Translational Update of Probability of Success in Drug Development

Size: px
Start display at page:

Download "Translational Update of Probability of Success in Drug Development"

Transcription

1 Translational Update of Probability of Success in Drug Development Francois Vandenhende, Ph.D. Chairman and Executive Consultant, ClinBAY Invited Lecturer, Institute of Statistics, UCL, Belgium Translational Medicine () 7 March 2007 Francois Vandenhende 1 / 27

2 Outline Presentation Highlights 1 Drug development framework 2 Data-driven p(success) estimation Sequential update from study to study Multicriteria decision for drug differentiation 3 Expected drug profitability 4 Summary and conclusion Translational Medicine () 7 March 2007 Francois Vandenhende 2 / 27

3 Drug Development Framework Sequential Experiments Drugs are developped in stages: 1 Sequential assessment of drug properties, and 2 Optimization of drug characteristics. Each study generates data on selected properties (eg, efficacy or safety). Decisions are made at intermediate milestones. Using preclinical, medical, manufacturing and commercial data. Translational Medicine () 7 March 2007 Francois Vandenhende 3 / 27

4 Drug Development Framework Decisions Pharmaceutical drug development is all about decision making. Selection of a lead among several candidates in discovery. Selection of a dosage regimen in phase I/II. Regulatory approval in phase III. Patient buy-in after launch. p(success) The overall probability of a successful launch combines all risk estimates at intermediate milestones: p(success) = p(launch III)p(III II)p(II I)p(I disco)p(disco). Translational Medicine () 7 March 2007 Francois Vandenhende 4 / 27

5 Drug Development Framework Illustration If success rate is 80% in each phase, and there are 5 phases to launch: Discovery,Phase I, II, III, and Launch = Then, overall probability of success till launch equals 80% 5 = 33% Uncertainty at each step adds up multiplicatively in the overall risk estimation. If a GO decision is made (let say start Phase II), the risk on that decision still needs to be carried forwards. All historical data matter for risk assessment! Translational Medicine () 7 March 2007 Francois Vandenhende 5 / 27

6 p(success) Definition p(success) = p(set of Drug Properties Range) Risk assessment quantifies the chance that the most relevant drug properties are within an acceptability range. The range may be determined based on: 1 targeted risk levels for efficacy/safety, 2 competition, 3 development costs, duration, 4 revenue,... Translational Medicine () 7 March 2007 Francois Vandenhende 6 / 27

7 p(success) Illustration New SNRI under development. Phase I SERT blockade trial using 11C-DASB. Occupancy of serotonin transporters as primary response. Success as a single criterion Success is defined as receptor occupancy (θ) larger that 65%. Investigated property (θ) is RO p(success) = p(θ > 65%) Environmental constraint is > 65%. May come from competitive analysis. Phase I PET data are used to estimate that probability. Translational Medicine () 7 March 2007 Francois Vandenhende 7 / 27

8 p(success) Estimation Prior to PET study, one may have a prior estimate on p(success): 1 Objective: based on preclinical studies, or 2 Subjective: based on perception or guesses. 3 Non informative: when prior is vague. Then study is run that generates data Observed RO data are 60, 70 and 80% in N=3 subjects. The updated probability of success is then calculated as p(success) = p(θ > 65% data {60, 70, 80%}). The posterior probability combines trial data and prior info, as p(θ data) p(data θ)prior(θ). Translational Medicine () 7 March 2007 Francois Vandenhende 8 / 27

9 p(success) Illustration: First Study in N=3 Assuming no prior information, p(success) is updated as based on N=3 data points. Translational Medicine () 7 March 2007 Francois Vandenhende 9 / 27

10 p(success) Illustration - Update after a second study We enroll 3 more subjects: RO data is also {60, 70, 80%}. Then, p(success) is updated from trial #1 to trial #2 as: p(θ data 1, data 2 ) p(data 2 θ)p(θ data 1 ) using the new data 2 and the initial update after data 1 as new prior. Overall probability increases Overall variability decreases better clear-cut GO decision. Translational Medicine () 7 March 2007 Francois Vandenhende 10 / 27

11 p(success) Update p(success) is updated as more data (ie, trials, subjects) are available. Accural may be stopped when one has sufficient insurance on success/failure. All cumulative data are used to assess p(success). Do not discard undesired data or outliers. They will vanish away if really spurious. Adjust trial size to target p(success) cut-off points for decision making. Translational Medicine () 7 March 2007 Francois Vandenhende 11 / 27

12 p(success) Dose-selection Translational markers such as Receptor Occupancy PET may be used to assist dosage selection prior to phase II. We illustrate the use of p(success) estimation for another type of decision than go/no go : dose selection. dose = arg min{p[success(dose)]) threshold} Dose-response PET Study RO PET in N=15 subjects. Multiple q.d. dosing of drug candidate to steady-state. Doses of 1, 10, 30 and 100mg. Scans at baseline and 24h post-last-dose. RO estimated in target brain region. Success is defined as θ > 80%. Translational Medicine () 7 March 2007 Francois Vandenhende 12 / 27

13 p(success) The dose producing θ = 80% is 20 mg. Probability of success at 20mg is 50% If one wants more insurance (let say 90%), the minimum dose should be 32mg. The target dose is precisely estimated here (pink curve is steep). Translational Medicine () 7 March 2007 Francois Vandenhende 13 / 27

