1. Rui (Sammi) Tang Biostatistics Associate Director, Vertex
Bayesian Model Screening Platform Design for Randomized Phase II Combo Therapy Oncology Trials
Rui (Sammi) Tang
Biostatistics Associate Director, Vertex
Collaborators: Jing Shen, Ying Yuan

2. Disclaimer
The presentation reflects presenter’s own views and do not represent the view of Vertex; The presenter is not speaking on behalf of Vertex.

3. Overview Motivation Challenges and Proposed Method Simulation Results
Summary
Conclusion and Discussion

4. An Era of Immunotherapies in Oncology
Source: Morrissey, K., Yuraszeck, T., Li, C.-C., Zhang, Y. and Kasichayanula, S. (2016), Immunotherapy and Novel Combinations in Oncology: Current Landscape, Challenges, and Opportunities. Clinical And Translational Science, 9: 89–104. doi: /cts.12391

5. Selected List of Combination Immunotherapies in Clinical Development
Source: Morrissey, K., Yuraszeck, T., Li, C.-C., Zhang, Y. and Kasichayanula, S. (2016), Immunotherapy and Novel Combinations in Oncology: Current Landscape, Challenges, and Opportunities. Clinical And Translational Science, 9: 89–104. doi: /cts.12391

6. Challenges
Combination therapies may dramatically improve the outcome for cancer patients
Combination is a future direction of cancer therapy development
Discovery of effective combinations is a challenging endeavor
Nearly 200 molecules approved by the FDA by 2016
Over 15 immunotherapy agents
Unfeasible: experimentally testing every possible combination
Urgently NEED: Innovative study design to efficiently identify effective combination therapies

7. Motivatonal Trial Setting and Challenges
Primary investigation compound 1
Multiple Back-up compounds (Similar MOA)
Multiple immunotherapies as backbone choices
The same disease indication
How to select the most efficient combination therapy(ies) quickly?

8. Screen Selection Design (SSD)
Recruit N patients
Randomized 1:1:1:…:1 (x arms)
Simon two stage approach applied to each arm
H0: P0 <=20%
H1: P1>40%
Arm1
Arm2 …
Arm x
Bayesian stopping rule
could also apply (BSD)
No arms meet the criteria to proceed to next stage
Treatments with positive results are recommended
Apply selection strategy with arm comparison to pick the winner
Can not take the advantage of combination therapy; Ignored the potential relationship between backbone or compound arms;

9. Proposed Design and Method Bayesian Model Screening Platform Design (BMSPD)
For simplicity and without loss of generality
Assume efficacy endpoint is an event of response following a Bernoulli distribution
Other efficacy endpoint could apply to the same concept of design
The Bayesian hierarchical logistic model is used to model the rate of response
Let i and j denote the compound and backbone
k indicate the index of patients assign to the compound i and backbone j.

10. Proposed Design and Method Cont.
The hierarchical model setting allows borrowing information across different backbone groups (j) within a compound (i)
Instead of setting default vague prior to 𝜎 2 , we allow choosing 𝜎 2 from 106, 104,10, 1, 0.1 to construct 5 model candidates representing weak to strong borrowing strength models
Using DIC based model selection criterion to select the best fitted model among the candidates after the first stage
Can add and early drop arm(s)
We propose to have n interim look during the trial and the model prior are determined after the sth interim (typically s=1)
At the end of the trial, the non stopping arm(s) who has the largest winning probability Pr(p_ij = max(p_ij) |data) will be selected for future development. Other threshold selection rule can also apply
The specifications of s will be determined by the study team according to practical reasons
An extreme case would be evaluating data after each subject's outcome like the Zhou et al. (2008) BATTLE trial

11. BMSPD Design Process

12. Simulation Set up Simulation runs =1000
Assume a set-up of a 3 (compound) by 3 (backbone) arms with pij corresponding to the response rate of compound i and backbone j combination.
Response rate
Backbone 1
Backbone 2
Backbone 3
Compound 1
P11
P12
P13
Compound 2
P22
P23
Compound 3
P31
P32
P33

13. Simulation Results for trivial cases
All arms are equally non-efficacious(bad) or with moderate efficacy
Case
Scenario
SSD
BMSPD
BSD
Sample size
Percentage of no arm selected 1.
All bad arms 81 1 63 2.
All moderate arms
100
0.71
103
0.86
104
0.88
SSD: Screen Selection Design
BMSPD: Bayesian Model Screening Platform Design( proposed method)
BSD: Bayesian Screening Selection (Basic Case of Bayesian framework)

14. Simulation Results for Cases at Least 1 Superior Arm Exists
SSD
BMSPD
BSD
Sample size
Selection Percentage 3.
Compound effect
137
0.78
144
0.88
120
0.62 4.
Backbone effect
140
0.73
152
0.79
105
0.49 5.
Total random
163
0.81
157
0.91
153
0.87 6.
121
0.71
117
0.80
112
0.74 7.
Additive effect
167
169
171 8.
Synergy effect
161
172
0.89
164
0.83 9.
Drug-drug interaction
141
0.85
143
0.90
142
Compound effect case: the low selection rate of BSD where the priors of variance depends on hyper prior of IG mistakenly borrowed too much information cross the backbones and neutralized the compound effect.
Similarly, in backbone effect case the BSD mistakenly borrowed too much information across compound and neutralized the backbone effect.
Selection rate= Percentage of selecting superior arm(s)

15. A R shiny tool is being develop for the method
It is a very interactive tool for trial fitting and simulation
Layout
Outline flowchart
Uploading files of trial set up
Trial summary
Trial fitting
Simulated trial

16. BMSPD – Uploading files

17. BMSPD – Fitting outputs

18. BMSPD – Trial simulation

19. Conclusion and Discussion
Our approach is flexible and can stop arm early if unsatisfied response rate observed.
In cases where no arm meets an acceptance rate, our method is comparable with BSM claiming no arm is selected, both are better than SSD.
In cases where superior arm(s) existed, our method can select the superior arm(s) with highest probability comparing to that of SSD and BSD.
The model tends to select the prior with high borrowing strength when response rates are homogeneous across backbone and/or compounds;
and select the prior with moderate/mild borrowing strength when the less homogeneity assumed across arms
The number of subjects enrolled were compatible across all three methods
Finally, a web based application is being developed and available to users soon.

21. Thank you

22. Acknowledgment Research Collaborators:
Jing Shen (Gilead) ; Ying Yuan (MD Anderson)
Thank my management Bo Yang (Vertex) and Yang Song (Vertex) for the review comments

23. Back up

24. Simulation set up detials
The sample size of SSD is determined by Simon's two stage design, ph2simon(p0=0.5, pa= 0.75, ep1 = 0.1, ep2=0.2): r1/n1 = 5/9, r/n =13/22, corresponding to 75% early stopping given p0 is correct after the first stage [Yap, Pettitt and Billingham (2013)]. The sample size of the two Bayesian approach BSD and BMSPD is defined as (7, 7, 8) for a 3-stage design. The early stopping rule for BMSPD and BSD are set as
𝑃𝑟 𝑝 𝑖𝑗 > 𝜃 1 𝑑𝑎𝑡𝑎)≤ 𝛿 𝑙
where 𝜃 1 = 0.7 and 𝛿 𝑙 = 0.1, which corresponds to 70% early stopping given p= 0.5 and prior is M0 at the end of the first stage.