1. Cyrus R. Mehta, Ph.D Cytel Inc. and Harvard University
Adaptive Population Enrichment for Oncology Trials with Time to Event Endpoints
Cyrus R. Mehta, Ph.D
Cytel Inc. and Harvard University

2. Adaptive enrichment designs in oncology Motivation
Outline of Talk
Adaptive enrichment designs in oncology
Motivation
Statistical methodology
Simulation guided operating characteristics
Future directions
August 7, 2014
JSM Boston

3. Oncology Products Approved in the US for Selected Patient Populations
Compound/Target
Indication (prevalence target)
Crizotinib (Xalkori®)/ ALK-rearrangement
Non-small cell lung cancer with ALK-rearrangements (5%)
Vemurafenib (Zelboraf®)/ BRAF mutation
Advanced melanoma with mutant BRAF (30-40%)
Trametinib (Mekinist™)/ MEK
Trastuzumab (Herceptin®); Lapatinib (Tykerb®)/ Her2
Her2 expressing breast cancer (25%)
Her2 expressing metastatic gastric cancer (20-30%)
Aromatase inhibitors (letrozole, exemestane)
ER(+) breast cancer (60-70%)
Rituximab (Rituxan®)/ CD20
CD20(+) B-cell lymphomas (90%+)
Cetuximab (Erbitux®); Panitumumab (Vectibix®) / EGFR
Advanced Head/neck cancer (~100%)
EGFR(+) metastatic colorectal cancer (60-80%)
KRASWT metastatic colorectal cancer (60%)
August 7, 2014
JSM Boston

4. Dilemma in Phase 3 Oncology Trials
High phase 3 failure rate in broad population but treatment is effective in subgroups
In many cases, responsive subgroups were identified retrospectively
Desirable to launch phase 3 trials of targeted agents in biomarker-defined subgroups
Risky to do so without empirical evidence of biomarker predictivity
Conclusion: First launch a phase 2 POC trial
August 7, 2014
JSM Boston

5. Design Considerations for phase 2 POC
Strength of preclinical evidence
Prevalence of the marker
Reproducibility and validity of assays
Time-to-event endpoint (PFS or OS)
Sample size limitations ( )
Utilize adaptive enrichment design
August 7, 2014
JSM Boston

6. Adaptive Enrichment Design: Notation
𝑺 and 𝑺 are biomarker subgroups
n denotes sample size
d denotes events
T denotes the logrank statistic
A priori biological and PK basis for large effects in 𝑺 and small effects in 𝑺
August 7, 2014
JSM Boston

7. Schematic Representation of Design
Subgroup 𝑆
.5 Treatment
.5 Control
𝑛 0 patients
Stop for Futility
𝑑0 events
ALL COMERS S T
RAT I F Y
INTERIM ANALYSIS
Continue with
S and 𝑆
FINAL ANALYSIS
Perform a closed test of S
Subgroup 𝑆
.5 Treatment
.5 Control
Continue with
S only
patients
events
If is dropped, randomize all remaining patients to subgroup S and increase its events
August 7, 2014
JSM Boston

8. Important Features of Design
Ensures that both 𝑺 and 𝑺 are adequately represented at the interim analysis
In keeping with FDA Guidance on Enrichment Designs (2012, Figure 5 A)
Makes better use of limited sample size by re-directing enrolment to 𝑺 if biomarker is predictive
August 7, 2014
JSM Boston

9. Time Line of S Subgroup 𝑛 0 𝑛 𝑠 𝑛 𝑠 𝑑 0 𝑑 𝑠 𝑑 𝑠 𝑇 0 𝑇 𝑠 𝑇 𝑠 𝑆 ′ cohort
𝑛 𝑠
𝑑 0
𝑑 𝑠
𝑑 𝑠
𝑇 0
𝑇 𝑠
𝑇 𝑠
Time Axis
𝑆 ′ cohort
𝑆 ′′ cohort
Interim
Analysis
Planned
Final
Analysis
Actual
Final
Analysis
August 7, 2014
JSM Boston

10. Time Line of 𝑺 Subgroup 𝑛 0 𝑛 𝑠 𝑑 0 𝑑 𝑠 𝑇 0 𝑇 𝑠 Drop the Subgroup
𝑛 0
𝑛 𝑠
𝑑 0
𝑑 𝑠
𝑇 0
𝑇 𝑠
Drop the Subgroup
if it has low conditional power 𝑺
Time Axis
𝑆 ′ cohort
𝑆 ′′ cohort
Interim
Analysis
Planned
Final
Analysis
August 7, 2014
JSM Boston

11. For closed testing both 𝑯 𝑺 and 𝑯 𝑺 ∩ 𝑯 𝑺 must be rejected
Hypothesis Testing at Final Analysis (a) If you do not drop 𝑺 at interim
R: rejection region for the
Intersection hypothesis 𝑯 𝑺 ∩ 𝑯 𝑺
T s
2.24
For closed testing both 𝑯 𝑺 and 𝑯 𝑺 ∩ 𝑯 𝑺 must be rejected
1.96
2.24
T s
Rejection region for the elementary hypothesis 𝑯 𝑺
August 7, 2014
JSM Boston

12. Hypothesis Testing at Final Analysis (b) If you do drop 𝑺 at interim
Rejection region for the elementary hypothesis 𝑯 𝑺
Reject if T 𝑺 > 𝒄 𝑐
T s
August 7, 2014
JSM Boston

