2014-10-22 EUGM | ROYCHAUDHURI | Phase I Combination Trials

Satrajit Roychoudhury

East User Group Meeting 2014

Oct 22nd 2014

A Bayesian Industry Approach to Phase I Combination Trials in Oncology

Beat Neuenschwander

Alessandro Matano

Zhongwen Tang

Simon Wandel

Stuart Bailey

Statistical Methods in Drug Combination Studies, Boca Raton, FL: Chapman & Hall/CRC Press (2015)

Book Chapter and Authors

Acknowledgements

2 | East User Group Meeting 2014 | Satrajit Roychoudhury | Oct 22, 2014 | Phase I Combination Trial

Table of Content

1. Introduction

2. Phase I Design Framework

3. Methodology

4. Applications

5. Implementation Issues

6. Conclusion


1. Introduction

In Oncology, challenges in Phase I trials are many: while keeping patient safety within acceptable limits, the trials should be small, adaptive, and enable a quick declaration of the maximum tolerable dose (MTD) and/or recommended phase II dose (RP2D).

The objective of this chapter is to provide a comprehensive overview, which includes;

• a rationale based on general clinical and statistical considerations,

• a summary of the methodological components, and

• applications


1. Introduction Challenges and Design Requirements

Phase I Trial Challenges Design Requirements

Untested drug or combination in treatment-resistant patients

Escalating dose cohorts with small number of patients (e.g. 3-6)

Primary objective: determine Maximum Tolerated Dose (MTD)

Accurately estimate MTD

High toxicity potential: safety first Robustly avoid toxic doses (“overdosing”)

Most responses occur 80%-120% of MTD

Avoid sub therapeutic doses while controlling overdosing

Find best dose for dose expansion (which generally becomes the recommended phase II dose)

Enroll more patients at acceptable, active doses (flexible cohort sizes)

Complete trial in timely fashion Use available information efficiently


2. Phase I design framework Traditional 3+3 design

Trial Data

0/3,0/3,1/3,...

Dose escalation

decision from

predefined

algorithm

Very simple BUT

• Poor targeting of true MTD

• Highly variable estimates

High attrition rates in phase II and III

• How many failures are due to the wrong dose coming

from phase I?


2. Phase I design framework Why Bayesian modeling?

Bayesian approach offers a way to quantify our knowledge and assess risk.

• Given what I have seen, what can I say about the true risk of dose limiting toxicity (DLT)?

Models allow us to share information between doses – efficient use of data.

Flexibility to include additional patients, unplanned doses, change schedules.

Higher chance to find the correct MTD (Rogatko et al. (2007)).


2. Phase I design framework Clinically driven, statistically supported decisions

DLT rates

p1, p2,...,pMTD,...

(uncertainty!)

Historical

Data

(prior info)

Model based

dose-DLT

relationship

Trial Data

0/3,0/3,1/3,...

Clinical

Expertise

Dose

recommen-

dations

Decisions Dose Escalation

Decision

Model Inference Decision/Policy

Responsible: Statistician Responsible: Investigators/Clinician

Informing: Clinician (Prior, DLT) Informing: Statistician (risk)


3. Methodology Single Agent Model (Neuenschwander et. al. 2008)

Data: (#DLT/#Patients): 𝑟𝑑~ Binomial(π𝑑, 𝑛𝑑)

Parameter Model: logit 𝜋𝑑 = log(α) + β log(d/d*)

Prior: (log(α), log(β)) ~ N2(m1, m2, s1, s2, corr)

Model parameter α and β can be interpreted as;

α is the odds, of a DLT at d*, an arbitrary scaling dose.

β > 0 is the increase in the log-odds of a DLT by a unit increase in log-dose.

Can be extended by adding additional covariates/relevant predictors representing different patient-strata.


3. Methodology Developing drug combinations: Paradigm change from single agent

There is no longer one MTD but a many

- e.g. given drug A = X mg, MTD of drug B is Y mg.

