Care and Use of Laboratory Animals” prepared byNational
Institutes of Health (NIH publication 86-23).achusetts Institute of
Technology, 77 Massachusetts Avenue, Cambridge, MA 02139, USA. Fax:
+1 617 258 0204.).n, CT, USA.ring, National Technical University of
Athens, Athens, Greece.nces, University of Nicosia, Nicosia,
Cyprus.nology, Merck & Co., North Wales, PA, USA.
ll rights reserved.
mailto:[email protected]://dx.doi.org/10.1016/j.taap.2009.04.002http://www.sciencedirect.com/science/journal/0041008X
-
318 B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
Introduction
Idiosyncratic drug hepatotoxicity is defined as drug-induced
liverinjury that occurs in a very small fraction of human patients,
isunrelated to the pharmacologic target of the drug, and is
hostdependent (Ganey et al., 2004; Kaplowitz, 2005; Uetrecht,
2007).Idiosyncratic drug hepatotoxicity is poorly predicted by
standardpreclinical cell culture and animal models as well as in
clinical trials,and, consequently, most idiosyncratic drug
hepatotoxicities are notevident until after approval for human use.
Due to the inability topredict idiosyncratic hepatotoxicities in
the drug developmentprocess, idiosyncratic drug hepatotoxicity
frequently leads to drugwithdrawal or “black box” warnings and
accounts for more than 10%of acute liver failure cases (Uetrecht,
2003; Kaplowitz, 2005). Multiplehypotheses have been suggested to
explain the mechanisms under-lying idiosyncratic drug
hepatotoxicity. These include (i) variations indrug metabolism,
particularly associated with alterations in theexpression and/or
activities of the cytochrome P450 family enzymes,due to variable
environmental conditions and/or genetic polymorph-isms in the
humanpopulation (Uetrecht, 2008); and (ii) a relationshipwith
concomitant liver inflammation associated with viral or
bacterialinfection or liver or inflammatory disease (Ganey et al.,
2004).Moreover, it is likely that multiple factors – both genetic
andenvironmental – contribute, at relative degrees which are
notpredictable at the present time, to a drug's hepatotoxicity
idiosyn-crasies (Peters, 2005).
A number of preclinical models have been developed in attemptsto
predict idiosyncratic drug hepatotoxicity, including the
assessmentof reactive metabolites through glutathione (GSH)
conjugation assaysand the evaluation of animals models by
toxicogenomic andmetabolonomic approaches to identify common
idiosyncratic hepa-totoxicity-associated biomarkers, with little
overall predictive success(Kaplowitz, 2005; Peters, 2005; Obach et
al., 2008). Rodent modelsadministered with bacterial
lipopolysaccharide (LPS) have beenrecently developed to assess
inflammation-associated idiosyncraticdrug hepatotoxicity. In these
rodent models, LPS exposure induces amild inflammatory response
that has been demonstrated to synergis-tically induce
hepatotoxicity in the presence of a number ofidiosyncratic
hepatotoxic drugs, including diclonfenac, sulindac,trovafloxacin,
ranitidine, chlorpromazine, but not non- or less-toxiccontrol drugs
(Buchweitz et al., 2002; Luyendyk et al., 2003; Deng etal., 2006;
Shaw et al., 2007). In rats, LPS administration upregulatesplasma
concentrations of the cytokines tumor necrosis factor-α
(TNF),interferon-γ (IFNγ), interleukin-1α and -1β (IL-1α/β),
interleukin-6(IL-6), and the chemokine interleukin-10 (IL-10)
(Bergheim et al.,2006). Of these, TNF, IFNγ, IL-1α/β, IL-6, and LPS
itself all stimulatehepatocyte signaling responses through the
activation of a diversity ofintracellular signal transduction
pathways, including the IKK–NF-κB,p38, and JNK pathways (associated
with TNF, IL-1α/β, and LPSsignaling) and the STAT1 and STAT3
pathways (associated with IFNγand IL-6 signaling, respectively),
which all are implicated inhepatocellular death in liver diseases
and injuries (reviewed inLuedde and Trautwein, 2006; Schwabe and
Brenner, 2006; Malhiand Gores, 2008; Tacke et al., 2009). In
LPS-administered rat models,synergistic induction of hepatocellular
death in the presence of theidiosyncratic hepatotoxicants
ranitidine and trovafloxacin has beenreported to be dependent on
TNF signaling (Shaw et al., 2007, 2009;Tukov et al., 2007). The
observations in LPS-administered rodentmodels suggest that
idiosyncratic drug hepatotoxicity can arise whenmild drug-induced
hepatocellular stresses synergize with LPS-induced inflammatory
cytokine signaling to elicit acute hepatocellulardeath (Ganey et
al., 2004; Kaplowitz, 2005). These stresses may beidiosyncratic in
nature in human patients due variations in drugmetabolism,
exposure, and/or clearance. The sensitizing role ofhepatocellular
stress is supported by the fact that drug-induceddepletion of
glutathione is known to sensitize hepatocytes to TNF-
induced apoptosis (Mari et al., 2008). Furthermore, both LPS
andinflammatory cytokine signaling can alter hepatocyte expression
ofcytochrome P450 enzymes and thus lead to dysregulated
drugmetabolism and clearance in conditions of LPS-induced liver
inflam-mation (Warren et al., 1999; Zolfaghari et al., 2007).
Although theyoffer promise for improved predictability of
idiosyncratic hepatotoxi-city in preclinical screening,
LPS-administered rodent models lacksufficient throughput for
preclinical screening of candidate pharma-ceuticals. Moreover, it
has been shown that animal models are ingeneral not highly
predictive of human drug hepatotoxicity, ascombined preclinical
testing in rodents, dogs, and monkeys can onlyidentify ∼50% of
known human hepatotoxicants (Olson et al., 2000).
Recent advances in the maintenance and characterization of
invitro hepatocyte culture systems offer substantial promise for
theirmore wide-spread utilization in high-throughput preclinical
screen-ing approaches for the prediction of both non-idiosyncratic
andidiosyncratic drug hepatotoxicity in humans. Amongst
hepatocyteculture systems that are commonly employed for
high-throughputpreclinical studies, primary human hepatocytes are
considered the“gold standard” for evaluating drug metabolism,
transport, andtoxicity (LeCluyse et al., 2005; Hewitt et al.,
2007). In comparison,primary rat hepatocytes, while more readily
available and similarlycapable of maintaining differentiated
hepatic function in time-scalesof a few days in vitro, do not
reproduce some aspects of human drugmetabolism (Xu et al., 2004;
Sivaraman et al., 2005). Immortalizedand transformed human cell
lines (e.g. HepG2 cells) are alsofrequently employed but have poor
maintenance of liver-specificfunctions and are relatively
insensitive to human hepatotoxicants insimple cytotoxicity assays
(Xu et al., 2004; O'Brien and Haskins, 2007).A small number of
hepatocyte cell culture models have been recentlydeveloped to
assess idiosyncratic drug hepatotoxicity. Of note, Xu et
al.utilized human hepatocyte cell culture models to assay four
sub-lethalhepatotoxicity injuries with high-throughput live-cell
microscopy forover 300 drugs, including many that cause
idiosyncratic liver toxicityin humans (Xu et al., 2008). Using a
well-calibrated random forestprediction model of the imaging data,
they were able to predict drughepatotoxicity with a ∼50%
true-positive rate and ∼5% false-positiverate. A rat
hepatocyte-Kupffer cell co-culture model has beendeveloped and
shown to successfully predict chlorpromazine idiosyn-cratic
hepatotoxicity through its synergistic induction of hepatocel-lular
death following LPS treatment (Tukov et al., 2006). The
furtherdevelopment and validation of hepatocyte cell culture models
wouldprovide much-needed tools for the preclinical evaluation of
idiosyn-cratic drug hepatotoxicity and could offer greater
predictive abilityand higher throughput than LPS-administered
animal models.
Here, we describe a model of inflammatory
cytokine-associatedidiosyncratic drug hepatotoxicity in three
standard hepatocyte cellculture systems amenable to high-throughput
preclinical screening —primary rat and human hepatocytes and the
HepG2 humanhepatoblastoma cell line. We initially validate this
model todemonstrate that a number of idiosyncratic hepatotoxic
drugs(ranitidine, trovafloxacin, nefazodone, nimesulide,
clarithromycin)synergistically induce hepatocellular death in vitro
when co-administered with a cytokine mix containing the
LPS-upregulatedcytokines TNF, IFNγ, and IL-1α, and LPS itself. We
then collect ahepatotoxicity data compendium comprised of
combinations of drugand cytokine mix co-treatments covering ∼1500
experimentalconditions and analyze it to identify informative
cytokine mixtreatments and hepatocyte cell systems for predicting
inflamma-tion-associated idiosyncratic drug hepatotoxicity. Using
this datacompendium, we show that in vitro drug–cytokine synergies
arepredominantly potentiated by TNF, IL-1α, and LPS within
thecontext of multi-cytokine mixes and that patterns of
drug–cytokinemix synergies across a landscape of multi-cytokine
environmentscan be shown to correlate to drug-induced sub-lethal
hepatocyteinjury signatures. Then, we demonstrate the screening
utility of this
-
319B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
drug–cytokine mix co-treatment model by assaying a set of
90drugs in human hepatocytes and show that a significantly
largerfraction of idiosyncratic hepatotoxicants synergize with a
singlecytokine mix at physiologically relevant dosing
concentrations thando non-toxic drugs. Lastly, we also employ an
information theoretictechnique to identify subsets of cytokine
co-treatment conditionsthat maintain the information contained
across the full set ofcytokine conditions in the compendium. We
show that theseinformative condition sets can be transferred across
cell systemsand act as better trainings sets for predicting drug-
and cytokine-induced hepatotoxicities in primary human hepatocytes.
Our resultsindicate promise for employing our approach for
efficient in vitroinvestigation of inflammation-associated
idiosyncratic drughepatotoxicity.
Methods
Drugs and cytokines. Most drugs were obtained from Sigma
(St.Louis, MO) or Sequoia Research Products (Pangbourne,
UK).Trovafloxacin was obtained from Pfizer's chemical sample
bank(Groton, CT). Unless otherwise noted, the following
drugconcentrations were used: 450 μM ranitidine, 450 μM
trovafloxacin,70 μM nefazodone, 450 μM nimesulide, 175 μM
clarithromycin, and175 μM telithromycin. These drug concentrations
were selected frominitial dosing studies based on the criteria that
the drugconcentration (i) elicit minimal drug-only hepatotoxicity,
(ii) inducerobust supra-additive hepatotoxicity synergy with a
representativecytokine mix, and (iii) be within a physiologically
relevant dosinglimit of 100-fold its Cmax value, which was
satisfied for all drugsexcept ranitidine, for which a dose of
317⁎Cmax was selected (seeFigs. 1 and S6 for additional details).
