1. Quantitative Decision Making (QDM) in Phase I/II studies
Kevin Gan, Jon Haddad, GlaxoSmithKline
2019 JSM 1

2. Agenda Challenges for clinical design in Pharma R&D
Benefits for Bayesian based QDM framework
Illustration for Phase 1 and Phase 2 studies

3. Challenges in Clinical Trial
Cost of bring new drugs to market increase
Low probability of success for novel candidates
High failure rate in phase II and III
Success rate across industry in Phase 2 is only 18%
Pharma R&D being pushed on several fronts:
Deliver more with less resource
Innovation on top of innovation
Streamline development and crisper decision making
More consideration for adaptive clinical trial and bring more Quantitative Decision Making Framework 3

4. Adaptive Trial
In theory – this design framework allows you to ‘adapt’ pretty much any aspect of a trial.
Some adaptations will be more acceptable than others.
Add/drop treatment arms
Sample size
Randomization ratio
Analysis schedules (inserting or dropping interims)
Combine trials from different phases
Hypotheses
Primary endpoint
Study population 4

5. Advantage of QDM based Adaptive Trial Design
More information earlier in development, with increased efficiency for quicker decision making
Kill ineffective compounds earlier
Take the right doses into Phase III
Improve characterization of dose response
Increase “probability of technical success”
To market faster – shift to seamless combined phase studies
Optimize patient treatment within a study
Minimize patient exposure to ineffective/unsafe treatment
Maximize patient exposure to effective treatment
Enhance design efficiency by learning as you go
Most valuable in situations of prior uncertainty
Answers the right question 5

6. Phase I Studies FTIH: single dose escalation trial (alternating)
2nd trial: Repeat dose escalation trial (parallel)
Food effect trial
Drug-drug interaction studies (DDI): probes study
Relative Bioequivalence (BE): compare different formulations, different dose forms 6

7. EX: FTIH trial First Time in Human (FTIH):
Single dose escalation study in healthy subjects, n=6~10
Double blind trial
Dose: mini predicted measurable dose to highest dose under the toxi coverage
Alternating panel design
Safety and pharmacokinectic evaluation at each dose
Dosage form: liquid, suspension 7

8. FTIH Study Design FTIH MAD – Part B FTIH SAD – Part A
Healthy Subjects
FTIH MAD – Part B
Healthy Subjects
Option: Combine Phase 1 and POC study
Dose 1 SD
Dose x QD
FTIH MAD – Part C
Patient Subjects
Bayesian decision after each dose
Safety & PK
Review
Dose 2X QD
Dose 2 SD
Safety and PK Review
Dose 4X QD
Dose 4X QD
Dose 3 SD
Dose 4 SD
Bayesian decision Rule: Eg. Dose escalation in Part A (SAD) will be stopped in the event that the Bayesian
predicted probability is >50% that any participant’s Cmax for the next subsequent dose will exceed X.xx μg/mL 8

9. Cohort 1: treatments A, B, C Subjects from Cohorts 1&2: F/G
EX: RBA (BE) study
Relative Bioavailability study, Relative Bioequivalence study
Evaluation &
Decision
Part A
Part B
Cohort 1: treatments A, B, C
Subjects from Cohorts 1&2: F/G
Cohort 2: treatments A, D, E
Treatments F/G: 3rd formulation with or without food
Treatment A: Gelucire formulation/with food (reference)
Treatments B/C: 1st formulation with or without food
Treatments D/E: 2nd formulation with or without food 9

10. Endpoint Considerations
Endpoint: R = test/reference, the ratio of the geometric least square mean (GLSmean) of PK parameters of test treatment vs reference treatment
Evaluation criteria (through simulations):
90% CI: ( ) or ( ): Not chosen as required large sample size
R >0.8 or R >0.9 for the point estimate: Chosen when small sample size
Based on posterior probability
Non-informative prior
Different decision rules
Interaction with clinical team 10

11. Phase 2a: POC study POC (proof of concept design)
Multiple active doses + placebo arms
The goal is to identify the best dose for Phase 2b studies
Finding exposure response curve is the main purpose of POC study 11

12. Example: POC Study Design
Part 1
Part 2
Dose (Max)
N=6~10
Dose 3 (ED90)
Dose 4 (ED75)
Dose 2 (ED50)
n = 6~10
Dose 5 ED30
PBO
Unblinded
Interim Analysis
Conducted by
Study Team
(Including PK, safety and
HIV Viral Load data)
PBO
Day 1 Randomization
Day x
Day 1 Randomization
Day x

13. Decision Rule for Interim Analysis
Futility rule for Interim analysis VL data
Set up Futility criteria, EX:
Adjust doses for Part 2 based upon PK: Decision Tree (one example below)

14. Decision Making
MV: Minimum Value TV: Target Value
MV: The true value of drug effect that would trigger development interest TV: The value would represent a really commercialisable and desirable medical product.
Endpoint 1: Viral Load drop
Endpoint 2: Safety endpoint:
EX: Decision Rules:
Go if P(∆>MV)>80%
STOP if P(∆<TV)>90%
Else CONSIDER
EX: Decision Rules:
Go if AESI rate< cutoff value 1
STOP if AESI rate> cutoff value 2
Else CONSIDER

15. Bayesian decision Rule: Viral load data Illustration
Bayesian Posterior of
Mean VLD based on trial data 15

16. GO Decision
Based on QDM rules (The example here used MV=1.3)

17. Stop Decision
Based on QDM rules (The example here used MV=1.3)

18. Simulation: Observed Trial Mean VLDs and Decisions
Based on Different Decision Rules: Interim Stop, Final stop, Final consider, and Final success
Illustration when true mean=1.3, all different observed mean and corresponding decision rules with 1000 simulation trials. 18

19. Operating Characteristics for QDM framework (Mock data)
Decision Rules: EOS
Go if P(∆>MV)>80%
STOP if P(∆≤TV)>90%
Else CONSIDER
MV: Minimum Value
TV: Target Value
Combined No Go decision
Interim decision rule:
STOP if Predictive Probability of Stop at EOS is high
True VLD
Possibility of achieving each decision
Interim NO GO (%)
Final
No GO (%)
Final Consider (%)
Final Go
(%)
No effect 1.0
82.5%
11.4%
6.1%
0.0% MV
18%
3.5%
75.1%
3.4% TV
2.0% 0%
48.9%
49.1%
High effect
0.2%
17.5%
82.3%
93.9% 19

20. Predictive Probability for Future at EOS
VL drop >1.7 or 2.0 based on Emax model of Ctau

21. Challenges related to adaptive Trial/QDM setting
Understanding well for medical profile, Ex: identification of MV or TV
logistics for conducting the trial could be complicated
Likely to involve more set-up time than a traditional study design
Requires real-time availability of data
Recruitment rate not too quick in relation to data availability
Requires prompt analysis and decision making
Usually requires more sophisticated statistical methodology
Likely to require simulations to understand study operating characteristics, which can be time consuming
Drug supplies can be complex
Uncertainty with final budget and study duration
Access to unblinded data
Accrual and time bias 21

22. Technical Challenges Understand the objectives Opportunities
Computation: FACTS, R, SAS and WinBugs
Time for simulation runs and set up the criteria 22

23. Questions?
Thanks 23