Assessment GSK R&D
Disclosure is an employee and shareholder of GSK Data presented is based on human research studies funded and sponsored by GSK 2
Outline 1. Motivation 2. GSK s Approach to Benefit-Risk Assessment 3. Bayesian Joint Modelling of Mixed Outcomes 4. A GSK Case Study 5. Summary 3
Motivation Growing industry and regulator awareness that more quantitative methods can contribute to transparency of benefit-risk assessment Efficacy and safety signals could be linked via exposure to active drug Joint modelling of efficacy and safety endpoints enables data driven quantitative assessment of the benefit-risk profile Bayesian inference provides a direct framework to build relevant and intuitive probability statements in the context of benefit-risk 4
GSK s Approach to Benefit-Risk Assessment A 3 Step Process... Frame Model / Analyse / Graph Conclude The endpoints are important because and we met them or we didn t and this is why the drug will matter, or why it shouldn t continue. Value Tree Graph, effects table, others BENEFIT-RISK PROFILE Based on this evidence, we believe the benefit outweighs the risk because 5
Subjects GSK s Approach to Benefit-Risk Assessment Graphical Methods Separate analysis of benefits and risks Results presented jointly Forest Plot Norton Heatmap Active Benefit and Risk Comparison Placebo 100 90 80 70 60 50 40 30 20 10 0 1 2 3 4 0 1 2 3 4 Weeks Benefit Only Benefit +AE Neither AE only Withdrew 6
GSK s Approach to Benefit-Risk Assessment More Complex Methods Joint analysis of benefits and risks Results presented jointly Utility Index Contour Plots 7
Bayesian Joint Modelling of Mixed Outcomes Motivation Strength of efficacy and safety signals likely to be linked at subject level via exposure to active drug: Higher exposure Higher efficacy Higher risk of AE continuous, count, binary, etc binary, count, continuous, etc E.g., antibody drugs for diabetes mellitus, increases in C-peptide (efficacy) and cytokine releases (safety) Approach that accounts for observed correlation at subject level between efficacy and safety signals is desirable Often efficacy and safety endpoints modelled using different distributions Further development conditional on magnitude of efficacy and safety effect 8
Bayesian Joint Modelling of Mixed Outcomes Approaches to Linking Mixed Outcomes Option 1: Use generalised linear mixed models Assume J different observations on same subject (each following some distribution) For subject i with mean response mu i, g(mu i ) = X i b + Z i u i, u i ~ N(0, G(X i )) Random effect u i is shared across all J observations for subject i thus modelling potential correlation When g j (.) is not identity function then fixed effects b are conditional on random effects u i Monte Carlo integration can be used to obtain marginal population effects Constraints may be necessary to ensure identifiability for certain distributions 9
Bayesian Joint Modelling of Mixed Outcomes Approaches to Linking Mixed Outcomes Option 2: Use copulas Copulas - distribution functions used to form new multivariate distributions given set of marginal distributions of interest (which are preserved) E.g., H(y 1,y 2 ) = C(F(y 1 ), G(y 2 ) θ), F(.) and G(.) CDF of marginal distribution of rv y 1 and y 2 C(.,. θ) is the copula function (e.g., Gaussian CDF) θ measures association between y 1 and y 2 Directly obtain marginal population effects for parameters of interest Choice of copula C(.) may impact results through different dependency assumptions Difficult to interpret beyond 3 dimensions (non-unique model definition) 10
A GSK Case Study Compound X in Adults With Schizophrenia Antipsychotic drugs can cause extra pyramidal side effects (EPS) Dystonia (involuntary muscle contractions) Akathisia (extreme, uncontrollable restlessness) risk of suicide Tardive dyskinesia (involuntary, repetitive movements)... Compound X novel antipsychotic - antagonist dopamine D2/D3 and 5HT2 receptors However, D2/D3 receptor antagonism is associated with EPS Clinical trial to evaluate safety and efficacy of compound X in acute schizophrenia Placebo and active comparator (Olanzapine ) In this talk, focus on compound X 120mg (54 subjects) vs placebo (52 subjects) Development of compound X was stopped due to preclinical finding of phospholipidosis 11