TPCA-1, an IKK-2 inhibitor, was
Fig. 1. Identification of drug dose-dependent hepatotoxicity
synergies between a cytokine mand HepG2 cells (panels F–J). Primary
rat hepatocytes and HepG2 cells were cultured, treatwere dosed at
varying concentrations in the presence or absence of a cytokine mix
containingwere fold-change normalized to DMSO/no cytokine control
samples from the same cell systchemical class and/or molecular
target are plotted together, with the less or
non-hepatotoxicpresented as mean±SEM of four biological samples.
Results from additional time-points,values, are shown in Figs.
S1–S5.
obtained from Tocris Bioscience (Ellisville, MS). All drugs
weresuspended in 0.25% final DMSO.
Recombinant rat or human cytokines were obtained from
R&DSystems (Minneapolis, MN) and were used at the
followingsaturating concentrations: 100 ng/ml tumor necrosis
factor-α(TNF), 100 ng/ml interferon-γ (IFNγ), 20 ng/ml
interleukin-1α(IL-1α), and 20 ng/ml interleukin-6 (IL-6).
Lipopolysaccharides(LPS) serotype 1 from E. coli 0111:B4 was used
at 10 μg/ml, aspreviously (Geller et al., 1993). In most cell
culture studies,especially involving monocyte or macrophages (e.g.
Kupffer cells),lower LPS concentrations (∼1–10 ng/ml) are used
(Bellezzo et al.,1996). Though hepatocytes express the LPS receptor
TLR4, they aresubstantially less responsive to LPS than are
macrophages (Geller etal., 1993; Bellezzo et al., 1996).
Consequently, a sufficiently high LPSconcentration was selected to
ensure that LPS was fully stimulatinghepatocytes and not just the
very small fraction of Kupffer cells(present at ∼0.4% in both human
and rat hepatocyte seedingisolates) that may have remained viable
2–3 days post-seeding.Unless noted, all reagents were obtained from
Sigma.
Drug hepatotoxicity classifications and pharmacokinetic
properties.Drug hepatotoxicity classifications were made according
to a drug-induced liver injury (DILI) scale (see Table S1) based on
clinical datacollected from PubMed searches, as in Xu et al.
(2008). For selectdrugs, idiosyncratic hepatotoxicity
classifications were assignedaccording to literature references
(see Table S2). Therapeuticallyappropriate drug exposure levels
were defined by average plasmamaximum concentration (Cmax) values
observed in humans uponsingle- or multi-dose administration at
commonly recommendedtherapeutic doses. Cmax values were obtained
from a combination ofliterature searches and available databases,
as in Xu et al. (2008),
ix and multiple idiosyncratic hepatotoxic drugs in primary rat
hepatocytes (panels A–E)ed, and assayed for LDH (at 24 or 48 h
post-treatment) as described in Methods. Drugs100 ng/ml TNF, 100
ng/ml IFNγ, 20 ng/ml IL-1α, and 10 μg/ml LPS. LDH release
values
em. (Note that LDH release axes are separately scaled for each
plot.) Drugs from similar“comparison” drug in blue and the more
idiosyncratic hepatotoxic drug in red. Data are
with drug doses plotted with respect to both molecular
concentrations and drug Cmax
-
320 B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
and are reported in Table S3. Unless noted otherwise,
aconcentration of 100-fold Cmax, encompassing a scaling factor
toaccount for human population pharmacokinetic and
toxicodynamicvariabilities, was considered as a therapeutically
relevant dosinglimit for each drug, as previously discussed (Xu et
al., 2008).
Liver cell isolation, culture, and stimulation. Primary rat
hepatocyteswere isolated from male Fisher rats using a modified
collagenaseperfusion and Percoll isolation, routinely yielding N90%
viability, asdescribed previously (Cosgrove et al., 2008). Cell
type purity wasassessed by flow cytometry and showed that rat
hepatocyte isolateswere routinely comprised of ∼97% hepatocytes
(albumin+-cytokeratin-18+ cells), ∼0.4% Kupffer cells (ED2/CD163+),
∼0.4%stellate cells (GFAP+), and ∼0.2% sinusoidal endothelial cells
(SE-1+),as previously reported (Cosgrove et al., 2008). Rat
hepatocytes wereseeded on collagen type I-coated 96-well plates (BD
Biosciences,Franklin Lakes, NJ) at 1×105 cells/cm2 in
insulin-containing, serum-free hepatocyte growth medium (HGM; as in
Cosgrove et al., 2008but supplemented with 1 μM trichostatin A).
One day post-seeding,rat hepatocytes were overlayed with 0.25 mg/ml
Matrigel (growthfactor-reduced; BD Biosciences) in fresh HGM. One
day followingMatrigel overlay, primary rat hepatocytes were
stimulated withdrugs and/or cytokines in fresh HGM. For rat and
human hepatocytestudies, multiple donors were used throughout this
work, with asingle donor used for each self-consistent data set.
For rathepatocytes, donor-to-donor variability was assessed by
comparingtwo drug- and cytokine mix-induced hepatotoxicity data
compendia(each consisting of the same 256 treatment conditions)
collectedfrom two separate primary rat hepatocyte isolations. The
twoseparate data compendia showed a high degree of
reproducibility(R=0.98; see Fig. S10).
Primary human hepatocytes were obtained in suspension
fromCellzDirect (Durham,NC). Cell typepuritywas
assessedbyflowcytometryand showed that human hepatocyte isolates
were routinely comprised of∼97% hepatocytes
(albumin+-cytokeratin-18+ cells), ∼0.4% Kupffer cells(CD163+),∼0.3%
stellate cells (GFAP+),∼0.03%endothelial cells (CD31+),and 0.03%
bile epithelial cells (cytokeratin-19+) (L.G. Alexopoulos,
B.D.Cosgrove, and P.K. Sorger, unpublished results). Human
hepatocytes wereseeded on collagen type I-coated 96-well plates at
1.5×105 cells/cm2 in“plating medium” consisting of Dulbecco's
modified Eagle's medium(DMEM) supplementedwith 5% fetal bovine
serum (FBS; Hyclone, Logan,UT), 100 U/ml penicillin, 100 μg/ml
streptomycin, 0.58 mg/ml L-glutamine, 1 μM trichostatin A, 0.5 μM
dexamethasone, and 5 μg/mlinsulin. One day post-seeding, human
hepatocytes were overlayed with0.25 mg/ml Matrigel in “culturing
medium” consisting of William's Emedium (WEM) supplemented with 15
mM HEPES, 100 U/ml penicillin,100 μg/ml streptomycin, 0.29 mg/ml
L-glutamine, 1 μM trichostatin A,0.1 μMdexamethasone, 5 μg/ml
insulin, 5 μg/ml transferrin, and 5 ng/mlsodium selenite. One day
following Matrigel overlay, human hepatocyteswere stimulated with
drugs and/or cytokines in “dosing medium”(consisting of “culturing
medium” but without transferrin and sodiumselenite).
HepG2 cells were obtained from ATCC (Manassas, VA) and
weremaintained per ATCC recommendations. HepG2 cells were seeded
oncollagen type I-coated 96-well plates at 1×105 cells/cm2 in
Eagle'sminimum essential medium (EMEM; ATCC) supplemented with
10%FBS, 100 U/ml penicillin, and 100 μg/ml streptomycin. One day
afterseeding, medium was changed to fresh EMEM without FBS. One
dayafter medium change, HepG2 cells were stimulated with drugs
and/orcytokines in fresh EMEM without FBS. All cells were
maintained at37 °C and 5% CO2.
Quantitative cell apoptosis and death assays. At 12, 24, or 48 h
post-drug and/or cytokine treatment, conditioned medium samples
werecollected to assay lactate dehydrogenase (LDH) release
(indicator ofnecrotic and apoptotic cell death) and cells were
assayed for caspase 3/7
activity (indicator of apoptotic cell death). LDH activity in
culturesupernatants was quantified using a CytoTox-ONE
HomogeneousMembrane Integrity Assay (Promega, Madison, WI)
according tomanufacturer's recommendations. Cellular caspase 3/7
activity wasquantified using a Caspase-Glo 3/7 Assay (Promega)
according tomanufacturer's recommendations. For each cell system
and time-point,LDH and caspase 3/7 activity assay results were
fold-change normalizedto the average DMSO control/no cytokine
treatment value from four ormore biological samples from the same
96-well culture plate.
Quantitative sub-lethal hepatotoxicity imaging assays.
Drug-inducedsub-lethal hepatotoxicity phenotypes were
quantitatively imaged inhuman hepatocytes in the absence of
cytokine co-treatment,essentially as described previously (Xu et
al., 2008). Briefly, humanhepatocytes at 24 or 48 h post-treatment
were stained with fourfluorescent probes: DRAQ5 (Biostatus,
Shapshed, UK) to stain nucleiand lipids, CM-H2DCFDA (Invitrogen,
Carlsbad, CA) to stain reactiveoxygen species (ROS), TMRM
(Invitrogen) to stain mitochondrialmembrane potential (MtMP), and
mBCl (Invitrogen) to stainglutathione (GSH). Automated live-cell,
multi-color image acquisitionwas performed on a Kinetic Scan Reader
(Cellomics, Pittsburgh, PA)using a 20× objective and a XP93 filter
set (Omega Optical,Brattleboro, VT). Quantitative image analysis
was performed usingImagePro Plus software (Media Cybermetrics,
Bethesada, MD). In eachimage, five features were quantified: nuclei
count and intracellularlipid (non-nuclear DRAQ5 stain), ROS, MtMP,
and GSH contents. Foreach feature, the summed intensity value from
each well wasnormalized by the total nuclei count (∼500 imaged per
well), andthen the intensity-per-cell values were fold-change
normalized to theaverage DMSO control value from eight or more
biological samplesfrom the same 96-well culture plate.
Factorial analysis. The drug–cytokine mix
hepatotoxicitycompendium was collected such that a full factorial
design of thefive cytokine or LPS treatments (25=32 mixes) was
included for eachdrug co-treatment in each cell system. For each
drug/cell system, thefold-change normalized toxicity assay values
were subjected tofactorial analysis. One-, two-, three-, four-, and
five-factor effectsand their associated errors were calculated
according to standardfactorial analysis formulae (Box et al.,
1978).