13. Preserving Type-1 Error: CER Method 1 ( Mullër and Schafër, 2001)
𝑛 0
𝑛 𝑠
𝑛 𝑠
𝑑 0
𝑑 𝑠
𝑑 𝑠
𝑇 0
𝑇 𝑠
𝑇 𝑠
Time Axis
𝑆 ′ cohort
𝑆 ′′ cohort
Interim
Analysis
Planned
Final
Analysis
Actual
Final
Analysis
August 7, 2014
JSM Boston

14. Preserving Type-1 Error: Method 2 (Mehta, Irle, Daniels, Schafër 2014) methodology)
𝑛 0
𝑛 𝑠
𝑛 𝑠
𝑑 0
𝑑 𝑠
𝑑 𝑠
𝑇 0
𝑇 0
𝑇 0
𝑇 𝑠
𝑇′ 0
𝑇 𝑠
𝑇′ 0
Time Axis
𝑆 ′ cohort
𝑆 ′′ cohort
Interim
Analysis
Planned
Final
Analysis
Actual
Final
Analysis
August 7, 2014
JSM Boston

15. Elicit Operating Characteristics by Simulation
Example: metastatic non-small cell lung cancer trial
N = 160 patients with 80 for each subgroup
Primary endpoint is progression free survival (PFS)
Median PFS for control arm is 5 months
Prior evidence that subgroup S (EGFR mutation) is predictive of PFS:
HR(S) = 0.5, 0.6 is plausible
HR( 𝑆 ) < 0.7 is unlikely
August 7, 2014
JSM Boston

16. Decision Rule for Dropping 𝑺 at Interim
Based on conditional power (CP) at the interim look
Drop 𝑆 if CP( 𝑆 ) < A and CP(S) > B
Use phase 2 results for go no-go phase 3 decision
August 7, 2014
JSM Boston

17. Use Phase 2 Simulations to Guide Phase 3 Go/No-Go/Enrich Decisions
Decision rules for initiating a Phase 3 trial based on the results of the Phase 2 adaptive enrichment trial
Phase 2 Outcome:
Decision Rule for Phase 3
Win on S; Lose on 𝑆
Initiate Phase 3 in S only
Win on S; Win on 𝑆
Initiate Phase 3 in S and 𝑆
Lose on S; Win on 𝑆
No Go/investigate
Lose on S; Lose on 𝑆
No Go
August 7, 2014
JSM Boston

18. Drop 𝑺 if CP( 𝑺 ) < 0.5 and CP(S) > 0
Assume HR(S) = 0.5
August 7, 2014
JSM Boston

19. Drop 𝑺 if CP( 𝑺 ) < 0.5 and CP(S) > 0.5
Assume HR(S) = 0.5
August 7, 2014
JSM Boston

20. Never Drop 𝑺 (apply closed test only)
Assume HR(S) = 0.5
August 7, 2014
JSM Boston

21. Incorporating Tumor Response Data into the Adaptive Decision Rules
Generalized logistic regression model for tumor response Y given subgroup Z and treatment t:
Mean of exponential survival function f(t|y, t, Z)
The mixture distribution of t unconditional on Y
Stop accrual to subgroup Z if
August 7, 2014
JSM Boston

22. Comparison with Jenkins et. al. (2011)
Jenkins et. al. (JSJ method):
Combines p-values with pre-specified weights
Test H_F and H_S
Mehta et. al. (MDSI method):
Uses conditional error rate approach
Tests H_S and H_Sbar
Which is more appropriate for targeted therapies in oncology?
August 7, 2014
JSM Boston

23. Comparing JNJ and MDSI for rejecting (H_F, H_S) when H_S=0. 5 and 0
Comparing JNJ and MDSI for rejecting (H_F, H_S) when H_S=0.5 and 0.5≤H_Sbar≤1
August 7, 2014
JSM Boston

24. When to use each approach
When one expects that the biomarker is predictive (i.e., new treatment has greater benefit than control in S but not in S_bar), then it is preferable to test (H_S, H_Sbar).
When one expects that the treatment works equally well in both subgroups then it is preferable to test (H_F, H_S)
August 7, 2014
JSM Boston

25. Concluding Remarks and Future Work
Phase 2 trial produces go/no-go/enrich Phase 3 decision
Simulate Phase 2 trial under scenarios where biomarker is predictive and where it is prognostic
Use simulation results to calibrate the performance of the go/no-go/enrich decision rules
Improve the criteria for dropping 𝑆 or for futility termination at interim analysis:
Utilize the information in the censored observations
Use Bayesian model incorporating tumor response and PFS for sharper criteria
Consider modeling both Phase 2 and Phase 3
August 7, 2014
JSM Boston

26. References and Acknowledgements
Design based on collaborations with the medical oncologists at Pfizer Inc.
Joint research with Sebastien Irle and Helmut Schäfer, University of Marburg, Germany
Expert computational help by Pranab Ghosh
References: Posch and Bauer (2004, SiM); Jenkins, Stone, Jennison (2011, Pharmaceut. Statist); Mehta, Schäfer, Irle (2014, SiM); FDA Guidance on Enrichment Designs (2012)
August 7, 2014
JSM Boston