- Critical to determine the MTD boundary and the set of acceptable doses.

Still in “learning phase”

- Flexibility (e.g., variable cohort sizes, split cohorts).

- Bayesian model summarizes knowledge on all dose pairs, sets upper bound on dosing.

- Actual decisions use additional information (e.g. efficacy, PK, biomarkers, later cycle AE) to select “best” dose pair(s) for next cohort.


3. Methodology Interaction: Antagonism, independence or synergy

There are a number of factors that may lead to increased or decreased risk of toxicity;

• Drug-Drug Interaction seen in the PK profiles, e.g.

- AUC or Cmax for one or both compounds is significantly increased/decreased compared to being given alone.

- May be competitively using/inhibiting CYP enzyme causing lower clearance.

• Overlapping toxicities, e.g.

- If drug 1 and drug 2 have similar toxicity profiles, giving them together may cause a significant increase in the observed rate.

- Equally, the combination may cause the same toxicity in the same subset of patients but at a higher grade, whilst other patients are unaffected.

- Two toxicities that, when they occur together increase the risk of DLT in a patient.

• Pathway interactions, feedback loops that either trigger additional toxicities or provide protective effects.


3. Methodology Dual and Triple Combination Model

Proposed model has two parts: (multiplicative odds model)

• Marginal effects: 2 parameters per agent representing single agent toxicities. • Interaction: 1 parameter for Dual interaction (Dual combo model). : 4 parameters for Dual and triple(extra) interaction (Triple combo model).

Extension of Thall et. al. Biometrics (2003).

Properties for proposed combination models

• Parsimonious since the number of tested dose combinations in phase I trials is usually fairly small.

• Have easily interpretable parameters similar to Base model.

• Allow for interaction.

• Have the ability to incorporate single-agent information.


3. Methodology Metrics for Dosing Recommendations

Posterior Distribution: The result of the Bayesian analysis is the posterior distribution of which, in turn , leads to the posterior of for each dose group. Our approach relies formally on this metric (using principle of escalation with overdose control (EWOC)) for suggesting feasible dose for next cohort.

Predictive Distribution: The predictive distribution of the number of DLT “r*” in the next cohort with size “n”. This metric is used informally in discussions with clinical teams.

A Formal Decision Analysis: We do not use formal decision analysis to identify the dose range for further dosing or the MTD/RP2D.


3. Methodology Choice of Prior: Using Contextual Information or Co-data

Key question: Can we use available information as prior?

Naive views

• The information is of no value because we simply don‘t know what‘s going to happen in our trial

• Or, transform these data into a prior that is worth N patients

• Eg. N=200, with observed rate = 0.10. Use a Beta(20,180) prior.

Between-trial heterogeneity discounting historical data

• After discounting, prior may be Beta(2,18)

• Don‘t use 10% as a rule of thumb for discounting!

Methodology: meta-analytic-predictive (MAP) • Random-effects meta-analysis of historical data/parameters 1,... H

• with a prediction for the parameter * in the new trial

• basic assumption: exchangeable parameters


3. Methodology Meta-analytic-predictive (MAP) Prior

Similarity Scenario

H Historical Trials with Effects 1,..., H

Data from Historical Trials Yh h=1,...,H

Similarity of Parameters 1, ..., H, * ~ N(µ,2)

Meta-Analytic-Predictive (MAP) Prior * | Y1,..., YH

Challenge: how large is ? (cannot be inferred if for small H)


3. Methodology MAP Prior for Oncology Phase I

Different types of a-priori information in Phase I

• Animal data (2 species, e.g. dog and rat) Requires discounting of information due to between-species

extrapolation.

• Historical data for same compound in another indication (with similar toxicity profile), one or more trials. Requires moderate to substantial discounting of information due to between-trial/indication extrapolation.

Typically these priors have relatively little weight, 3 to 9 pts.

• Compared to phase I sample sizes of 15 to 30 patients, this is still relevant information.


