Predicting the Probability of Developing a Successful Cancer Therapy Tony Sabin, CSG Feb 2010 Contents Brief Overview of Clinical Drug Development Process – Aims of Clinical Development – Stages of Development – Cost Decision Making in the Pharmaceutical Industry – Historical Perspective Bayesian Model for Predicting Success in Pancreatic Cancer 2 Aims of Clinical Development Stop development of poor treatments as soon as possible – – – – avoids giving patients ineffective/unsafe therapies avoids escalating costs of development allocate funds to develop more promising treatments make more patients available for promising treatments Decision making process should possess a – High chance of failing a poor treatment – Low chance of failing a good treatment 3 Clinical Development Process Phase 1 First in Human FIH Phase 3 Phase 2 Proof of Concept PBA Learn Dose Find Confirm EOP2 CTF Confirm Filing Launch CTL Launch Post marketing Phase 4 4 First in Human Trials Single centre, healthy volunteer studies – Occasionally done in patients e.g. when potential toxicity precludes use of normal volunteers Single dose safety Increasing doses in small cohorts based on safety Single dose PK - Determine absorption, distribution, metabolism & excretion Questions Is it possible to dose without gross safety signals? What is the single dose PK profile? 5 Proof of Concept Trials Volunteer/patient studies, few centres Repeat dose safety Repeat dose PK/PD Different formulations, schedules and doses Increasing doses in cohorts based on safety and PD/efficacy Questions Is there a range of safe doses where it may be possible to observe the intended benefit? What is the best formulation and schedule of dosing? Is there any sign of the intended efficacy? 6 Dose Optimization Trials Patient studies, more centres Clinical durations of treatment Short term biological/clinical endpoints that are good predictors of likely phase III endpoints Few doses against a control (active and/or placebo) Questions Which dose is likely to give the best balance between efficacy and safety? Are we likely to be successful in a phase III? What is the likely cost effectiveness of treatment? 7 Confirmatory Trials Patient studies, many centres and regions Clinically relevant endpoints One or two doses compared against a control Studies may take a long time to run for chronic treatments Prior agreement of design with regulatory authorities Questions Can we provide robust evidence of clinical efficacy? Is there any evidence of ‘case identifiers’? What is the safety profile of treatment in the target population? Can we successfully file for marketing authorization? 8 Observational Trials How is the treatment actually being used in practice Is the treatment being used on label Are there patients who should be getting the treatment 9 Decision Making Quote: If we knew what it was we were doing it wouldn’t be called research, would it? (Albert Einstein) Success Rates from FIH to Registration by Therapeutic Area Kola and Landis. Nature Rev Drug Discov 2004;3:711. 11 Success Rates by Phase and Therapeutic Area Kola and Landis. Nature Rev Drug Discov 2004;3:711. 