Hierarchical clustering. The drug–cytokine mix
combinatorialhepatotoxicity compendium was fused across all cell
systems andassay types to generate a hepatotoxicity matrix spanning
192“experimental” conditions (i.e., combinations of cell type,
assay readout,and cytokine treatment) and 8 drug treatments. For
each combination ofcell system and assay type, the fold-change
normalized values werelinearly mapped to a scale from the minimum
observed value (set to 0)and the maximum observed value (set to 1).
The fused datacompendiumwas subjected to two-wayclusteringusing
theunweightedpair groupmethodwith arithmetic mean and a Pearson
distancemetric.
Joint entropy-based representative subset selection. To
identifysubsets of cytokine treatments that maximally maintain the
diversityof hepatotoxicity responses across different cell systems
and assays, arepresentative subset selection technique was applied
using aninformation theoretic scoring function. All subsets
considered wereconstrained to include the no cytokine treatment as
well as alltreatments containing only a single cytokine or LPS. To
this set of sixconditions, all possible combinations of other
cytokine treatmentcombinations were added. Each candidate subset
was then scored bycomputing the joint entropy of all drug
hepatotoxicity responses(using the fold-change normalized data
fromboth the caspase 3/7 andLDH release assays) across the subset
using data from: (i) only primaryhumanhepatocytes, (ii) only
primary rat hepatocytes, (iii) only HepG2cells, or (iv) both rat
hepatocytes andHepG2 cells. Joint entropieswere
-
321B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
computed by first discretizing each drug response into
threeequiprobable bins and then applying the second-order
MaximumInformation Spanning Tree approximation (King and Tidor, in
press).Having scored all cytokine condition subsets, the maximum
entropyset of each size (i.e. number of cytokine treatment
conditions) wasidentified. Additionally, a consensus set for each
size was selected byincluding the conditions that were most
prevalent in the top 100subsets. See Supplementary methods for
additional details.
Partial least-squares (PLS) regression modeling. To examine
theutility of themaximum entropy sets described above, we used them
asa basis for training predictive PLS models of hepatotoxicity
across, andwithin, cell types and drug treatments. PLS models were
built usingthe PLSREGRESS function in MATLAB release 2008a (The
MathworksInc., Natick, MA), which implements the SIMPLS algorithm.
Allvariables were variance-scaled (with respect to the training
set)prior to learning the model and all models used two
principalcomponents. Separate models were built treating the
hepatotoxicityresponse profile of each drug in human hepatocytes
across thecytokine conditions as the dependent variable. One class
of models(“single drug”) was built in which only the response
profiles to thesame drug in rat hepatocytes and/or HepG2 cells were
used asindependent variables. In a second class of models, all drug
profiles inrat hepatocytes and/or HepG2 cells were used. Models
were builtusing the training sets (containing 25 of the 32 cytokine
conditions)and then used to predict the values of the test sets
(containing theremaining 7 conditions), and the correlation
coefficient betweenpredicted and observed values in the test set
was computed. For therandomly chosen sets, all sets were enforced
to include the nocytokine and all five single cytokine/LPS
conditions, as this constraintwas also applied to the maximum
entropy consensus sets. SeeSupplementary methods for additional
details.
Statistical analysis. To identify drug–cytokine mix
co-treatmentconditions that elicited supra-additive hepatotoxicity
synergies,additive projections of drug–cytokine mix co-treatments
wereestimated by adding mean values of drug-only and cytokine
mix-only toxicities and propagating their associated variances.
Supra-additive synergies were identified for conditions in which
theobserved drug–cytokine mix co-treatment results exceeded
theadditive projections as assessed by a two-sample,
one-tailed(Student's) t test with a false discovery rate correction
for multiplecomparison testing for multiple drug doses or multiple
cytokinemixes. The statistical significance of each factorial
effect and itsassociated error was assessed using a one-sample,
two-tailed t testwith a false discovery rate correction for
multiple cytokine mixes.Statistical significance of drug-induced
sub-lethal hepatotoxicitieswas assessed by a Student's t test. In
the 90-drug study, a thresholdtwo-fold above the additive
projection was used to identify supra-additive drug–cytokine mix
synergy due to the limited number ofreplicate samples instead of a
Student's t test. The statisticalsignificance of the observed
number of synergistic drugs in the eachhepatotoxic group was
assessed using a hypergeometric test (seeTable 1 for details). All
tests were performed at a significance level ofα=0.05. False
discovery rate-corrected P-values were calculated
as:P=α·(N+1)/(2N), where N is the number of comparisons.
Results
Several idiosyncratic hepatotoxic drugs, but not their
control-pairedcompounds, exhibit drug–cytokine mix hepatotoxicity
synergies in vitro
We developed an in vitro model of
inflammation-associatedidiosyncratic drug hepatotoxicity by
co-administering drug com-pounds with known idiosyncratic
hepatotoxicities in humans with avariety of inflammatory cytokines
mixtures (comprised of the
cytokines TNF, IFNγ, IL-1α, and IL-6, along with LPS) in
multiplehepatocellular cell culture systems (primary human and rat
hepato-cytes and HepG2 human hepatoblastoma cells). In developing
thismodel, we investigated drug–cytokine mix hepatotoxicity
synergiesfor six idiosyncratic hepatotoxic drugs: ranitidine,
trovafloxacin,nefazodone, nimesulide, telithromycin, and
clarithromycin (a “com-parison” compound for telithromycin also
with idiosyncratic hepato-toxicity). For each drug compound
associated with idiosyncratichepatotoxicity, a less- or
non-hepatotoxic “comparison” controlcompound was used. In this
study, the term “comparison” compoundwas applied to drugs with
similar molecular target and clinicalindication and, where
possible, similar chemical structure. See TableS2 for additional
information on these drugs and their correspondingcomparison
compounds. Initially, this in vitro drug–cytokine mix co-treatment
model was applied to primary rat hepatocytes and HepG2cells treated
with five pairs of drug compounds in the presence orabsence of a
single cytokine mix containing TNF, IFNγ, IL-1α, and LPSand assayed
for LDH release as a marker of both apoptotic and necroticcell
death (Figs. 1 and S1–S5). Synergistic induction of
hepatocellulardeath was assessed by a supra-additive synergy
criterion thatcompares the experimentally observed cell death
induced by drugand cytokine co-treatment to the additive projection
of cell deathobserved for drug-only and cytokine mix-only
treatments (Fig. S6).
In this co-treatment model, we observed drug–cytokine
mixsynergies for ranitidine but not cimetidine or famotidine (data
notshown) in rat hepatocytes (but not HepG2 cells), matching
similarobservations in a LPS-administered rat model (Luyendyk et
al., 2003).We observed drug–cytokine mix synergies for
trovafloxacin but notlevofloxacin in both rat hepatocytes and HepG2
cells, again matchingsimilar observation in a LPS-administered
mouse model (Shaw et al.,2007). For drugs not previously examined
in LPS-administered animalmodels, we observed drug–cytokine mix
synergies for nefazodone (butnot buspirone) and clarithromycin in
both rat hepatocytes and HepG2cells, and nimesulide (but not
aspirin) and telithromycin in only HepG2cells. In this initial
study, drug–cytokine mix synergies were observedonly for the more
idiosyncratic hepatotoxic drugs, except for clarithro-mycin and
telithromycin, which both synergized with cytokine mix co-treatment
and both have associated idiosyncratic hepatotoxicity
withtelithromycin having a greater incidence (see Table S2 and
Peters, 2005;Clay et al., 2006). Drug–cytokine mix hepatotoxicity
synergies wereobservedwithin 24 h following co-treatment except for
ranitidine in rathepatocytes,which required48hof co-treatment to
elicit hepatotoxicitysynergy, demonstrating that, at the drug and
cytokine treatmentconcentrations used, this in vitro model captures
acute, rather thanchronic, hepatotoxicity responses. The delay in
ranitidine–cytokinemixsynergy compared to other compounds, in
concert with the observationthat it onlyoccurs in rathepatocytes
andnotHepG2cells, indicates that amore prolonged mechanism (e.g.
requiring significant accumulation ofranitidine metabolites) may be
required to potentiate ranitidine–cytokine hepatotoxicity
synergy.
Specific concentrations and time-points for each drug
wereselected for further investigation (see summary in Table S2)
basedon the criteria that the concentration induce robust
supra-additivehepatotoxicity synergy with this representative
cytokine mix (Fig. S6)and elicit minimal drug-only hepatotoxicity.
This selection criteriaallowed for identification of drug
concentrations within a physiologi-cally relevant dosing limit of
100-fold its Cmax value (see Methods foradditional explanation) for
all cytokine-synergizing drugs exceptranitidine, for which a dose
of 450 μM or 317⁎Cmax was used.
Collection and analysis of a combinatorial drug- and cytokine
mix-induced hepatotoxicity compendium from multiple hepatocyte
cellsystems
To characterize drug–cytokinemix synergies in amore diverse set
ofcytokine environments and tomake comparisons across hepatocyte
cell
http://dx.doi.org/doi:10.1093/bioinformatics/btp109
-
322 B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
culture systems, we collected a multi-cue data compendium from
allcombinations of six idiosyncratic hepatotoxic drugs from the
initialstudy – each at one concentration and time-point – and aDMSO
control,and the 32 combinatorial mixtures of TNF, IFNγ, IL-1α,
IL-6, and LPS.(Note that IL-6, not included in the initial study
due to its presumed pro-survival effects (Luedde and Trautwein,
2006), was included in thiscombinatorial study.) Experiments were
performed in rat hepatocytes,human hepatocytes, and HepG2 cells and
then assayed for both LDHrelease and caspase 3/7 activity, amarker
specific to apoptotic cell death(Figs. 2 and S7–S9). The
hepatotoxicity data compendium, comprised of∼1500 combinations of
cell system, assay type, and drug–cytokinetreatment, was observed
to contain a diverse array of drug–cytokinesynergy patterns not
clearly interpretable by inspection alone, so wesubjected it to
four analytical approaches. (i) We discretized thehepatotoxicity
data compendium into conditions that did or did notelicit
supra-additive drug–cytokine mix synergy (Fig. S11). (ii)
Wesubjected the hepatotoxicity data compendium to factorial
analysis toidentify which underlying cytokine treatment factors
potentiate celldeath across the entire combinatorial landscape of
cytokine environ-ments (Figs. S12–S15). (iii) We employed
hierarchical clustering of thehepatotoxicity data compendiumwith
respect to both drug treatmentsand “experimental” conditions (i.e.,
combinations of cell type, assayreadout, and cytokine treatment;
Fig. 3A). (iv) We identified subsets ofcytokine conditions
thatmaximallymaintained the information contentof the response data
(Fig. 5).
