Platform Trials and Precision Medicine in Early Oncology Drug - - PowerPoint PPT Presentation

Platform Trials and Precision Medicine in Early Oncology Drug Development – BI Experience

Yihua (Mary) Zhao, Bushi Wang

I (BW) would like to thank my BI colleagues for their contributions to the talk:

• Dr. Mary Zhao
• Dr. Daniela Fischer
• Dr. Frank Fleischer
• Dr. Birgit Gaschler-Markefski
• Dr. Miaomiao Ge
• Dr. Natalja Strelkowa
• Dr. Kathrin Stucke-Straub
• Dr. Wenqiong Xue

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 2

Acknowledgement

• BI experience with basket trials
• BI experience with platform trials
• Go/No-go decision with patient selection biomarkers in basket trials
• Go/No-go decision with continuous biomarker

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 3

Outline

BI Experience with Basket Trials

• Multiplicity and Bias
• “Best” cohort(s) considered for further development
• Homogeneity or Heterogeneity
• Expectation vs. reality
• More factors can contribute to heterogeneity than possibly measurable: different

prevalence of biomarker, prognostic difference, treatment landscape difference, etc.

• If high heterogeneity is expected, how to implement in model?
• Early stopping
• How to facilitate futility/interim analysis?
• Logistics
• Biomarker test turn around time
• Recruitment rate difference and timing of interim and final analysis

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 5

Basket Trials – General Design Considerations

• First (easy solution)

– Consider a single cohort only for Go/No-go decision (although we may have four in real life)

• Example – observed outcomes in dose expansion (basket of multiple

single arm cohorts, assume 25% suffice for Go)

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 6

Basket Trial Example – Go/No-go Decision

Cohort NSCLC CRC Melanoma xxx #Patients 30 30 30 30 #Responders 3 7 9 8

• Obs. ORR

10.0% 23.3% 30.0% 26.6%

• Example revisited (assume 25% as max. suffice for Go)

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 7

Shrinkage Estimators

Cohort NSCLC CRC Melanoma xxx #Patients 30 30 30 30 #Responders 3 7 9 8

• Obs. ORR

10.0% 23.3% 30.0% 26.6% Shrinkage est. 18.3% 23.3% 26.2% 24.7%

Shrinkage estimator based on a prior for ~ 2.0

• Go/No-go criterion

– Should be based on shrinkage estimates (=adjusted) – Overall Go if at least one estimate achieves Go – Increase in correct decision rates due to information borrowing – Shrinkage estimate is far less biased then looking at max. observed ORR

• Clearly define in the presentations to management of Go/No-go

boundaries

– Single or multiple cohorts considered – Shrinkage or observed estimate considered – How is multiplicity addressed regarding

• Time points
• Number of cohorts/indications

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 8

Basket Trial Example - Summary

BI Experience with Platform Trials

• Concept

– Beyond the concept of umbrella trial which focus on one particular cancer type. – Exploration of multiple regimens in multiple tumour indications/settings,

• by including patient cohorts with a variety of immunobiological baseline characteristics
• to better understand how regimen efficacy depends on cancer immunobiology

– Exploration of IO-retreatment after failure of prior IO therapy

• Scope

– all current and future BI IO-combinations

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 10

Scope of Platform Trial Development in BI

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 11

Design with two IO combinations (expandable)

Indication 1: IO + B Indication 1: IO + A Indication 2: IO + A Indication 2: IO + B IO + A IO + B IO + A IO + B

Patients will be able to cross over to any other arm they are eligible for.

IO pre-treated IO naive

Tumour types TBD

Prior IO benefit Primary IO failure

• For a certain tumor type with treatments entering the study at different

time points

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 12

Treatment Assignment without Patient Selection

PD1 + A PD1 + B PD1 + C Etc. Tumor type I

Equal allocation

Need an updated randomization list whenever a new treatment enters the study

• For a certain tumor type with enrichment is desired in a certain arm,

e.g., PD1 + C

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 13

Treatment Assignment with Patient Selection

PD1 + A PD1 + B PD1 + C Etc. Tumor type II BM+

Equal allocation with e.g., 50% in PD1 + C

BM-

Equal allocation with no slots in PD1 + C

Randomization ratio needs to be adjusted in the BM+ stratum whenever a new treatment enters the study

• Possible treatment switch after PD

– multiple treatments are available

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 14

Treatment Assignment after Initial Treatment

PD1 + X PD PD1 + A PD1 + B Etc. PD

Go/No-go Decision with Biomarker

Part III Expansion

(35 pts per cohort)

Part II Combination dose-finding (12-18 pts) Part I Monotherapy dose-finding (6-12 pts)

advanced solid tumour advanced solid tumour Cohort A: Indication A Cohort B: Indication B Cohort C: Indication C

Trial Design investigating an IO combination

The dose-finding will be guided by Bayesian Logistics Regression Model (BLRM) and the final decision will be made by the Safety Monitoring Committee

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 16

• Assumed response rates in the clinical trial protocol:
• All the observed RR will be transformed into the scale that historical

control is ca. 10% RR

• Potential Go requirement is to add 20% on top of the historical control
• Hence the decision boundaries will be 30% for the cohorts A-C

Decision framework for efficacy - Assumptions for ORR

PD-1 mono Presumed combi Transformation Difference for BHM A 10% 30% As is + 20% B 25% 45% Obs RR - 15% + 20% C 20% 40% Obs RR - 10% + 20%

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 17

N=35 per indication expansion cohort, 3 indications BHM estimated ORR of > 30% for Go and ≤ 20% for NoGo

Numbers are based on 100 simulations, therefore they are still approximations and can vary ca. +/-5%