12 Cost and Number of Compounds at Each Stage Cost $ # Compounds 304M 5000 – 10000 Screened in Discovery 96M 250 Entering pre-clinical 72M 10 Entering Phase 1 48M 3-5 Entering Phase 2 224M 2 Entering Phase 3 56M 1 Regulatory submission 800M 1 Approved and launched Source: PhrMA and EFPIA Pharm Report 2007 13 There is a high attrition rate within drug development Considerable amount of attrition in late phase development which means incurring the majority of drug development costs. – This is not good for R&D productivity – This is not good for cost of medicines Finding ways to make better decisions earlier on in development would be a big step in improving R&D productivity – How can statistics help? – Focus on the end of Phase 2 decision point – Bayesian Model for Predicting Success in Pancreatic Cancer 14 Bayesian Approach to Predicting Success in Phase 3 Pancreatic Cancer Pancreatic Cancer Fourth leading cause of cancer-related deaths in the US – 32,240 deaths and 42,470 new cases in the US during 2009 Usually diagnosed at a late stage and is largely unresponsive to current medical therapy – Highest cure rates in people with localized disease which are amenable to surgical resection (10-24% 5 years survival rate) – 90% present with unresectable or metastatic disease with a 5year survival rate < 1% Current acceptable treatment is gemcitabine (chemotherapy) – 1000mg/m2 once a week for 3 weeks then one weeks rest (repeated until disease progression) 16 Pharmacology of Apoptosis RTK inhibitors Bcl-2 inhibitors Bcl2, BclXL, Mcl1 Akt Flip FLIP Akt/PI3K inhibitors IAP Smac mimetics 17 PI3K Common Endpoints in Oncology Monitor the change in tumour size over time and classify as progressed, stable, partial response or complete response at a time point • Progression Free survival time (Time to progression or death). Compare trts using a Hazard Ratio, difference in PFS rate at a time point • Objective Response Rate (% best response of CR or PR). Compare trts using difference in proportion of responders Overall Survival Time (Time to death). Compare trts using Hazard ratio or the difference in OS rate at a time point 18 The Model for Decision Making at EOP2 Probability of Success wrt Efficacy Project Deliverables Probability of Success wrt Safety 19 Anticipated Competitive Situation Combine Information Anticipated Regulatory Situation Anticipated Reimbursement Situation GO? NO GO? Current Business Situation (Amgen) End of Phase 2 Evidence Diagram Literature Control Group Response Rate Predictors Time Trends Differences Predictors Prior Belief Phase 2-3 Endpoint Relationships Prior 20 Phase 2 Phase 3 Treatment Treatment R Diff Control R Diff Control Phase 2 Endpoint Phase 3 Endpoint New Data Prediction The competition to be first to market is fierce: Reliable short term endpoints in P2 are a must Preclinical Phase I Phase II 1 6 2 3 4 5 7 21 Phase III Launched Our Goal: Probability of Success in Pancreatic Cancer 100 Probability of Success in Ph 3 90 80 70 Better than industry average Uninformative Sceptical 60 50 Optimistic Light Sceptical 40 Light Optimistic 30 20 10 0 -15 -10 -5 0 5 10 15 Diff in 6m Survival Rate in Phase 2 (Test - Control) 22 20 The Process Step 1: Literature search and data abstraction Step 2: Determine the expected treatment difference for the Phase 2 endpoint and population – Factor in the observed results from the P2 study, prior knowledge of the control group behaviour from the literature in the P2 population – Factor in prior belief on the whether the drug will work Step 3: Use the relationship between P2 and P3 endpoints to calculate the treatment difference in the P3 endpoint Step 4: Determine the probability of success in Phase 3 – Incorporate the Phase 3 design 23 Pancreatic Cancer Example Conduct a randomized controlled Phase II study: – Test+gemcitabine combination versus gemcitabine alone – Primary Phase 2 endpoint is 6 month survival rate – PFS originally not thought