From the first analysis approach, examining discretized
drug–cytokine synergyclassification (Fig. S11), it is evident that
higher-order(four- or five-factor) cytokine environments were more
efficient atidentifying possible drug synergies (∼50% of possible
synergies acrosscombination of all cell systems and drugs) than
were lower-order(one-, two-, or three-factor) environments (∼15–35%
of possiblesynergies). Of note, there are higher-order cytokine
mixes other thanthe mix of TNF, IFNγ, IL-1α, and LPS (which was
used in the initialstudy) that are more efficient at synergizing
with these idiosyncraticdrugs in human hepatocytes. This is in part
due to the fact that theinitial cytokine mix is mildly toxic by
itself for human primaryhepatocytes, limiting its ability to
synergize with drug co-treatmentsin a supra-additive manner.
Instead, slightly less-toxic five-factormixes (in particular, the
five-factor mixes that instead do not containeither TNF or IFNγ
[the latter noted as “2” in Fig. S11A]) are far moreefficient at
eliciting supra-additive hepatotoxicity synergieswith thesesix
idiosyncratic drugs in human hepatocytes, and therefore wouldlikely
serve as a more predictive cytokine environment for
assessingdrug–cytokine synergies in human hepatocytes.
In the second analysis approach, we applied factorial analysis
to thehepatotoxicity data compendium to identify underlying
cytokineeffects potentiating drug–cytokine hepatotoxicity synergies
acrossthe entire landscape of cytokine environments (Figs.
S12–S15). Asimplemented here, factorial analysis calculates the
effect of theaddition or removal of component treatment
“variables”, eachcontaining one-to-four cytokines and/or LPS, from
all treatmentconditions in which they are present or absent, and,
as such,summarizes the average effect of each cytokine treatment
“variable”within the context of all other cytokine co-treatment
conditions (Boxet al.,1978). Higher-order factorial effects (those
containingmore thantwo-cytokine variables)were generallymodest,
and, instead, one- andtwo-cytokine factorial effects dominated the
observed drug- andcytokine mix-induced hepatotoxicities. The most
significant effects,across all cell types and drug co-treatments,
arose from the singlecytokine treatment variables of TNF and, to a
lesser extent, IL-1α andLPS (see Supplementalmethods and Results
for additional discussion).The identification via factorial
analysis of TNF as a potentiallyimportant cytokine mediator of the
inflammation-associated hepato-toxicity of multiple idiosyncratic
drugs is in agreement with reportsthat TNF mediates the LPS-induced
hepatotoxicity of both ranitidine(Tukov et al., 2007) and
trovafloxacin (Shaw et al., 2007, 2009).
Clustering of the drug–cytokine mix hepatotoxicity compendium
toassess hepatotoxicity patterns with respect to cell system, assay
type,and drug-induced hepatocellular injuries
Factorial analysis of the hepatotoxicity data compendium
sug-gested a significant degree of variability in cytokine factors
potentiat-ing idiosyncratic drug hepatotoxicity synergies in
different drugbackgrounds and cell systems. To further assess these
differences, wefused the hepatotoxicity data compendium into a
single data matrix of192 “experimental” conditions (comprised of
all combinations of threecell systems, two assay types, and five
cytokine/LPS treatmentvariables) by eight “drug” conditions (six
idiosyncratic drugs andtwo DMSO controls). This hepatotoxicity data
matrix was subjected totwo-way hierarchical clustering to assess
patterns of drug–cytokinemix synergies across both the 192
experimental conditions and the 8drug or DMSO backgrounds (Fig. 3).
Pearson clustering yielded themost distinct separationwith respect
to assay readouts due to the factthat they are poorly correlated
(R=0.18) across the entire data set. Asecond notable groupingwas
that of the different cell types, with largesections of each assay
type cluster consisting solely of the conditionsfrom each cell
system, showing that there was little overlap betweenthe three
hepatocyte cell systems. Inspection within the LDH andcaspase data
clusters revealed that neither rat hepatocytes nor HepG2cells were
distinctly better correlated with human hepatocytes in thisdata
compendium.
It has been proposed that a conserved mechanism of
inflammation-associated idiosyncratic toxicity is that sub-lethal
hepatocellular injuriesinduced by idiosyncratic drugs and/or their
metabolites sensitizehepatocytes to undergo cytokine-stimulated
cell death (Kaplowitz,2005). We hypothesized that there might exist
correlations betweenthepatterns of drug–cytokinemix lethal
hepatotoxicities and a set of foursub-lethal hepatocyte injury
measurements for the six idiosyncraticdrugs in this study.
Drug-induced sub-lethal hepatocyte injuries weremeasured in human
hepatocytes, in the absence of any cytokines, using
ahigh-throughput live-cell microscopy approach (Xu et al., 2008),
whichquantifies lipid content, reactive oxygen species (ROS),
mitochondrialmembrane potential (MtMP), and glutathione (GSH)
depletion (Fig.S16). It was difficult to discern clear correlations
between sub-lethalinjuries (measured in human hepatocytes) and
cytokine synergyclustering patterns (across all three hepatocyte
systems) for nimesulide,clarithromycin, and nefazodone due to the
numerous sub-lethalhepatotoxicities induced by these drugs (Fig.
3A), and thus synergycorrelations possibly reflect the convolution
of multiple sub-lethalinjury–cytokine synergy mechanisms. In
contrast, the only sub-lethalinjury induced by both telithromycin
and trovafloxacin that wasstatistically significant was MtMP
depletion (Figs. 4C and S16). Inhuman hepatocytes, telithromycin
and trovafloxacin elicited markedlysimilar patterns of cytokine
synergy as assayed by caspase 3/7 activityand represented through
factorial analysis (Fig. 3B). For both drugs,cytokine synergy
effects were evident, in decreasing magnitude, for LPSand IL-1α but
not other treatment variables. This pattern of cytokinesynergy
effectswas not shared by any other drugs at 24 h post-treatmentin
human hepatocytes (Fig. S14). This unique and specific
sub-lethalinjury–cytokine synergy relationship suggests that
drug-induced mito-chondrial injury may sensitize hepatocytes to
apoptosis induced by LPSand IL-1α, as has been similarly
hypothesized for alcoholic hepatitis-inducedmitochondrial injury
inhepatocytes (Hoek andPastorino, 2002).
Large-scale screen in primary human hepatocytes demonstrates
theutility of cytokine co-treatment synergy model as a tool for
identifyingidiosyncratic hepatotoxic drugs
To test the drug–cytokine mix hepatotoxicity synergy model as
ascreening tool for identifying inflammation-associated
idiosyncraticdrug hepatotoxicity, we assayed drug–cytokine mix
synergy for 90drugs in human hepatocytes. This set of 90 drugs
included 53
-
Fig. 2. A drug- and cytokine mix-induced hepatotoxicity data
compendium. Primary rat hepatocytes (far left), primary human
hepatocytes (left center), and HepG2 cells (rightcenter) were
cultured, treated, and assayed for caspase 3/7 activity (top) or
LDH release (bottom) at 24 or 48 h post-treatment as described
inMethods. In rat hepatocytes, the DMSOcontrol and ranitidine
treatment conditions at t=48 h are shown in expanded bar plots for
both assay types (far right). Bar plot graphs for all combinations
of cell type, assay type,and treatment condition are shown in Figs.
S7–S9. Caspase 3/7 activity and LDH release values were both
fold-change normalized to DMSO/no cytokine samples from the same
cellsystem. In the heatmaps, mean toxicity assay values of three to
six biological samples are plotted using linear color-scales
indexed separately to the minimum and maximumobserved value for
each combination of cell system and assay type. In the bar plots,
data are plotted as mean±SEM of four biological samples, with all
conditions demonstratingstatistically significant supra-additive
drug–cytokinemix synergy labeled (⁎) (see Methods). The cytokinemix
(TNF, IFNγ, IL-1α, and LPS) used in Figs. 1 and S1–S5 is noted as
“Mix”.Abbreviations: Cla, clarithromycin; Tel, telithromycin; Nef,
nefazodone; Tro, trovafloxacin; Nim, nimesulide; Ran,
ranitidine.
323B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
hepatotoxic drugs from DILI classes P1, O1, and P2 and 36
non-hepatotoxic drugs from DILI classes O2, N3, N2, and N1 (see
Tables S1and S3 for additional details). DILI class P2 is
substantially comprisedof drugs with idiosyncratic hepatotoxicities
in humans and thereforeassumed for analysis purposes here and
previously (Xu et al., 2008) to
represent idiosyncratic drugs. The non-hepatotoxic group (DILI
O2,N3-N1) is used to provide corresponding non-toxic control
com-pounds, although we note that the idiosyncratic drugs
clarithromycinand ranitidine used in the initial study here are in
DILI classes N1 andN3, respectively. In this 90-drug screen,
comparisons were made by
-
Fig. 3. Hierarchical clustering of the drug–cytokine mix
hepatotoxicity compendium. (A) The drug–cytokine mix combinatorial
hepatotoxicity compendium was fused across allcell systems and
assay types into a single data matrix, which was then subjected to
two-way Pearson clustering (top left; see Methods for additional
details). First, clusteringwas used to re-sort a matrix of 192
“experimental” conditions, comprised of combinations of three cell
systems, two assay types, and five cytokine treatment variables
(topright). Second, this clustering was used re-sort to a
sub-lethal hepatotoxicity data matrix of eight drug conditions and
four drug (only)-induced sub-lethal hepatotoxicities(bottom). The
sub-lethal hepatotoxicities (measured by quantitative imaging in
primary human hepatocytes; see Fig. S16) are plotted in the bottom
heatmap using linear color-scales indexed separately to the minimum
and maximum observed toxicity value for each assay type. (Note that
the MtMP and GSH assay scales are inverted compared to
Figs.S16K–L.) Conditions used for the large-scale primary human
hepatocyte toxicity study (see Fig. 4) are noted: (1) no cytokines
and (2) TNF, IL-1α, IL-6, and LPS. (B) Factorialeffects±errors of
all one- and two-cytokine effects from the caspase 3/7 activity
data at t=24 h in primary human hepatocytes for DMSO control,
telithromycin, andtrovafloxacin drug treatments. Statistically
significant factorial effects (see Methods and Fig. S12) are
labeled (⁎). (C) Sub-lethal hepatotoxicities measured in primary
humanhepatocytes treated with DMSO control, telithromycin, or
trovafloxacin are plotted on a normalized scale as in panel (A).
