1.  Adaptive Enrichment Designs for Confirmatory Clinical Trials Specifying the Intended Use Population and Estimating the Treatment Effect
Richard Simon, D.Sc
R Simon Consulting
Noah Simon, Ph.D.
Department of Biostatistics
University of Washington

2. Changed Strategies for Drug Development and Clinical Trial Design
Inter-patient variability of cancers of the same histologic type
Different oncogenic mutations that drive them
Different biologic behavior and response to treatment
Most new cancer treatments are very expensive & only work for a subset of patients with the disease

3. Need to Modify Standard Paradigm of Phase III Clinical Trials
Broad eligibility
Focus on ITT analysis

4. Standard Paradigm May Lead to
False negative studies
Positive studies with:
Large NNT
Small average treatment effects
High economic costs to society

5. Principle of New Paradigm
A phase III trial needs a primary analysis plan that preserves the study-wise type I error level, but that analysis need not be comparison of treatments for the all-randomized population

6. When The Drug Inhibits a Mutated Oncogene
The (targeted) enrichment design is the design of choice

7. Enrichment Design Enrichment design
Develop Predictor of Benefit from New Drug
Using phase II data, develop predictor of response to new drug
Patient Predicted Responsive
Patient Predicted Non-Responsive
Off Study
New Drug
Control
Enrichment Design En
Enrichment design

8. Successful use of targeted enrichment design
Trastuzumab, pertuzumab, ado-trastuzumab emtansine for HER2 over-expressed or amplified breast cancer
Vemurafinib, dabrafinib, trametinib for BRAF mutated melanoma
Crizotinib and ceritinib in ALK translocated NSCLC
Afatinib in EGFR mutated NSCLC

9. Advantages of enrichment design
Targets larger treatment effect less diluted by non-sensitive tumors
Often requires many fewer randomized patients than standard designs
Avoids exposing patients to adverse effects of drug until drug is shown effective for those whom it is supposed to benefit
Clarity of interpretation

10. Cancer biology is complex and it is not always possible to have the right single predictive classifier identified with an appropriate cut-point by the time the phase III trial of a new drug is ready to start accrual

14. “Adaptive keep your job design”

15. Candidate Predictive Biomarkers for PD-1 T Cell Checkpoint Therapy
Percentage of tumor cells expressing PD-L1 ligand
Presentation of immunogenic neo-antigens by tumor cells

16. Adaptive enrichment involves restricting the eligibility criteria during the course of the trial based on interim results
Pre-specified interim analysis times
Pre-specified adaptive enrichment algorithm
The Simon & Simon framework includes cut-point based enrichment, single binary marker enrichment or enrichment based on multi-marker modeling

17. Example

18. Example
At interim analysis determine the mle’s of the regression coefficients and their estimated covariance matrix
Compute approximate mean and variance of

20. One statistical significance test is performed at the end of the trial.
It includes all randomized patients

21. Tests strong null that the treatments are equivalent for all patients
Fixed sample size regardless of changes in eligibility
Except for early termination for futility
Significance test preserves the study-wise type I error even with time dependent and data dependent changes to covariate distributions

22. Survival Data Use intermediate endpoint for adapting eligibility
PFS or response
zk=standardized log-rank statistic for patients accrued in period k but evaluated at final analysis time

23. Our simulations show that adaptive enrichment can substantially increase the statistical power with adaptive threshold determination and multi-biomarker modeling

24. p0=.2, p1=.5, K=5, Ntot=200, all pts 100/yr
True cut-point
Power adaptive
Power non-adaptive
Accrual adaptive
Accrual non-adaptive
.25
.968
.955
2.55
2.25 .5
.897
.726
3.19
3.25
.67
.768
.424
3.97
4.75
Single quantitative biomarker uniform on (0,1)
Two period cut-point enrichment
Binary response; control p0
Response rate for E is p1 if B≥cut-point

26. Example Specify prior distributions for the regression coefficients
At interim analysis determine the posterior distributions
Exclude eligibility for patients with covariates such that

36. Description of Bootstrap De-biasing for a Two-period clinical trial
Outcome model as function of covariates is estimated using full dataset.
Bootstrap sample of outcomes for first period patients is generated from estimated outcome model
Eligibility restriction is determined for second period
Treatment effect in first period for patients eligible for second period is empirically determined
It is compared to treatment effect for those patients based on the model used to generate the data; the difference is an estimate of bias
Estimates of bias are averaged over multiple bootstrap samples

37. Simulation of J Disjoint Strata with Binary Response
True response probabilities for different strata are not modeled. They are estimated separately from the data.
pT(j)=response probability of test rx for stratum j
pC(j)=response probability of control for stratum j

38. Simulation of J Disjoint Strata with Binary Response
Restrict 2nd period eligibility to stratum j if the likelihood of the stratum j period 1 data for the model pT(j)=pC(j)+δ* relative to the model
pT(j)=pC(j) is < a threshold ϒ where δ* and ϒ
are pre-specified.

39. 100 pts/period; nsim=1000,nboot=100
Half of strata with rx effect (p0=.2,p1=.5)
δ*=γ=.25

43. Extensions We have subsequently extended the analysis plans to
Improve precision of estimate of treatment effect (TE) for the eligible population selected at the final interim (the intended use population)
Identify conditions in which rejection of the global null hypothesis permits rejection of the null for the eligible population selected at the final interim analysis

44. Thank you