to be relevant in this disease (see later – this process can change medical opinion) Proposed Frequentist Phase III study design: – 2 arms (Test+gemcitabine, gemcitabine alone) – Primary endpoint is OS. – Analyze after 379 events Our focus is on relating the difference in 6m survival rate to the OS hazard ratio 24 Step 1 – Pancreatic Cancer Literature Search Inclusion criteria – Adult patients with locally advanced or metastatic pancreatic cancer – All randomized controlled comparative studies that were published in English in year 2000 or later, in which gemcitabine was used either alone or in combination with other therapies. Exclusion criteria – Studies where patients were given concurrent radiotherapy or local regional modalities such as surgery, which might have influenced survival – Cross over studies where the assessment of survival times was impaired – Non randomized study – Information on patient survival times was not available 25 Step 1 – Pancreatic Cancer Literature Search Search Method – Studies were identified by targeting Medline, Embase, the American Society of Clinical Oncology web site, published meta-analyses and the internal knowledge of Amgen’s clinical and regulatory groups. – Studies identified were screened for inclusion by both Amgen Biostatistics and Amgen Clinical. 26 Step 1 – Pancreatic Cancer Literature Abstraction 27 Step 1 - Literature Search results 136 hits 30 studies identified for detailed analysis 22 with Gemcitabine only control Use this data to develop the prior inputs to our model and check the choice of endpoint makes sense 28 Meta Analysis of Gemcitabine 6 Month Survival Rate (P2 endpoint – metastatic + LA subjects) Meta Analysis Forest Plot 6 Month Surv iv al (Proportion) (Gemcitabine Control) Study Reference 31 29 33 32 3 17b 30 18 2 1 12 14 4 7 20 16 19b 19b 15 21 9 6b Heterogeneity Q = 39.1 p= 0.0096 I-Sq = 46.3% Estimate [95% CI] 0.480 [ 0.312, 0.648] 0.560 [ 0.412, 0.708] 0.600 [ 0.452, 0.748] 0.440 [ 0.293, 0.587] 0.470 [ 0.350, 0.590] 0.500 [ 0.383, 0.617] 0.570 [ 0.467, 0.673] 0.510 [ 0.409, 0.611] 0.410 [ 0.322, 0.498] 0.450 [ 0.373, 0.527] 0.600 [ 0.524, 0.676] 0.620 [ 0.544, 0.696] 0.510 [ 0.435, 0.585] 0.570 [ 0.495, 0.645] 0.520 [ 0.446, 0.594] 0.520 [ 0.456, 0.584] 0.530 [ 0.471, 0.589] 0.420 [ 0.362, 0.478] 0.520 [ 0.462, 0.578] 0.490 [ 0.432, 0.548] 0.490 [ 0.437, 0.543] 0.490 [ 0.439, 0.541] Fixed Random Weight 0.9% 1.2% 1.2% 1.2% 1.8% 1.9% 2.4% 2.6% 3.3% 4.4% 4.5% 4.5% 4.6% 4.7% 4.7% 6.4% 7.5% 7.6% 7.7% 7.7% 9.4% 10.0% 0.508 [ 0.492, 0.524] 0.511 [ 0.487, 0.534] 0.2 Program: pan_group_forest.sas Output: meta_group_m6.cgm (Date Generated: 07JAN2010: 13:48) 29 0.3 0.4 0.5 0.6 0.7 0.8 Gemcitabine 6m survival predictions for the exact population (control priors) Number of Studies Estimate 95% CI All studies 22 0.511 (0.487, 0.534) Metastatic Subgroups 6 0.464 (0.400, 0.514) Metastatic Subgroups and Studies with ≥ 90% Metastatic Subjects 9 0.472 (0.431, 0.514) Meta-Regression Estimating for 100% Metastatic Subjects 22 0.482 (0.445, 0.517) Method 30 Do the endpoints predict survival? SPEED Pancreatic Cancer All Treatment Group Responses Median Ov erall Surv iv al by Median Progression-Free Surv iv al 10 9 Median OS v Median PFS Median OS(Months) 8 7 6 5 4 3 2 0 3 4 Cancer 5 SPEED Pancreatic All TreatmentMedian Group PFSResponses (Months) Median Ov erall Surv iv al by 6 Month Surv iv al 1 2 6 7 8 The size of the circles are proportional to the number of patients. 