Decision framework for efficacy Cohorts A-C

ORR

Assumed RR for 3 cohorts ( %, %, %) No go Consider Go Negative scenario: (10%, 10%, 10%) (20%, 20%, 20%) 71% 33% 27% 63% 2% 4% Mixed response cohorts (30%, 25%,5%) Nugget scenario (40%, 5%,5%) 17% 0% 61% 11% 22% 89% Positive scenario: (27%, 27% ,27%) (30%, 30%, 30%) 0% 3% 53% 27% 47% 70%

Obs RR ≥ 20%‐≤30% Obs RR < 20% Obs RR > 30%

Decision probabilites in the negative and positive scenario Red: wrong decision Green: correct decision

• Cohorts A-C, Selection biomarker probably PD1-driven

Decision framework for biomarkers - Patient populations & expected prevalences

Prevalence

(exp)

#patients

per cohort (total)

RR

(exp)

#responders

per cohort (total)

Overall trial population 100% 35 (105) 30% 11 (33) BMX+ 30% 11 (33) 70% 8 (24) BMX- 70% 24 (72) 13% 3 (9)

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 19

• Our hypothesis is that Biomarker X is associated with response across

cohorts

• If the biomarker is predictive/prognostic of clinical response, it is expected to

work across cohorts

• Borrowing of information across cohorts possible
• Bayesian Hierarchical Model with the Biomarker X as covariate
• Association is indicated if regression coefficient for Biomarker X ≠ 0 with

sufficient posterior probability

• If association is indicated -> Biomarker X may be used prospectively in

Phase II

Bayesian Hierarchical Model (BHM)

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 20

• The decision is based on the model slope parameter of BHM
• The slope parameter estimate is provided as posterior probability

distribution

• The association between clinical response and biomarker is concluded if

the posterior probability of the slope parameter in BHM is located above zero with the probability 97.5%

The cohort heterogeneity is expressed via the parameter τ. Its distribution is assumed to be half-normal with zero mean and standard deviation 2. This setting corresponds to large heterogeneity between biomarker subgroups and indications. The prior for the slope parameter beta is set to 2 for the binary case.

Stats - Illustration of decision criteria

PD SD+PR logit(response rates) biomarker biomarker PD SD+PR logit(response rates) biomarker biomarker Posterior dist. for model parameter Average difference in response rates Posterior dist. for model parameter Average difference in response rates

N=35 per indication, 11 assumed BM+, 3 indications - overall N=105 Decision rule: The association between clinical response and biomarker is concluded if the posterior probability of the slope paramater in BHM is located above zero with the probability 97.5%

Note: Nugget and Mixed association scenarios are more difficult to differentiate from random fluctuations compared to clean scenarios, but we still have an acceptable probability to conclude association in extreme cases. Numbers are based on 100 simulations*

Operating characteristics – Bayesian hierarchical model with a covariate for BMX+ for cohorts A-C

Assumed RR for 3 cohorts (BM-,BM+) (BM-,BM+) (BM-,BM+) Prob to conclude no association Prob to conclude association No association scenario: (30%,30%) (30%,30%) (30%,30%) 97% 3% Mixed association cohorts (10%,60%) (20%,35%) (5%,5%) Nugget scenario (20%,85%) (5%,5%) (5%,5%), 23% 6% 77% 94% Strong association in all cohorts: (13%,70%) (13%,70%) (13%,70%) 0% 100%

• Given the study design, sample size and the assumption that biomarker

is associated with response across cohorts A-C, it is possible to detect the association with high probability

• On the other hand, if there is no association, the decision framework

suggests to stop investigation of the biomarker with high probability

• Remarks:

– We are investigating association between the biomarker and the clinical response – If association is detected, the biomarker could be only prognostic, then it is not a suitable patient selection biomarker – Randomized Phase II study needed to investigate if the biomarker is suited for patient selection

Stats - Conclusion

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 23

Go/No-go Decision with Continuous Biomarker

• There is a trade off between marker positive size and efficacy signal;
• Objectives:

– Proof of Clinical Principle that treatment is efficacious in at least a subset of the patient population – Determine a biomarker cut-off for further evaluation

• Literature

– Liu et al. (2016). Thresholding of a continuous companion diagnostic test confident of efficacy in targeted population. Statistics in Biopharmaceutical Research, 8(3): 325-333. – Jiang, Freidlin and Simon (2007). Biomarker-adaptive threshold design: A procedure for evaluating treatment with possible biomarker-defined subset effect. Journal of the National Cancer Institute, 99, 1036-1043.

Size of marker positive subset Strength of efficacy

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 25

Continuous Patient Selection Biomarker

• Proof of clinical principle is

most important for early

• ncology development,

therefore biomarker positive only or enriched trial is of interest.

• Enrolment is challenged

without a cut-off for the continuous biomarker.

• A wrong cut-off may

lead to recruitment failure for a marker+

• nly trial
• All-comer enrolment

may fail to prove clinical principle if the potential marker- positive population is small and under- represented.

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 26

Adaptive Enrollment

Adaptive Enrollment Illustration

12 47 4 56 77 8 32 89 95 113 56 65 38 77 108 66 63 112 142 10 135 76 120 83 54 34 59 35 86 62 135 29 27 136 155 48 6 23 16 48 27 66 31 124 147 111 109 75 44 20 29 82 28 32 37 28 20 44 50 29 41 16 57 7 124 48 40 69 209 115 65 45 1 44 75 60 20 76 116 114 44 30 34 117 71 63 122 17 3 102

• Figure to the right: density function of biomarker

distribution and histogram of sample from all-comer enrollment.

• Figure below: density and histogram of sample

from adaptive enrollment.

Platform Trials and Precision Medicinedate, 5/30/2018, Bushi Wang and Yihua Zhao 27

Presentation title, date, author 28

Questions?