Data are presented as mean±SEM of five biologicalsamples. For each
assay type, treatments significantly different from the DMSO
control are labeled as significant (⁎) if Pb0.05 by a Student's t
test. Abbreviations: RH, primaryrat hepatocytes; HH, primary human
hepatocytes; G2, HepG2 cells; Cla, clarithromycin; Tel,
telithromycin; Nef, nefazodone; Tro, trovafloxacin; Nim,
nimesulide; Ran, ranitidine;ROS, reactive oxygen species; MtMP,
mitochondrial membrane potential; GSH, glutathione.
324 B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
examining the differences between the idiosyncratic group (DILI
P2)and the non-hepatotoxic group as not all idiosyncratic drugs
could beindividually paired with “comparison” control drugs.
Due to practical limitations in conducting
medium-to-high-throughput screens in primary human hepatocytes, we
assesseddrug–cytokine mix synergy only for a single cytokine mix
(TNF, IL-1α, IL-6, and LPS), which was equally effective at
inducing hepato-toxicity synergies across the six idiosyncratic
drugs as the full set ofcytokinemixes in the initial
drug–cytokinemix data compendium (seeFig. S11A,mix noted as “2”).
Humanhepatocyteswere treatedwith oneof 90 drugs, each dosed between
0 and 150 μM, in the presence orabsence of TNF, IL-1α, IL-6, and
LPS and assayed for LDH release at 24 hpost-treatment (Figs. 4 and
S17). Supra-additive drug–cytokinesynergywas assessedwith regard to
two differentmethods of defininga physiologically relevant dosing
limit: (i) using each drug's own100⁎Cmax concentration, or (ii)
using multiples (33× or 100×) of the
median Cmax concentration for all drugs in this study (0.91 μM)
as ageneral estimate of physiological exposure limit, which may be
anecessary approximation if clinical human pharmacokinetic data
isunavailable.
For doses less than each drug's own 100⁎Cmax
concentration,drug–cytokine mix synergy was observed for the P1
compoundsbenzbromarone, demeclocycline, azathioprine, amiodarone,
retinoicacid; the O1 compound menadione; the P2 compounds
trovafloxacin,diclofenac, quinine, chlorpromazine, riluzole,
mexiletine, clomipra-mine, nortriptyline; and the N1 compound
clarithromycin (which hasreported idiosyncratic hepatotoxicity in
humans (Peters, 2005; Clay etal., 2006) and was used as a test
idiosyncratic hepatotoxicant in theinitial study here). Among these
cytokine mix synergy compounds,three of the six overtly hepatotoxic
drugs (P1 compounds benzbro-marone and azathioprine and the O1
compound menadione) and twoof the eight idiosyncratic hepatotoxic
drugs (P2 compounds quinine
-
Fig. 4. Large-scale drug–cytokine mix hepatotoxicity study in
primary human hepatocytes demonstrates the utility of cytokine
co-treatment approach for identifying idiosyncratichepatotoxic
drugs. Primary human hepatocyteswere cultured, treated, and assayed
for LDH release (at 24 h post-treatment) as described inMethods.
Ninety drugs (see Table S3)wereeach dosed at seven non-zero
concentrations (2.5× serial dilutions from a high concentration of
150 μM) in the presence or absence of a cytokine mix containing
TNF, IL-1α, IL-6, andLPS. The differential between+and− cytokinemix
co-treatment for each drug dosewas calculated and is plotted in the
heatmaps (see Fig. S17 for rawdata and additional details).
Theheatmaps are split into hepatotoxic (DILI classes P1, O1, and
P2; left) and not or minimally hepatotoxic (DILI classes O2, N3,
N2, and N1; right) drug groups, with these DILI classessorted in
order of decreasing hepatotoxicity (see Tables 1 and S1 for
additional details). Note that DILI class P2 is substantially
comprised of drugswith idiosyncratic hepatotoxicities inhumans.
Within each DILI class, drugs are sorted in order of 100⁎Cmax value
(a physiologically relevant dosing limit). Drug 100⁎Cmax values are
plotted in an overlayed line plot, withvalues exceeding 150 μMnot
shown. Individual drug doses that exhibited supra-additive
drug–cytokinemix synergy (seeMethods and Table 1) at concentrations
less than their drug's100⁎Cmax limit are highlightedwith gray
boxes. Drugs with one ormore dose exhibiting drug–cytokinemix
supra-additive toxicity synergy at less than their 100⁎Cmax
concentrationare listed in red font. A representativeDILI P2 drug
(chlorpromazine) displaying drug–cytokinemix synergy at dosing
concentrations less than 100⁎Cmax is shown in the expanded plotat
the bottom right (data presentedmean±SEMof twobiological samples).
TPCA-1, a smallmolecule IKK inhibitor (IKKi), was used (at ten-fold
lower concentrations than arenoted bythe axis labels for the other
drugs) as a positive control for drug–cytokine mix synergy, but is
not labeled in red as its Cmax is unknown.
325B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
and chlorpromazine) also induced significant drug-only
hepatotoxi-city at doses less than each drug's own 100⁎Cmax
concentration (Fig.S17 and data not shown). In this data set,
drug-only hepatotoxicity
was defined as greater than two-fold increase in LDH release.
Usingthis approach to physiological concentration limit, a
significantlylarger fraction of the idiosyncratic hepatotoxic drugs
(8 of 43=19%)
-
326 B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
demonstrated hepatotoxicity synergy with the cytokine mix than
didthe non-hepatotoxic drugs (1 of 36=3%; see Table 1). In
contrast,using 100-fold the median Cmax concentration (91 μM) as a
generalestimate of physiologically relevant dosing limit, the
idiosyncratichepatotoxic drugs (9 of 43=21%) did not elicit more
frequenthepatotoxicity synergy then the non-hepatotoxic drugs (7
of36=19%). But in using 33-fold the median Cmax concentration(30
μM), the idiosyncratic hepatotoxic drugs (7 of 43=16%) didelicit
more frequent hepatotoxicity synergy then the non-hepatotoxicdrugs
(0 of 36=0%).
This demonstrates that synergistic induction of
hepatotoxicitywith a cytokine mix, even when limited to a single
hepatocyte cellsystem and cytokine mix, can be utilized as a
screening tool forevaluating inflammation-associated idiosyncratic
drug hepatotoxicity.As implemented here, optimized assessment
requires knowledge ofthe drug's Cmax value, which necessitates
human clinical pharmaco-kinetic data, but a reduced set of
idiosyncrasies can be reproducedwith a more conservative estimate
of liver exposure (33⁎Cmax) basedon a generalized Cmax estimate
calculated frommany drugs. Moreover,idiosyncratic hepatotoxic drugs
(P2 compounds) largely induce drug–cytokine mix synergies in the
absence of drug-only hepatotoxicities,which aremore often evident
for synergizing drugs that are associatedwith overt hepatotoxicity.
This dependency on Cmax to optimallycalibrate the drug–cytokine mix
synergy model to distinguishbetween drugs with idiosyncratic
hepatotoxicity and those witheither no or overt hepatotoxicity is
in concert with Paracelsus' conceptthat “exposure makes a
poison”.
Joint entropy analysis identifies informative cytokine
environments in rathepatocytes and HepG2 cells for characterizing
human hepatocyte drug–cytokine mix death synergies
The poor correlation in drug–cytokine hepatotoxicity
patternsbetween all three hepatocytic cell types (Fig. 3) indicated
that it
Table 1Drug–cytokine hepatotoxicity synergies in the large-scale
human hepatocyte toxicitystudy evaluated by DILI class and
physiological dosing limit.
Drug hepatotoxicity classificationby DILI categoriesa
N Drugs with one or more dose with cytokinesynergy within the
applied physiologicaldosing limitb
Using each drug'sCmax value
Using the medianCmax value
100⁎Cmax 33⁎Cmax 100⁎Cmax 33⁎Cmax
Idiosyncratic hepatotoxic(DILI P2)
43 8 (19%) 7 (16%) 9 (21%) 7 (16%)
Not or minimally hepatotoxic(DILI O2, N3-N1)
36 1 (3%) 0 (0%) 7 (19%) 0 (0%)
Hypergeometric test P-valuec 0.028 0.011 0.55 0.011
a DILI categories are described in Table S1.b Hepatotoxicity in
this large-scale primary human hepatocyte study was assayed by
LDH release (see Figs. 5 and S17). Drug–cytokine synergies that
were at least two-fold(with respect to the LDH release, reported as
fold-change compared to the DMSO/nocytokine control condition)
greater than the calculated supra-additive synergythreshold were
characterized as synergistic. This “rule-of-thumb” threshold was
usedinstead of a supra-additive Student's t test (see Methods and
Fig. S6) as two or threebiological samples per condition were used
due to the screening nature of this large-scale study, thus
limiting the ability to satisfy statistical significance tests.
Drug–cytokine synergy was assessed with regard to two different
methods of defining aphysiologically relevant dosing limit: (i)
using each drug's own 33⁎Cmax or 100⁎Cmaxconcentration, or (ii)
using the median 33⁎Cmax (30 μM) or 100⁎Cmax (91 μM)concentration
for all drugs in this study (as a general estimate of physiological
exposurelimit). If a drug demonstrated cytokine synergy at one or
more dosing concentrationwithin the physiological dosing limit
applied, it was included in the aggregate for itshepatotoxicty
class. See Fig. 5 for a list (those drugs in red font) of the drugs
thatsatisfied synergy condition using each drug's 100⁎Cmax
concentration limit.
c The statistical significance of the observed number of
synergistic drugs in theidiosyncratic hepatotoxic class was
assessed using a hypergeometric test with a nullhypothesis that
synergistic drugs would not preferentially populate either of
thehepatotoxicity groupings.
would be difficult to make accurate predictions of drug- and
cytokine-induced hepatotoxicities in primary human hepatocytes from
primaryrat hepatocyte and/or HepG2 data. To examine this further,
we askedwhether a subset of the 32 cytokine conditions present in
the drug–cytokine mix data compendium could be identified that
maintainedthe diversity of the hepatotoxicity responses present in
the full set ofexperiments. Such a subset could prove useful for
motivating futureexperiments in a more compact experimental scope.
