1. Oncology Biostatistics
Design strategies to assess benefit for biomarker sub-populations in Phase III clinical trials
Bharani Dharan, Ekkehard Glimm
Joint Statistical Meeting, Denver
July 31, 2019

2. Outline Background Factors impacting biomarker study designs
Types of biomarker designs
Simulation Results
Summary

3. Background: Biomarker based Phase III designs
A confirmatory trial with a biomarker subpopulation
Multiple populations of interest : All patients, Biomarker positive (B+), Biomarker negative (B-)
Multiple statistical comparisons involved due to multiple populations
Overall type 1 error rate needs to be controlled at one-sided 2.5% level
Role of B+ subgroup is key to assess any differential effect between the biomarker subgroups
Freidlin B, Sun Z, Gray R, et al: Phase III clinical trials that integrate treatment and biomarker evaluation. JCO 31: , 2013
Goteti S, Hirawat S, Massacesi C, Dharan B. Some practical considerations for phase III studies with biomarker evaluations. JCO 32:854–855, 2014

4. Factors that can impact the study design
Biomarker prevalence, targeted effect, biomarker assay
Prevalence of the biomarker subgroup
If B+ comprises a higher percentage of patient population, it is expected that treatment effect in “all patients” will be influenced by treatment effect in B+
Targeted treatment effect in biomarker subgroups and all patients
A conservative targeted effect can result in over-powering of the endpoints.
Assay to determine the biomarker status
Treatment effect needs to be demonstrated by CDx assay in addition to clinical trial assay
Assay sensitivity and specificity is crucial
Minimize unknown status or randomize only the known biomarker status patients

5. Biomarker based Phase III designs
Traditional All patients design
Flexible efficient testing procedures e.g. alpha split approach tailored to match the study objective
Treatment
Enrollment in All patients
(stratified by Biomarker status)
Primary analysis (All, B+)
Control

6. Biomarker based Phase III designs
Biomarker Stratified design
Allows to ‘independently’ test the biomarker positive and negative subpopulations
Treatment
Primary analysis (B+)
B+ Randomization
Control
Biomarker status determination
Treatment
Primary analysis (B-)
B- Randomization
Control
SCREENING
Treatment & Follow up

7. Biomarker based Phase III designs
Adaptive enrichment design
Allows selection of the population (All patients or biomarker positive) based on pre-specified decision rules at the interim analysis
Stop for futility
Treatment
Enrollment in All patients (stratified by biomarker status)
Pre-specified decision rules at interim
Continue in
Full population
Control
Stop recruitment in B- and Enrich B+ subpopulation

8. Traditional All patient design:
Simulation set-up
Study enrolled ‘all patients’ which include biomarker positive and negative sub-populations
Primary endpoints: Progression-free survival (PFS) in All patients and biomarker positive
Key secondary endpoints: Overall survival (OS) in All patients and biomarker positive
Gatekeeping approach to test the multiple endpoints
Assumptions:
Target hazard ratio for PFS: 0.67 in all patients and 0.60 in B+ sub-population
Target hazard ratio for OS: 0.75 in all patients and in B+ sub-population
Accrual rate: 30 patients/month with ~40% B+

9. Traditional design: Statistical testing approach
Alpha split approach tailored to the hypotheses of interest
Several testing strategies are explored
Scenario 1: More alpha for overall and less alpha for biomarker +: 2%, 0.5%
Scenario 2: Equal alpha split between overall and biomarker +: 1.25%,1.25%
Scenario 3: Less alpha for overall and more alpha for biomarker +: 0.5%, 2% .
Scenario 1:
: Endpoint meets success : Endpoint fails to meet success

10. Statistical testing strategy
Alpha split approach tailored to the hypotheses of interest
Scenario 2
Scenario 3

11. Statistical impact of the three testing strategies
Scenario 1 has the optimal power for both OS and PFS; Scenario 3 is optimal from sample size;
PFS adequately powered for all scenarios
Lower sample size for Scenario 3
90% power for OS for “All patients” in Scenario 1
Scenario 1: More alpha for All patients and less alpha for B+: 2%, 0.5%
Scenario 2: Equal alpha split between All patients and B+: 1.25%,1.25%
Scenario 3: Less alpha for All patients and more alpha for B+: 0.5%, 2%

12. Testing strategy for biomarker negative
Decision rule for B-
Biomarker negative sub-population will be assessed to determine whether the treatment effect in All patients is driven solely by biomarker positive subpopulation or by both positive and negative subpopulation
Should there an alpha allocated to test the biomarker negative? or
Will a pre-defined clinically relevant threshold based on a decision rule suffice?
E.g. Hazard ratio <0.67 and probability(HR<1)>97.5%