A GSK Case Study The Value Tree Symptoms Decrease in PANSS TS Benefits Benefit-Risk Balance Individual Risks Weight Gain Glucose Intolerance / Insulin Resistance High Dyslipidemia Identified Benefit or Risk Category Identified benefit/ risk Outcome Akathisia Risks Nervous System Disorders / EPS Tremor Dystonia Dyskinesia AIMS Scale Barnes Akathisia Scale Simpsons Angus Scale 12
A GSK Case Study Efficacy and Safety Endpoints and Prior Distributions PANSS Total Score (TS) = Positive and Negative Syndrome Scale measures symptom severity in patients with schizophrenia Highest possible score is 210 = most severe measure of schizophrenia Lowest possible score is 30 = subject not suffering from schizophrenia Change from baseline (CFB) in PANSS TS at week 6 ~ N(µ, σ 2 ) Efficacy threshold for further development: treatment difference (TD) = -8 pt, but clinically meaningful TD is ~ -15 Adverse events (AEs) associated with Nervous System Disorders I AE = 1 if subject has an AE, 0 otherwise, I AE ~ Bernoulli(p), p = Prob(AE) Safety threshold for further development : odds ratio (OR) = 1.5, huge level of unmet need, large number co-morbidities, top 5 causes of disability in individuals < 25* Non-informative priors assumed for all parameters in the model Covariates in the model include age, baseline PANSS TS and treatment Hermes et al, 2012 * Nasrallah et al, 2013 13
A GSK Case Study Exploratory Analysis Correlation between efficacy and safety All outcomes: the subjects who had a nervous system Corr X = -0.18, and Corr Placebo = -0.15 AE and took drug X had an improvement from baseline in PANSS TS by week 6 - Improvement in score associated with AEs? 14
A GSK Case Study Estimated Joint Posterior Distribution of Treatment Difference and Odds Ratio How much evidence exists from the data to support the Benefit-Risk profile: TD < - 8pt AND OR < 1.5? Joint Posterior Median Parameter Median (95% Crd.I.) TD -7.01 (-15.91, 1.90) OR 2.65 (0.76, 12.39) Corr X -0.10 (-0.38, 0) Corr Placebo -0.077 (-0.30, 0) Safety Threshold: OR = 1.5 Efficacy Threshold: TD = -8 Prob(+ BR Profile) = 7% 15
A GSK Case Study Benefit-Risk Evaluation and Decision-Making How much evidence (probability) exists to support a chosen Benefit-Risk profile? Higher probabilities for lower benefit thresholds and/or higher risk thresholds 16
A GSK Case Study Predicting Response and AE event for a New Subject A new patient, aged 50, is diagnosed with schizophrenia (baseline PANSS score of 119) and the GP is considering whether to prescribe compound X......will compound X be effective, do nothing, or harm this patient? We can predict both the efficacy and safety responses for this new subject given what we have learned from the study data Predicted CFB in PANSS TS after 6 weeks = -19.22 (95% CI [-58.46, 19.91]) Predicted probability of AE of nervous system disorder = 0.30 17
A GSK Case Study Predicting Response and AE event for a New Subject Use predictive distribution for CFB PANSS TS to make predictive probabilistic statements around particular levels of efficacy Prob (Observed CFB PANSS TS at week 6 < 0 NAA Data) = 84% Prob (Observed CFB PANSS TS at week 6 < -15 NAA Data) = 58% Although this patient has a high probability of benefitting from drug X, it has just over 50% chance of achieving a clinically meaningful improvement, whilst having a 30% chance of experiencing an AE The GP may still choose to prescribe drug X if no alternative therapies are available 18
Summary Benefit-Risk assessment is both qualitative and quantitative Bayesian inference based on joint models of mixed outcomes is a powerful tool for Benefit-Risk assessment Explore dependency between benefit and risk thresholds for decision-making Joint (and conditional) probabilistic statements Predicting responses for a new subject conditional on what was learned from study data Benefit-Risk profile is a combination of two different quantities: Set of thresholds for efficacy and safety define Benefit-Risk profile Level of evidence (posterior probability) to support Benefit-Risk profile 19
Acknowledgements Thomas Drury (Integral Statistics Limited) GSK: Susan Duke Colleen Russell John Krauss 20
Thank you