To quantify theinformation content of a set of cytokine
co-treatments, we computedthe joint entropy of all drug
hepatotoxicity responses across bothassays and in all three cell
systems using Maximum InformationSpanning Trees, a framework for
approximating high-dimensionalinformation theoretic statistics with
limited sample sizes (King andTidor, in press). This joint entropy
metric quantifies the diversity ofhepatotoxicity responses across
the entire data compendium,accounting for relationships between
drugs, assays, and cell systems.We assessed the information content
of all possible subsets ofcytokine co-treatments using the joint
entropy metric, and selectedthe highest scoring subset of each size
(Fig. 5A) in the humanhepatocyte data set, as well as a consensus
set that considered the 100most information-rich subsets of each
size (data not shown). Thisanalysis identified subsets containing
16–19 cytokine co-treatmentswith comparable information content
(91–99%) to the full set of 32.The consensus subsets, while not as
information-rich, also had similarinformation content to the full
set (89–98%).
While the above results suggest that the number of
conditionsused could be reduced by as much as half while
maintaining muchof the information content, the selection scheme
relied uponknowledge of the full data set. While primary human
hepatocytesare considered to be the cell system with most
predictive of humanhepatotoxicity (Gomez-Lechon et al., 2003;
LeCluyse et al., 2005),they are also less amenable to
high-throughput investigation due totheir higher cost and lower
availability. We therefore asked whethera compact subset of highly
informative treatment conditions forhuman hepatocyte studies could
be identified using only theexperimental data from primary rat
hepatocytes and HepG2 cells.We repeated the joint entropy analysis
described above using datafrom (i) only HepG2 cells, (ii) only rat
hepatocytes, or (iii) both rathepatocytes and HepG2 cells (Fig.
S18). We then evaluated thetreatment subsets chosen from each data
set by their joint entropyin the human hepatocyte data (Figs.
5B–C). While the sets chosenwithout the human hepatocyte data were
less informative than theoptimal sets (e.g. 95% of the total
possible entropy for the optimal16-treatment set compared to 90%
using a 16-treatment set chosenby rat hepatocyte and HepG2 data),
they still performed signifi-cantly better than randomly chosen
sets (which averaged 77% of thetotal possible entropy for
16-treatment sets). Additionally, the setschosen using both the rat
hepatocytes and HepG2 data generallyoutperformed sets chosen using
either data set alone, suggestingthat both cell systems were
separately informative of the humanhepatocyte hepatotoxicity
observations. The results demonstratethat reduced-condition
treatment sets can be identified fromexhaustive data sets collected
in cell systems more amenable tohigh-throughput investigations (rat
hepatocytes and HepG2 cells)for more focused, but still fully
informative, experiments in higher-cost systems (human
hepatocytes).
Maximum entropy sets act as better training sets for
partialleast-squares regression models of drug- and
cytokine-inducedtoxicity in primary human hepatocytes
To examine the utility of the maximum entropy sets
describedabove, we used them as a basis for training a predictive
model. Wetook the approach of building statistical regression
models topredict the drug- and cytokine mix-induced
hepatotoxicityresponses in human hepatocytes, based upon the
analogous
http://dx.doi.org/doi:10.1093/bioinformatics/btp109http://dx.doi.org/doi:10.1093/bioinformatics/btp109
-
Fig. 5. Representative subset identification using joint entropy
analysis. In panel A, subsets of cytokine co-treatments that
maximally maintained the diversity of the full experimentalhuman
hepatocyte (HH) data set were identified by exhaustively scoring
all possible subsets that contained the no cytokine and single
cytokine/LPS treatment conditions. Thetreatments contained in the
highest scoring set of each size are indicated by the white boxes.
Red bars represent the joint entropy of the maximally informative
set. After 16–19 co-treatments are selected, additional
co-treatments do not increase the joint entropy, indicating that
the diversity of the full data set can be capturedwith awell-chosen
set of 16–19 co-treatments. See Fig. S18 for maximum entropy subset
plots for the rat hepatocyte (RH) and HepG2 (G2) data sets. In
panel B, maximally informative consensus subsets were chosenusing
only RH data, only G2 data, both RH and G2 data, or only HH data.
The mean performance of the top 100 subsets chosen from each cell
systemwhen scored for joint entropy inthe HH data is plotted, along
with themean and standard deviation joint entropy for all possible
subsets for the HH data. Sets chosen based on RH and G2 data still
performwell whenscored against the HH data. In panel C, a single
consensus set of each size was chosen from each cell system and the
scored for joint entropy in the HH data. The probability ofrandomly
choosing a subset with higher joint entropy is plotted as a
function of set size. Low values indicate that it is unlikely to
randomly select a set with higher informationcontent than the
evaluated set. The dashed line represents the average of all
possible subsets.
327B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
measurements in the rat hepatocyte and HepG2 cell systems.
Foreach drug and hepatotoxicity assay, we built a partial
least-squares(PLS) regression model that predicted the
hepatotoxicity responsesin human hepatocytes based upon training
data from rat hepato-cytes and/or HepG2 cells. We built the models
using two differentsets of inputs that reflect different
experimental modalities. In onecase (“single drug”), when building
a model of the response of agiven drug across the various cytokine
backgrounds in humanhepatocytes, we used as training data only the
response to the samedrug in both other cell systems and both
assays. In a second case(“all drugs”), the response profiles of all
drugs were used to predicteach individual drug response. To
evaluate the predictive perfor-mance of these PLS models, a subset
of the cytokine conditionscontaining 25 cytokine conditions
(training set) were used to trainthe model, and the trained model
was then used to predict theresponse in the remaining 7 cytokine
conditions (test set). These setsizes showed the best predictive
performance as well as the mostimprovement using maximum
entropy-selected sets compared torandom sets (see Fig. S19). As a
comparison, we chose 1000 sets ofeach size at random and trained
models from these sets, with themean performance across all drug
responses reported.
We examined whether the information-rich maximum entropysets
would act as better training sets when learning the PLS models.All
maximum entropy consensus sets had higher mean correlationbetween
predictions and observed values than the randomly chosensets (Fig.
6, Table S4). Importantly, the consensus sets identified usingthe
data from HepG2 cells or both rat hepatocytes and HepG2 cellsboth
lead to PLS models that performed much better than
randomly-selected subsets. Additionally, the models trained using
all of the drugdata consistently outperformed the single drug
models, suggestingthat parallel drugs studies may prove useful in
predicting the behaviorof individual drugs. These modeling results
show that, in addition tothe identity of information-rich cytokine
treatment conditions beingtransferable across cell systems, the
subsets themselves can also act assuperior training sets for
predicting hepatotoxicity responses inunmeasured conditions.
Discussion
Hepatotoxicity is a major cause of failures in both the
clinicaland post-approval stages of drug development and thus
representsa major challenge for the pharmaceutical industry
(Kaplowitz,
-
Fig. 6. Predictive performance of partial least-squares (PLS)
models trained onconsensus maximum entropy sets and evaluated
across cell systems. PLS regressionmodels were built using specific
cytokine condition sets containing 25 of the 32 possiblecytokine
treatments, selected based on data from rat hepatocytes (RH) and
HepG2 cells(G2). They were evaluated for their ability to predict
responses of the 7 remainingcytokine treatment conditions observed
in human hepatocytes (HH). Pearsoncorrelations (R) between the
observed and predicted responses in HH of both thecaspase 3/7
activity (panel A) and LDH release (panel B) responses are shown.
Cytokinecondition sets selected from RH and G2 data were chosen
using either the consensusmaximum entropy treatment sets or random
treatment sets, of which the meanperformance of 1000 random
treatment sets is plotted, and these conditions sets wereselected
and evaluated using data from either all six drugs or single
drugs.Abbreviations: Cla, clarithromycin; Tel, telithromycin; Nef,
nefazodone; Tro, trovaflox-acin; Nim, nimesulide; Ran,
ranitidine.
328 B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
2001; Lee, 2003). Furthermore, drug hepatotoxicity represents
aserious public health problem, as it is the leading cause of
acuteliver failure in the United States (Ostapowicz et al.,
2002).Idiosyncratic drug hepatotoxicity – a hepatotoxicity subset
thatoccurs in a very small fraction of human patients (∼1 in
10,000)and accounts for ∼10% of acute liver failure cases – is
poorlypredicted by standard preclinical models and in clinical
trials andfrequently leads to post-approval drug failure
(Kaplowitz, 2005).Thus, the development and validation of novel
preclinical tools thatdemonstrate successful identification of
idiosyncratic drug hepato-toxicity is a paramount need for the
pharmaceutical industry andthe public health. Recent findings in
LPS-administered rodentmodels suggest that idiosyncratic drug
hepatotoxicity can arisewhen mild drug-induced hepatocellular
stresses synergize withinflammatory cytokine signaling to elicit
hepatocellular death, butthese models lack sufficient throughput
for preclinical hepatotoxi-city screening (Ganey et al., 2004;
Kaplowitz, 2005).
In this work, we developed and evaluated in vitro hepatocytecell
culture models of idiosyncratic drug hepatotoxicity, which aremore
suitable to the high-throughput demands of
preclinicalpharmaceutical screening. In our cell culture model,
synergisticinduction of hepatocellular death by drugs and
inflammatorycytokines is used to reproduce known idiosyncratic drug
hepato-toxicities. The in vitro cytokine synergy model developed
hereinand other complementary cell culture (Xu et al., 2008) and
animalmodels (Buchweitz et al., 2002; Luyendyk et al., 2003; Shaw
et al.,2007) offer much-needed preclinical tools for the assessment
andprediction of idiosyncratic drug hepatotoxicity. This
cytokinesynergy model would be most useful simply for its ability
toidentify candidate idiosyncratic hepatotoxicity
phenomenologies.Identification of candidate idiosyncratic
hepatotoxicants in cell
culture allows for more detailed follow-up experiments to
parsethe mechanisms of particular candidate idiosyncratic
hepatotox-icants to help guide drug compound development. We
focused thedevelopment of our cell culture model on the simple case
of static,simultaneous drug–cytokine mix co-exposure over an acute
time-scale (24–48 h). Cell culture models more consistent
withidiosyncratic drug hepatotoxicity physiology would need to
includethe relative dynamics of both drug and cytokine
exposure.Clinically, drug exposure can be quite cyclical (i.e.,
oscillatingbetween Cmax and Cmin for a repeatedly administered drug
at apre-defined dosing interval). While cytokine exposure can be
static(due to, for example, a pre-existing inflammatory event),
cytokinelevels can also be highly cyclical, especially in the case
ofhepatotoxicant-induced inflammation (Horn et al., 2000; Ganey
etal., 2004).