13. Stratified design Feasible option when either B+ or B- is of interest
Treatment
Primary analysis (B +) B+
1:1
Control
Treatment
Primary analysis (B -) B-
1:1
Control
A feasible option when either B+ or B- is of interest
Ensures unknown status doesn’t confound the interpretation
Better control of prevalence rate and can avoid over-powering resulting from traditional design
Gatekeeping approach required for testing B+ and B– subgroups

14. Adaptive enrichment study design
Adaptation decision on enrichment at interim analysis
Stop for futility
Patient Population:
(stratified by biomarker status: + and -)
Treatment
Pre-specified decision rules at interim analysis
Continue in
All patients population
Control
Stop recruitment on B- and Enrich B+ subpopulation

15. Adaptive enrichment design: Decision Rules at Interim analysis Bayesian predictive power (PP) & Posterior probability for futility
Decision Rules
PPAll < x & PPB+ < y
Stop for futility
PPAll ≥ x & PPB+ < y
Continue in All patients population
Interim analysis
PPAll ≥ x & PPB+ ≥ y & rule in B-* not met
PPAll < x & PPB+ ≥ y
Stop recruitment on B- and Enrich B+ subpopulation
PPAll ≥ x & PPB+ ≥ y & rule in B-* met
PPAll: probability to meet success in either (or both) PFS in the full population or PFS in the B+ subpopulation at final analysis given interim data when study continued in full population
PPB+: probability to meet success in PFS in the B+ subpopulation given interim data when adapted to B+ subpopulation
*rule in the B- (Posterior probability P(HR>HR1)>z )

16. Adaptive enrichment design: Methodology
Multiplicity issues
Testing in All patients & B+ populations
Two looks; Interim -Adaptation, Futility; Final - Efficacy analysis
Closed testing procedure
Hochberg p-value qi for intersection hypothesis H0F ∩ H0S at stage i
qi=min{2 min(piS, piF), max(piS, piF)}
At final analysis
If trial continues in full population
Tests in F and S of equal interest (H0F & H0S co-primary)
Elementary hypotheses at stage 2:
Reject H0F ∩ H0S if C(q1, q2)≥c2 then:
reject H0F if C(p1F, p2F) ≥c2 & reject H0S if C(p1S, p2S) ≥c2
If trials is adapted to the B+ subpopulation
Reject H0F ∩ H0S if C(q1, p2S)≥c2 then
Reject H0S if C(p1S, p2S) ≥c2
*F refers to “All patients” & S refers to biomarker positive subpopulation

17. Adaptive enrichment design: Methodology*
Inverse normal method
p1 is the p-value based on log-rank z-statistic Z1 from stage-1.
p2 is obtained from the difference in log-rank statistics z2-z1, where z2 is the log-rank z-statistic based on the cumulative data at stage-2.
Independent p-values from 2 stages combined: inverse normal method (Lehmacher & Wassmer 1999, Cui, Hung & Wang 1999)
Pre-specified fixed weights based on “All patients” information fractions:
𝑤1= (𝑛1/𝑁) 1/2 𝑤2= (𝑛2/𝑁) 1/2 , 𝑤1 2 + 𝑤2 2 =1, 𝑛𝑖/𝑁 event fraction at stage i
*Brannath W et al. (2009) Confirmatory adaptive designs with Bayesian decision tools for a targeted therapy in oncology, Stats. Med; 28:
𝐶 𝑝1,𝑝2 =𝑤1ф −1 (1−𝑝1)+𝑤2 ф −1 (1−𝑝2)

18. Adaptive enrichment design Simulation set-up
Prevalence of biomarker positive
30%
40%
50%
Posterior probability in the biomarker negative subpopulation:
Probability cutoff: 0.75
Hazard Ratio cutoff: 0.80
Predictive power cutoff (Full population, biomarker positive subpopulation): (0.35, 0.35)
Hazard ratio scenarios for biomarker positive and negative
Sample size of 524 patients; Interim analysis at 125 PFS events; Final analysis at 374 events in the All patients population and 281 events if enriched to the B+ subpopulation

19. Adaptation decision at Interim analyses: Simulation results for different prevalence rates
Hazard ratio
Probability to B+ B-
Stop for futility (%)
Go with All (%)
Go with B+(%)
30%
40%
50%
0.5
0.4
0.2
100
99.4
99.3
0.8 3 2 72 74 75 25 24 23
1.0 6 4 33 38 42 61 58 56
0.67 7 88 87 85 8 13 12 66 67 21 20 19 16 29 31 36 50 48 68 59 63 71 14 18 15 19