10 Program: panc_group_plots.sas Output: all_os_pfs.cgm (Date Generated: 17DEC2009: 11:21) 9 Median OS v 6 Month Surv Rate Median OS(Months) 8 7 6 5 4 3 2 0.0 0.1 0.2 0.3 0.4 0.5 6 Month Survival The size of the circles are proportional to the number of patients. Program: panc_group_plots.sas Output: all_os_m6.cgm (Date Generated: 17DEC2009: 11:21) 31 0.6 0.7 0.8 0.9 1.0 2 Developing the Endpoint Relationships Phase 2 Endpoints – 6 month survival rate – PFS Phase 3 Endpoint – Overall Survival Predict the relationship between the treatment difference in the P2 endpoint with the treatment difference in the P3 endpoint 32 OS Hazard Ratio and 6M Survival Rate Difference SPEED Pancreatic Cancer Phase 2 - Phase 3 Endpoint Relationship Ov erall Surv iv al by 6 Month Surv iv al 1.6 Overall Survival (HR) 1.4 1.2 1.0 0.8 0.6 0.4 -0.4 -0.3 -0.2 -0.1 0.0 Difference in 6 Month Survival The size of the circles is inversely proportion to the OS SE(LnHR) Program: panc_reln_plots.sas Output: diff_oshr_6m.cgm (Date Generated: 14JAN2010: 14:17) 33 0.1 0.2 0.3 0.4 OS Hazard Ratio and PFS Hazard Ratio SPEED Pancreatic Cancer Phase 2 - Phase 3 Endpoint Relationship Ov erall Surv iv al by Progression-Free Surv iv al 1.6 Overall Survival (HR) 1.4 1.2 1.0 0.8 0.6 0.4 0.4 0.6 0.8 1.0 PFS (HR) The size of the circles is inversely proportion to the OS SE(LnHR) Program: panc_reln_plots.sas Output: diff_oshr_pfshr.cgm (Date Generated: 14JAN2010: 14:17) 34 1.2 1.4 1.6 Do we need to adjust the HR for other predictors of OS Hazard Ratio? SPEED Pancreatic Cancer Ov erall Surv iv al Hazard Ratio by Percentage of Metastatic Subj ects 1.6 OS HR v % Metastatic Overall Survival (HR) 1.4 1.2 1.0 0.8 0.6 0.4 40 50 SPEED Pancreatic Cancer 70 Ov60erall Surv iv al Hazard Ratio by Percentage of ECOG 0/1 Subj ects Percentage of Metastatic Subjects 80 90 100 80 90 100 1.6 of the circles is inversely proportion to the OS SE(LnHR) The size Program: panc_reln_plots.sas Output: diff_oshr_meta.cgm (Date Generated: 14JAN2010: 14:17) OS HR v % ECOG 0/1 Overall Survival (HR) 1.4 1.2 1.0 0.8 0.6 0.4 40 50 60 70 Percentage of ECOG 0/1 Subjects The size of the circles is inversely proportion to the OS SE(LnHR) Program: panc_reln_plots.sas Output: diff_oshr_ecog.cgm (Date Generated: 14JAN2010: 14:17) 35 Meta-Regression for OS Hazard Ratio SPEED Pancreatic Cancer Meta Regression for Ov erall Surv iv al HR by Month 6 Difference in Surv iv al Rate 1.6 Overall Survival Hazard Ratio 1.4 1.2 1.0 0.8 0.6 0.4 -0.20 -0.15 -0.10 -0.05 0.00 0.05 0.10 0.15 Month 6 Difference The radius of the circle is inversely proportion to SE(LHR) Predicted Mean HR (95%CI) for a fixed Month 6 difference is shown Program: panc_reln_model.sas Output: mreg_os_m6diff.cgm (Date Generated: 20JAN2010: 12:10) 36 Parameter Posterior Distribution Mean SD 95% Cr.Int. b (slope) t (random effects SD) -1.777 0.0233 0.3419 0.01759 (-2.518, -1.164) (0.0015, 0.0664) 0.20 Step 2: The expected difference in 6m survival between treatments Ph 2 RCT Gem Lit 6m Gem P2 6m AMG P2 6m This is our best estimate of the trt difference Equivalent to our Ph2 population diff 6m survival Eg. If the gem P2 < expected then the 6m rate in both the gem and AMG arms are reduced by the same amount 6m survival 37 The observed difference is maintained. The variance alters 0.2 0.0 -0.2 -0.4 Rsq=75.3% Phase 2 result modulated variance for the control behaviour -0.8 -0.6 log Hazard Ratio (Overall Survival) 0.4 Step 3: Predicting the OS result at the EOP2 (the posterior distribution for the OS HR) -0.4 -0.2 0.0 38 Difference in 6 mnth Survival Rate 0.2 0.4 Step 2: The expected difference in 6m