3. Methodology MAP Priors for Phase I Combination Studies

Prior for single-agent parameters

• Bivariate normal prior: from co-data (historical/external data), with appropriate discounting due to between-trial variability using MAP.

Prior for interaction parameter (odds-multipliers) ’s

• Normal (no restriction) or log-normal (synergistic toxicity)

• Pre-clinical/Clinical information can be incorporated in the prior of interaction parameter. But often no information is available. Therefore weakly-informative priors are used.


3. Methodology Robust Mixture Prior

The use of historical data requires an extrapolation from the past to the future. Predictions are always difficult.

What if there is a possibility that the chosen prior will be in conflict with the actually observed data?

Possible robustification: MAP Mixture Priors

• Adding a weakly informative component to MAP may address the issue.

• Mixture weights represent the degree of similarity between codata and new trial.

• Mixture priors are more robust against prior-data conflict, and should be used more often (rarely seen in practice).


4. Applications Practical consideration for Implementation

What is our starting combination?

• Generally determined ad-hoc... 50% of each MTD?

• What about interaction? Uncertainty?

No evidence of synergistic or antagonistic interaction between the two compounds

• But don’t want to rule it out

• Normal distribution for interaction

parameter:

- Antagonism

- Independence

- Synergy


4. Applications Dual-Combination Trial: Background

Combination of two new compounds

• Compound 1: d1 = 3, 4.5, 6, 8; d1* = 3 (all mg)

• Compound 2: d2 = 33.3, 50, 100, 200, 400, 800, 1120; d2* = 960 (all mg)

• Dose-limiting data (DLT): days 1 - 28

• Toxicity intervals for 𝜋𝑑: [0-0.16), [0.16-0.35), [0.35-1.00]

• EWOC: P(𝜋𝑑 >= 0.35) < 0.25

Actual dosing decisions

• must respect EWOC criterion.

• should also acknowledge other relevant information (PK, efficacy, ...).

Additional design constraints

• Cohort size 3 to 6

• Only one compound's dose can be increased; max. increment of 100%.


4. Applications Rules to determine MTD

Dose escalations continue until declaration of the MTD.

A dose is potential candidate for MTD when;

1. At least 6 patients have been treated at this dose

2. This dose satisfies one of the following conditions:

- The probability of targeted toxicity exceeds 50%:

- or, a minimum of number of 15 patients have already been treated in the trial.


4. Applications Dual-Combination Trial: Prior derivation

Relevant (single-agent) data was identified

MAP approach to derive priors based on this data: discussion with clinical team about similarity

No a-priori evidence for interaction between the two compounds, but considerable uncertainty


4. Applications Dual-Combination Trial: Full history (1/2)


4. Applications Dual-Combination Trial: Full history (2/2)

Cohort 1

3 mg / 400mg

0 DLT / 3 Patients

Cohort 2a

3 mg / 800mg

1 DLT / 3 Patients

Cohort 2b

6 mg / 400mg

1 DLT / 3 Patients

Cohort 3a

3 mg / 800mg

1 DLT / 3 Patients

... Cohort 3b

6 mg / 400mg

0 DLT / 3 Patients


5. Applications MTD Declaration

3-800: grade 2 adverse events, moderate increase in magnitude and duration of QT.

4.5-600: experienced grade 1 and/or grade 2 adverse events, No QT prolongation. BUT none of the patients had a relevant tumor lesion decrease.

6-400: experienced grade 2 adverse events, but easily managed with appropriate dose reductions, No QT prolongation. Moreover, encouraging efficacy signals were observed.

Based on clinical and statistical findings : 6-400 mg was declared as the MTD and RP2D.

| East User Group Meeting 2014 | Satrajit Roychoudhury | Oct 22, 2014 | Phase I Combination Trial 25

4. Applications Triple-Combination Trial: Background

Triple combination

• One «backbone», two experimental

• Agent 1: 100, 200, 300, 400 mg

• Agent 2: 10, 20, 30, 40 mg

• Agent 3: 250 mg (fixed)

Stage-wise approach considered reasonable

• Find MTD for agent 2 + 3

• Escalation of agents 1 and 2 (agent 3 fixed)

Prior derivation similar to dual combination case.