20. Overall trial power: Simulation results for different prevalence rates
Hazard ratio
Unconditional Power B+ B-
Overall
power
Prob. success in All patients population
Prob. Success in B+ sub-population
30%
40%
50%
0.5
100.0
99.6
99.8
99.7
99.4
99.2
95.4
98.0
0.8
94.7
97.5
68.7
72.4
74.8
92.3
96.7
97.9
1.0
91.2 95
97.6
21.2
32.0
39.4
91.1
94.9
0.67
89.2
89.5
89.8
84.0
83.8
82.4
57.1
67.4
76.1
72.3
76.9
79.4
51.5
57.3
59.4
57.5
67.2
73.4
62.0
66.4
72.2
10.3
15.0
23.2
60.9
65.2
71.4 20

21. Conditional Power: Simulation results for different prevalence rates
Hazard ratio
Conditional Power B+ B-
Prob. Success / go All
Prob. Success in All /go All
Prob. Success in B+ /go All
Prob. Success in B+/go B+
30%
40%
50%
0.5
100.0
100
99.9
95.4
98.4
99.8
0.8
97.2
99.3
99.7
94.9
99.5
93.9
98.2
1.0
91.5
96.9
99.4
64.3
84.7
95.0
91.1
96.6
0.67
95.7
96.4
96.5
95.6
96.3
59.1
70.9
80.4
89.1
91.6
94.6
80.7
86.4
78.2
85.0
88.1
58.2
71.9
80.3
93.4
56.1
66.1
78.1
35.0
47.8
64.7
52.1
62.4
75.8
91.3
91.4
92.3 21

22. Hazard Ratio plot No treatment effect (HR=1)in both B+ and B- sub-populations
Hazard ratio All population
Hazard ratio biomarker negative
Hazard ratio biomarker positive
Hazard ratio biomarker positive
+ Go B+
X Futile
O Go All
Decision at interim
Probability
Go All patients
15.7%
Go B+
21.2%
Stop for futility
63.1%

23. Hazard Ratio plot Treatment effect (HR=0.67) only in B+ sub-population
Hazard ratio All population
Hazard ratio biomarker negative
Hazard ratio biomarker positive
Hazard ratio biomarker positive
+ Go B+
X Futile
O Go All
Decision at interim
Probability
Go All patients
29.4%
Go B+
49.9%
Stop for futility
20.7%

24. Hazard Ratio plot Treatment effect (HR=0
Hazard Ratio plot Treatment effect (HR=0.67) in both B+ and B- sub-populations
Hazard ratio biomarker negative
Hazard ratio All population
Hazard ratio biomarker positive
Hazard ratio biomarker positive
+ Go B+
X Futile
O Go All
Decision at interim
Probability
Go All patients
87.8%
Go B+
5.9%
Stop for futility
6.3%

25. Adaptive enrichment design: Operational elements
Biomarker prevalence, DMC and enrollment hold
Biomarker prevalence can impact the decisions
Data Monitoring Committee (DMC) should be involved in the decision making for adaptation
Following DMC decision, randomization scheme will need to take into account the impact on randomization and stratification
Investigators should be informed of the possible actions and events in the course of the study which could result from the decision making in this trial
Follow-up of ongoing biomarker negative patients in case the study adapts to biomarker positive sub-population at the time of interim analysis
Enrollment hold during DMC review for adaptation

26. Summary
Prevalence of biomarker groups plays a crucial role in the study design and statistical testing sequence
Should be continuously monitored during the study
Inclusion/exclusion of ‘unknowns’ should be carefully considered at the study design stage
Designs such as gate-keeping procedures, adaptive designs or stratified designs can be used depending on the strategy

27. Acknowledgments Frank Bretz Patrick Urban Nathalie Fretault
Emmanuel Zuber
Sasikiran Goteti
Samit Hirawat
Cristian Massacesi
Willi Maurer
Sylvie Le Mouhaer
Kannan Natarajan
Yuanbo Song

28. References
Bretz F, Maurer W, Brannath W, et al: A graphical approach to sequentially rejective multiple test procedures. Stat Med 28: , 2009
Brannath W, Zuber E et al: Confirmatory adaptive designs with Bayesian decision tools for a targeted therapy in oncology, Stats. Med; 28: , 2009
Freidlin B, Sun Z, Gray R, et al: Phase III clinical trials that integrate treatment and biomarker evaluation. JCO 31: , 2013
Goteti S, Hirawat S, Massacesi C, Fretault N, Bretz F and Dharan B: Some practical considerations for phase III studies with biomarker evaluations. JCO 32:854–855, 2014
Rothmann et al, Testing in a Pre-specified Subgroup and the Intent-to-Treat Population: DIA, , 46(2), 2012