We demonstrated that numerous idiosyncratic hepatotoxicdrugs,
but not comparison non-toxic control compounds, synergis-tically
induce death in multiple hepatocyte cell systems when
co-administered with multi-cytokine mixes associated with
LPS-induced liver inflammation (Fig. 1). In primary rat and
humanhepatocyte cultures in particular, drug–cytokine mix synergies
weremost frequently observed for higher-order (containing four or
fivecytokines or LPS) cytokine mixes (Fig. S11), whose
hepatotoxicitywas potentiated in a drug- and cell system-specific
manner by theadditive combination of the single-factor effects of
TNF, IL-1α, and/or LPS (Fig. S12). Our results demonstrate that LPS
itself can amplifydrug hepatotoxicity, but this effect is
predominantly observed whenLPS is co-administered with other
cytokines in the presence ofmulti-cytokine mixes (Fig. S11A). When
LPS was administered inthe absence of other cytokines (mirroring in
vivo studies when LPSis administered upon inhibition or removal of
Kupffer cells), ityielded very infrequent drug synergies. Together,
these observationsconfirm that other cytokines are necessary for
LPS to have an effectin inducing hepatotoxicity in combination with
idiosyncratichepatotoxicants). This agrees with previous reports
(Shaw et al.,2007; Tukov et al., 2007) that Kupffer cell-mediated
cytokinerelease is necessary for LPS-induced hepatotoxicity in
combinationwith idiosyncratic hepatotoxicants in animal models.
Potentiation ofdrug–cytokine synergy by TNF, IL-1α, and LPS, more
so than by IFNγor IL-6, suggests that signal transduction pathways
that are similarlydownstream of the cytokine receptors TNFR and
IL-1R and the LPSreceptor TLR4 (Liu et al., 2002), such as
IKK–NF-κB, p38 (Deng et al.,2008), and JNK (Wu and Cederbaum,
2008), are likely criticalcomponents of hepatocellular toxicity
responses to idiosyncraticdrug-inflammatory cytokine
co-exposure.
Idiosyncratic hepatotoxicants are hypothesized to induce
adiversity of sub-lethal injuries that sensitize hepatocytes
toinflammatory cytokine-induced cell death (Kaplowitz, 2005).
Thishypothesis is supported by the demonstration that
acetaminophen(APAP), at high doses, can elicit an
idiosyncratic-like hepatotoxicitythat is dependent on cytokine
signaling as part of the innateimmune response (Kaplowitz, 2005).
At high doses, accumulationof a cytochrome P450-dependent APAP
metabolite leads todepletion of GSH in hepatocytes, which is known
to sensitizehepatocytes to TNF-induced apoptosis (Mari et al.,
2008). Inlimiting our analysis to hepatotoxicants that induced only
a singletype of sub-lethal injury, we found clear correlations
betweeninjury mechanism and hepatotoxicity effects by different
cytokinetreatment variables via factorial analysis. In human
hepatocytes,telithromycin and trovafloxacin both induced only one
type of sub-lethal injury, mitochondrial membrane potential
depletion (Figs. 3Cand S16) and elicited markedly similar patterns
of cytokine synergyas assayed by caspase 3/7 activity and
represented throughfactorial analysis that were not shared by other
drug treatments(Fig. S16). This unique and specific relationship
between drug-induced injury and drug–cytokine synergy suggests that
drug-
-
329B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
induced mitochondrial injury may sensitize hepatocytes to
apop-tosis induced by LPS and IL-1α, as has been similarly
hypothesizedfor alcoholic hepatitis-induced mitochondrial injury in
hepatocytes(Hoek and Pastorino, 2002), but not TNF. Alternatively,
idiosyn-cratic drug hepatotoxicity could arise when idiosyncratic
hepato-toxicant-induced stresses amplify the hepatotoxic effects of
pre-existing inflammatory cytokine signaling. Our findings do
notdistinguish between these hypothesized mechanisms. Instead
theysimply demonstrate that co-exposure of idiosyncratic
hepatotox-icants and inflammatory cytokines frequently leads to the
syner-gistic amplification of mild hepatotoxicities caused by each.
Furtherinvestigation is necessary to clarify the mechanisms of
inflamma-tion-associated idiosyncratic drug hepatotoxicity both
generally andfor specific drugs.
The cytokine mix-specific responses evident in the
hepatotoxi-city data compendium collected here (Figs. 2 and S11)
suggest thatinflammation-associated idiosyncratic drug
hepatotoxicities mightbe avoided by limiting drug treatments to
patients that do not haveplasma cytokine signatures (due to
pre-existing inflammatoryepisodes, for example) corresponding to
known synergizing inflam-matory environments. Further investigation
of drug–cytokine mixsynergies over across greater number of drug
compounds andcytokine environments, beyond those associated with
LPS-inducedinflammation, would have to be conducted to more
thoroughlycapture the diversity of patient-specific drug–cytokine
interactionsin humans. This suggestion of “personalized” or
“stratified” drugtreatment (Trusheim et al., 2007) to avoid
toxicity would likely benecessitated only for drugs for which
comparably efficaciouscompounds are not available and could be
combined withpharmaco-metabolonomic phenotyping approaches (Clayton
et al.,2006) to avoid both inflammation- and
metabolism-associatedidiosyncratic hepatotoxicities in a
patient-specific manner.
In a 90-drug screen in human hepatocytes, ∼20% of idiosyn-cratic
hepatotoxicants (those compounds associated with DILIcategory P2;
see Table S1) elicited hepatocellular death synergywith a cytokine
mix compared to only 3% of non-hepatotoxic drugswhen using each
drug's 100⁎Cmax concentration as a physiologicaldosing limit (Fig.
4). Using a generalized physiological dosing limitof 30 μM (based
on 33-fold the median Cmax concentration acrossall drugs in the
study), ∼15% of idiosyncratic hepatotoxicants andnone of
non-hepatotoxic drugs elicited synergy. This demonstratesthat,
given drug pharmacokinetic parameters to define a physiolo-gically
relevant dosing window (ideally individually defined foreach drug),
in vitro drug–cytokine hepatocellular death synergy canbe utilized
as a much-needed preclinical tool for identifyingpotential
inflammation-associated idiosyncratic hepatotoxicantswith
reasonable throughput. As conducted here, identification
ofinflammation-associated idiosyncratic drug hepatotoxicity based
onin vitro hepatocellular models depends on human
pharmacokineticdata and would be most reasonably used within an
iterativepreclinical–clinical toxicity assessment paradigm.
Furthermore, thiswork demonstrates the utility of a physiologically
relevant drugdosing limit of 100⁎Cmax to obtain a low
false-positive rate, asmany non-hepatotoxic drugs synergistically
induced human hepa-tocyte death at concentrations exceeding
100⁎Cmax (see Table 1).
At least for a subset of six drugs, this study demonstrates
that, inaddition to human hepatocytes, both rat hepatocytes and
HepG2cells can be useful hepatocellular systems for reproducing
inflam-mation-associated idiosyncratic drug hepatotoxicities. This
unex-pected success in using hepatocellular systems more amenable
tohigh-throughput screening suggests that cytokine mix
synergyscreens may be implemented for preclinical drug evaluation,
butthe utility of these cell systems may be limited to drugs
whosetoxicity is not dependent on in vivo-like metabolite
formation.Moreover, we demonstrate an information theoretic
approach thatcan identify particularly informative subsets of
cytokine treatments
in rat hepatocytes and HepG2 that are not only
equivalentlyinformative of larger data sets in human hepatocytes,
but also arehighly effective at training PLS regression models to
predict drug-and cytokine mix-induced hepatotoxicities across these
cell systems(Figs. 5 and 6). The utility of rat hepatocytes and
HepG2 cells forscreening inflammation-associated idiosyncratic drug
hepatotoxicitywill need to be evaluated for a greater diversity
drug compounds togenerate more confidence in their accuracy.
This work suggests numerous improvements in the
furtherdevelopment of high-throughput cell culture models used to
predictinflammation-associated idiosyncratic drug hepatotoxicity.
In thelarge-scale screen conducted here (Fig. 4), the limited
number ofcytokine synergies with idiosyncratic hepatotoxicants (8
of43=19%) was likely due to the use of only one cytokineenvironment
and one hepatocyte cell system and the fact that notall DILI P2
drugs have idiosyncratic hepatotoxicities associated
withinflammation. Increases in identification accuracy could be
obtainedusing a multi-variate predictive model (Xu et al., 2008)
calibratedfrom expanded measurements of drug–cytokine synergies at
multi-ple doses (up to the 100⁎Cmax limit) across multiple
hepatocyte cellsystems, additional cytokine environments, and/or
toxicity assays,which all were limited in the 90-drug screen here.
Additionally,hepatocyte cell culture models, such as
three-dimensional micro-reactor cultures using primary rat
hepatocytes, that better maintainhepatic drug metabolism and
biliary transport characteristics over achronic time-scale (more
than 7 days) and are scalable to medium-throughput screening
demands (Domansky et al., 2005; Sivaramanet al., 2005; Hwa et al.,
2007) could be utilized to develop morephysiologically relevant
models of inflammation-associated idiosyn-cratic drug
hepatotoxicity. These systems could better capture themix of
chronic and acute hepatocyte responses to drugs andinflammatory
cytokines (Ganey et al., 2004) and, moreover, couldbe used to
recapitulate the cyclic changes in both drug and cytokinelevels in
a well-controlled manner, which were not examined in thisstudy.
Nonetheless, the work presented here validates the use
ofsynergistic induction of hepatocellular death by idiosyncratic
hepa-totoxicants and an inflammatory cytokine environment as a
much-needed in vitro tool for assessing inflammation-associated
idiosyn-cratic drug hepatotoxicity and provides a framework for
furtherdevelopment of such in vitromodels to capture a greater
complexity ofand to elucidate the mechanistic basis of
inflammation-associatedidiosyncratic drug hepatotoxicity.
Conflict of interest statementB.S.H. is employed by and holds
stock in Pfizer. J.J.X. was a past employee of Pfizer, isemployed
by Merck & Co., and owns stock in Pfizer, Merck & Co., and
otherbiopharmaceutical companies.
Acknowledgments
The authors thankArthur Smith andMargaret Dunn for assistance
inconducting hepatotoxicity imaging assays and David de Graaf,
SteveTannenbaum, Ajit Dash, Walker Inman, Justin Pritchard, and
BrianJoughin for helpful discussions. The authors acknowledge
fundingsupport from Pfizer Inc., theMIT Center for Cell Decision
Processes (NIHgrant P50-GM68762; D.A.L., P.K.S.), the MIT
Biotechnology ProcessEngineering Center (L.G.G.), the MIT Center
for Environmental HealthSciences (NIH grant U19ES011399; L.G.G.), a
NIH/NIGMS BiotechnologyTraining Program Fellowship (NIH grant
T32-GM008334; B.M.K.), and aWhitaker Foundation Graduate Fellowship
(B.D.C.).