4. Applications Triple-Combination Trial: Model assessment (1/3)

Stage I: Data scenarios for Dual Combination



Stage II: Data scenarios for Triple Combination



Operating characteristics: Assumed true toxicity rates

Metrics


5. Implementation issues Practical guide

Some of the points that need to be considered when implementing the proposed Bayesian adaptive design in clinical practice.

Discussion with clinical colleagues

Expertise in Bayesian Statistics

Computations

Study Protocols

Review Boards


5. Implementation issues Discussion with clinical colleagues

Discussing important features with clinical colleagues: key responsibilities of a statistician when designing and implementing clinical trials.

Recommended to explain the main features of the approach using visual illustrations, non-statistical language, and common sense.

A good preparation includes a mimicked dose-escalation meeting using data scenarios to highlight various possible outcomes.


5. Implementation issues Expertise in Bayesian Statistics and Computation

Expertise in Bayesian Statistics: Implementing the proposed approach requires:

• Technical skills related to the concrete implementation

• the ability to explain in simple language

• Training/workshop/case studies for clinician and statistician

• project-related regular interactions between statisticians and clinicians

Computation: Require a tool that allows to extract the relevant metrics for inference and decision making:

• WinBUGS/R

• JAGS

• Stan


5. Implementation issues Study Protocols

Requires more attention when Bayesian methods are used.

A clear description of the model and prior needs to be provided in protocol.

If an informative prior based on external information is used, the derivation of the prior should be clear in protocol.

The actual process for dose-escalations needs to be described as well.

Further technical details including data scenarios and operating characteristic may be covered in a statistical appendix.


5. Implementation Issues Review Boards

Experience shows that discussions with HA and IRB/IEC around these issues lead to a better understanding and appreciation of the rationale and intent of the proposed approach.

“Typical” questions include;

• A 25% risk of overdose is too high.

• The design allows too many patients to be exposed at once to a new dose level.

• The design must be set up in a way that observation of 2 DLT at a dose level means that this dose cannot be used again.

• The design makes a recommendation, but the clinician decides the dose. This implies that a decision may overrule the original recommendation, and patients may be dosed at unsafe dose levels.


6. Conclusion

Detailed explanation of phase I (combination) approach

• Concepts

• Mathematical background

• Examples

• Practical implementation

Covers about 10 years of practical expertise

• Methodological work

• Operational experience (> 100 trials)

• Technical implementation

Standard at Novartis Oncology (others just taking up)


References (1/2)

[1] Neuenschwander, Matano, Tang, Roychoudhury, Wandeland Bailey. A Bayesian Industry Approach to Phase I Combination Trials in Oncology. Statistical Methods in Drug Combination Studies, Boca Raton, FL: Chapman & Hall/CRC Press 2015.

[2] Neuenschwander, Branson, Gsponer. Critical aspects of the Bayesian approach to phase I cancer trials. StatMed 2008.

[3] Rogatko, Schoeneck, Jonas et al. Translation of innovative designs into phase I trials. JCO 2007.


References (2/2)

[4] Babb, Rogatko, Zacks. Cancer phase I clinical trials: efficient dose escalation with overdose control. Statistics in Medicine 1998.

[5] Thall, Lee. Practical model-based dose-finding in phase I clinical trials: methods based on toxicity. Int J Gynecol Cancer 2003.

[6] Thall, Millikan, Mueller, Lee. Dose-finding with two agents in phase I oncology trials. Biometrics 2003.

[7] O’Quigley, Pepe, Fisher. Continual reassessment method: a practical design for phase I clinical trials in cancer. Biometrics 1990.


2014-10-22 EUGM | ROYCHAUDHURI | Phase I Combination Trials

Health & Medicine

recommended phase

design requirements

design trial data

dose cohorts

best dose

dose expansion

dose rp2d

wrong dose