14 p(success) Adaptive Design Assume this was not the case. Can we generate more data to improve precision of target dose? How many subjects? At which doses? Bayesian c-optimal design Criterion: precision of selected property. Here: Our dose-response model is a 3-parameter Emax. Dose producing RO= 80% is Dose 80% = (80% E 0)ED 50 Emax 80% + E 0 We add + 4 subjects at doses so that drop in var( Dose 80% ) is maximized. Translational Medicine () 7 March 2007 Francois Vandenhende 14 / 27

15 p(success) Dose selection Optimal choice is to place all 4 subjects at 10mg. Translational Medicine () 7 March 2007 Francois Vandenhende 15 / 27

16 p(success) Dose selection As expected, limited benefit in precision. Study may be stopped. Translational Medicine () 7 March 2007 Francois Vandenhende 16 / 27

17 p(success) Summary We have defined success as Drug Properties Acceptability Range. We decide based on a simple and standardized measure: p(success). We estimate and update p(success) using all available data: p(success) = p(drug Properties Acceptability Range Data). Possible decisions: 1 Candidate selection: Take drug with largest p(success) 2 Go-no go decision: if p(success) > threshold 3 Drug characteristic optimisation such as dosage selection: dose = arg min{p[success(dose)]) threshold} 4 Adaptive design and optimal sizing of trials. Translational Medicine () 7 March 2007 Francois Vandenhende 17 / 27

18 Multicriteria decision Let us now discuss multi-criteria decision making in the framework of drug differentiation. Multicriteria decision Clinical research generates a lot of data. Most of which are non-informative. When deciding based on multiple responses, it is good to focus on a limited set of responses, including: 1 The most relevant, 2 The most sensitive. Decision making is invariant to the non-informative responses when there is: 1 No signal at all, or 2 A complete saturated signal. Translational Medicine () 7 March 2007 Francois Vandenhende 18 / 27

19 Multicriteria decision Example: Efficacy & safety Phase II study in diabetes. Design: 1 Placebo controlled, parallel, 12-week treatment (N=92). 2 4 dose levels: 0.04, 0.20, 0.80, 1.20 mg. Efficacy: fasting serum glucose (FSG) change at endpoint. Safety: body weight change from baseline. Success Ranges: δfsg > 1.5mmol/L and δbwt < 2kg. Translational Medicine () 7 March 2007 Francois Vandenhende 19 / 27

20 Multicriteria decision Joint p(efficacy & safety) Efficacy & safety are analysed using a Bayesian bivariate analysis of variance model. We estimate the probability of success given trial data: 1 For Efficacy alone (blue), 2 For Safety alone (pink), and 3 For both efficacy and safety (yellow). The optimal dose range is between 1 and 1.2 mg. Translational Medicine () 7 March 2007 Francois Vandenhende 20 / 27

21 Multicriteria decision Sensitivity analysis p(success) depends on the thresholds that were selected for each response: δfsg > 1.5mmol/L and δbwt < 2kg. In the illustration (blue curve), high p(success) indicates that the drug beats the thresholds easily. Let us study changes to p(success) when increasing the efficacy and safety hurdles. Translational Medicine () 7 March 2007 Francois Vandenhende 21 / 27

22 Multicriteria decision Comparison to Standard of Care When a relevant threshold is found, p(success) may be compared with an active comparator. The chart below presents 4 possible scenarios: 1 P(success) is overall inferior or superior to that of standard of care, or 2 Each drug has a its own pros and cons. Translational Medicine () 7 March 2007 Francois Vandenhende 22 / 27

23 Multicriteria decision Summary When dealing with multiple properties, Filter the most revelant and informative only. Then, p(success) can be updated in the same way for each property separately, or for a combination of properties (including all). The same updating method is utilized when collecting more data. Decisions are again possible based on p(success). Sensitivity analysis is useful to fully understand Drug performance, and Competition landscape. Let us now illustrate some financial components related to decision making. Translational Medicine () 7 March 2007 Francois Vandenhende 23 / 27

24 Expected Benefit Revenue and Costs vs Threshold Let us focus on one variable (ie, efficacy). A standard measure of success is the effect size versus placebo. With large effect sizes, chances are hight that the drug becomes a blockbuster. The revenue should increase with ES. To the contrary, development costs should decrease with increasing effect sizes (ie, lower sample size, fewer studies). Translational Medicine () 7 March 2007 Francois Vandenhende 24 / 27

25 Expected Benefit Benefit and Effect Size The benefit at any ES equals Benefit(ES) = Revenue(ES) - Cost(ES) Knowing the posterior distribution of ES : p(es data) It is possible to calculate the expected benefit as E[benefit] = Benefit(ES)dp(ES data) E[B]=0.32B$ in this case. Translational Medicine () 7 March 2007 Francois Vandenhende 25 / 27

26 Conclusions p(success) for decision making We have emphasized the value of using data-driven p(success) in decision making. We showed how to update p(success) sequentially based on prior and new data. Types of decisions: Go/no go Optimisation of drug characteristics Product differentiation Development plan optimisation Portfolio management Simple mathematics known as Bayesian techniques are the basis to all calculations. Feel free to contact me for more info Translational Medicine () 7 March 2007 Francois Vandenhende 26 / 27

27 Acknowledgments Eli Lilly: David Manner Ming-Dauh Wang Cytel Inc: Nitin Patel Ann Cleverly Fabian Tibaldi Suresh Ankolekar Thank you! Any question? Translational Medicine () 7 March 2007 Francois Vandenhende 27 / 27