Appendix A. Supplementary data
Supplementary data associated with this article can be found,
inthe online version, at doi:10.1016/j.taap.2009.04.002.
http://dx.doi.org/doi:10.1016/j.taap.2009.04.002
-
330 B.D. Cosgrove et al. / Toxicology and Applied Pharmacology
237 (2009) 317–330
References
Bellezzo, J.M., Britton, R.S., Bacon, B.R., Fox, E.S., 1996.
LPS-mediated NF-kappa betaactivation in rat Kupffer cells can be
induced independently of CD14. Am. J. Physiol.270, G956–G961.
Bergheim, I., Luyendyk, J.P., Steele, C., Russell, G.K., Guo,
L., Roth, R.A., Arteel, G.E.,2006. Metformin prevents
endotoxin-induced liver injury after partial hepatect-omy. J.
Pharmacol. Exp. Ther. 316, 1053–1061.
Box, G.E.P., Hunter, W.G., Hunter, J.S., 1978. Statistics for
Experimenters. Wiley, NewYork.
Buchweitz, J.P., Ganey, P.E., Bursian, S.J., Roth, R.A., 2002.
Underlying endotoxemiaaugments toxic responses to chlorpromazine:
is there a relationship to drugidiosyncrasy? J. Pharmacol. Exp.
Ther. 300, 460–467.
Clay, K.D., Hanson, J.S., Pope, S.D., Rissmiller, R.W., Purdum
3rd, P.P., Banks, P.M., 2006.Brief communication: severe
hepatotoxicity of telithromycin: three case reportsand literature
review. Ann. Intern. Med. 144, 415–420.
Clayton, T.A., Lindon, J.C., Cloarec, O., Antti, H., Charuel,
C., Hanton, G., Provost, J.P., Le Net,J.L., Baker, D.,Walley, R.J.,
Everett, J.R., Nicholson, J.K., 2006.
Pharmaco-metabonomicphenotyping and personalized drug treatment.
Nature 440, 1073–1077.
Cosgrove, B.D., Cheng, C., Pritchard, J.R., Stolz, D.B.,
Lauffenburger, D.A., Griffith, L.G.,2008. An inducible autocrine
cascade regulates rat hepatocyte proliferation andapoptosis
responses to tumor necrosis factor-alpha. Hepatology 48,
276–288.
Deng, X., Stachlewitz, R.F., Liguori, M.J., Blomme, E.A.,
Waring, J.F., Luyendyk, J.P.,Maddox, J.F., Ganey, P.E., Roth, R.A.,
2006.Modest inflammation enhances diclofenachepatotoxicity in rats:
role of neutrophils and bacterial translocation. J. Pharmacol.Exp.
Ther. 319, 1191–1199.
Deng, X., Lu, J., Lehman-McKeeman, L.D., Malle, E., Crandall,
D.L., Ganey, P.E., Roth, R.A.,2008. p38 mitogen-activated protein
kinase-dependent tumor necrosis factor-alpha-converting enzyme is
important for liver injury in hepatotoxic interactionbetween
lipopolysaccharide and ranitidine. J. Pharmacol. Exp. Ther. 326,
144–152.
Domansky, K., Inman, W., Serdy, J., Griffith, L., 2005. Perfused
microreactors for livertissue engineering. Conf. Proc. IEEE Eng.
Med. Biol. Soc. 7, 7490–7492.
Ganey, P.E., Luyendyk, J.P., Maddox, J.F., Roth, R.A., 2004.
Adverse hepatic drug reactions:inflammatory episodes as consequence
and contributor. Chem. Biol. Interact. 150,35–51.
Geller, D.A., Nussler, A.K., Di Silvio, M., Lowenstein, C.J.,
Shapiro, R.A., Wang, S.C.,Simmons, R.L., Billiar, T.R., 1993.
Cytokines, endotoxin, and glucocorticoids regulatethe expression of
inducible nitric oxide synthase in hepatocytes. Proc. Natl.
Acad.Sci. U. S. A. 90, 522–526.
Gomez-Lechon, M.J., Donato, M.T., Castell, J.V., Jover, R.,
2003. Human hepatocytes as atool for studying toxicity and drug
metabolism. Curr. Drug Metab. 4, 292–312.
Hewitt, N.J., Lechon, M.J., Houston, J.B., Hallifax, D., Brown,
H.S., Maurel, P., Kenna, J.G.,Gustavsson, L., Lohmann, C.,
Skonberg, C., Guillouzo, A., Tuschl, G., Li, A.P., LeCluyse,E.,
Groothuis, G.M., Hengstler, J.G., 2007. Primary hepatocytes:
current under-standing of the regulation of metabolic enzymes and
transporter proteins, andpharmaceutical practice for the use of
hepatocytes in metabolism, enzymeinduction, transporter, clearance,
and hepatotoxicity studies. Drug Metab. Rev. 39,159–234.
Hoek, J.B., Pastorino, J.G., 2002. Ethanol, oxidative stress,
and cytokine-induced liver cellinjury. Alcohol 27, 63–68.
Horn, T.L., O'Brien, T.D., Schook, L.B., Rutherford, M.S., 2000.
Acute hepatotoxicantexposure induces TNFR-mediated hepatic injury
and cytokine/apoptotic geneexpression. Toxicol Sci. 54,
262–273.
Hwa, A.J., Fry, R.C., Sivaraman, A., So, P.T., Samson, L.D.,
Stolz, D.B., Griffith, L.G., 2007. Ratliver sinusoidal endothelial
cells survive without exogenous VEGF in 3D perfusedco-cultures with
hepatocytes. FASEB J. 21, 2564–2579.
Kaplowitz, N., 2001. Drug-induced liver disorders: implications
for drug developmentand regulation. Drug Saf. 24, 483–490.
Kaplowitz, N., 2005. Idiosyncratic drug hepatotoxicity. Nat.
Rev. Drug Discov. 4,489–499.
King, B.M., Tidor, B., in press. MIST: Maximum Information
Spanning Trees forDimension Reduction of Biological Data Sets.
Bioinformatics. doi:10.1093/bioinfor-matics/btp109.
Lee, W.M., 2003. Drug-induced hepatotoxicity. N. Engl. J. Med.
349, 474–485.LeCluyse, E.L., Alexandre, E., Hamilton, G.A.,
Viollon-Abadie, C., Coon, D.J., Jolley, S.,
Richert, L., 2005. Isolation and culture of primary human
hepatocytes. MethodsMol. Biol. 290, 207–229.
Liu, S., Gallo, D.J., Green, A.M., Williams, D.L., Gong, X.,
Shapiro, R.A., Gambotto, A.A.,Humphris, E.L., Vodovotz, Y.,
Billiar, T.R., 2002. Role of toll-like receptors in changesin gene
expression and NF-kappa B activation in mouse hepatocytes
stimulatedwith lipopolysaccharide. Infect. Immun. 70,
3433–3442.
Luedde, T., Trautwein, C., 2006. Intracellular survival pathways
in the liver. Liver Int. 26,1163–1174.
Luyendyk, J.P., Maddox, J.F., Cosma, G.N., Ganey, P.E.,
Cockerell, G.L., Roth, R.A., 2003.Ranitidine treatment during a
modest inflammatory response precipitatesidiosyncrasy-like liver
injury in rats. J. Pharmacol. Exp. Ther. 307, 9–16.
Malhi, H., Gores, G.J., 2008. Cellular and molecular mechanisms
of liver injury.Gastroenterology 134, 1641–1654.
Mari, M., Colell, A., Morales, A., Caballero, F., Moles, A.,
Fernandez, A., Terrones, O.,Basanez, G., Antonsson, B.,
Garcia-Ruiz, C., Fernandez-Checa, J.C., 2008. Mechanismof
mitochondrial glutathione-dependent hepatocellular susceptibility
to TNFdespite NF-kappaB activation. Gastroenterology 134,
1507–1520.
Obach, R.S., Kalgutkar, A.S., Soglia, J.R., Zhao, S.X., 2008.
Can in vitro metabolism-dependent covalent binding data in liver
microsomes distinguish hepatotoxic fromnonhepatotoxic drugs? An
analysis of 18 drugs with consideration of intrinsicclearance and
daily dose. Chem. Res. Toxicol. 21, 1814–1822.
O'Brien, P., Haskins, J.R., 2007. In vitro cytotoxicity
assessment. Methods Mol. Biol. 356,415–425.
Olson, H., Betton, G., Robinson, D., Thomas, K., Monro, A.,
Kolaja, G., Lilly, P., Sanders, J.,Sipes, G., Bracken, W., Dorato,
M., Van Deun, K., Smith, P., Berger, B., Heller, A.,
2000.Concordance of the toxicity of pharmaceuticals in humans and
in animals. Regul.Toxicol. Pharmacol. 32, 56–67.
Ostapowicz, G., Fontana, R.J., Schiodt, F.V., Larson, A.,
Davern, T.J., Han, S.H., McCashland,T.M., Shakil, A.O., Hay, J.E.,
Hynan, L., Crippin, J.S., Blei, A.T., Samuel, G., Reisch, J.,
Lee,W.M., 2002. Results of a prospective study of acute liver
failure at 17 tertiary carecenters in the United States. Ann.
Intern. Med. 137, 947–954.
Peters, T.S., 2005. Do preclinical testing strategies help
predict human hepatotoxicpotentials? Toxicol. Pathol. 33,
146–154.
Schwabe, R.F., Brenner, D.A., 2006. Mechanisms of Liver Injury.
I. TNF-alpha-inducedliver injury: role of IKK, JNK, and ROS
pathways. Am. J. Physiol. Gastrointest. LiverPhysiol. 290,
G583–G589.
Shaw, P.J., Hopfensperger, M.J., Ganey, P.E., Roth, R.A., 2007.
Lipopolysaccharide andtrovafloxacin coexposure in mice causes
idiosyncrasy-like liver injury dependenton tumor necrosis
factor-alpha. Toxicol. Sci. 100, 259–266.
Shaw, P.J., Ganey, P.E., Roth, R.A., 2009. Tumor necrosis factor
alpha is a proximalmediator of synergistic hepatotoxicity from
trovafloxacin/lipopolysaccharidecoexposure. J. Pharmacol. Exp.
Ther. 328, 62–6