survival between treatments Gem Lit 6m Gem P2 6m AMG P2 6m Prior Belief of the ability of the drug 6m survival Sceptical Uninformative Optimistic Diff is pulled towards the prior belief 39 Prior Distributions for Phase 2 Treatment Difference Uninformative Prior No prior knowledge assumed, but somewhere between a 50% reduction and a 50% increase. -60 -45 -30 -15 0 15 30 45 60 Phase 2 Treatment Difference Sceptical Prior Expect treatment difference to be zero, but a 20% chance that difference could be > 15%. -60 -45 -30 -15 0 15 30 45 60 Phase 2 Treatment Difference Optimistic Prior Expect treatment difference to be 15%, but a 20% chance that difference could be negative. -60 -45 -30 -15 0 15 Phase 2 Treatment Difference 40 30 45 60 Light Sceptical/Optimistic: Downweight above to 25% P2 sample size Incorporating prior belief ultimately allows us to build up a bound for the PoS If we observe <X% diff in our P2 study the PoS is only Y% for even the most optimistic of you. If we observe >X% diff in our P2 study the PoS is still Y% for the real sceptics 0.2 0.0 -0.2 -0.4 Rsq=75.3% Expected Diff in 6m survival rate -0.6 log Hazard Ratio (Overall Survival) 0.4 Step3: Converting the P2 to the P3 endpoint -0.8 S -0.4 -0.2 U O 0.0 41 Difference in 6 mnth Survival Rate 0.2 0.4 Phase 3 Endpoint Treatment Difference 6M - OS Relationship Probability Probability true HR<1 (i.e we have an effect) 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 Overall Survival Hazard Ratio 42 1.4 1.5 1.6 1.7 1.8 1.9 2.0 Step 3: Posterior Distribution for OS Hazard Ratio Posterior Distribution for OS Hazard Ratio by Month 6 Difference in Surv iv al Rate and Type of Prior Distribution 2.0 1.8 OS Hazard Ratio (95CrI) 1.6 1.4 1.2 1.0 0.8 0.6 0.4 Prior Distribution Non-Informative Optimistic Sceptical 0.2 0.0 -0.20 -0.15 -0.10 -0.05 0.00 Month 6 Difference Program: panc_posterior_reln_plots.sas Output: post_os_m6diff_all.cgm (Date Generated: 02FEB2010: 14:57) 43 0.05 0.10 0.15 0.20 Determining the PoS in Phase 3 PoS will depend upon the Phase 3 study design – The power, significance level and hypothesis drive the number of events required – The more events you include in your analyses, the smaller the variance and the more likely you are to reach statistical significance We draw a sample result from the Phase 3 endpoint distribution and apply the variance associated with the chosen Phase 3 study design. We then determine if the result is statistically significant. We repeat the process multiple times and determine the proportion of times that we see a statistically significant result. This is the probability of success. 44 Step 4: Probability of Success 100 We can also change the Phase 3 design to optimize the PoS – do a larger study Probability of Success in Ph 3 90 80 70 Better than industry average Uninformative Sceptical 60 50 Optimistic Light Sceptical 40 Light Optimistic 30 20 10 0 -15 -10 -5 0 5 10 15 Diff in 6m Survival Rate in Phase 2 (Test - Control) 45 20 Summary The PoS we select will depend upon the current business situation and our strategy to risk The industry average for oncology is a 40% success rate in P3 If we observe < 7% difference the PoS < 40% for the optimists If we observe >10% difference then the PoS > 40% for the sceptics 46 The power of the research: Examples 0.2 -0.2 0.0 Rsq=0.3% -0.4 log Hazard Ratio (Overall Survival) 0.4 Using PFS for Soft Tissue Sarcoma -0.4 47 -0.2 0.0 log Hazard Ratio (PFS) 0.2 0.4 The power of the research: Examples Log Hazard Ratio (Overall Survival) ORR in NSCLC 0.4 0.2 0.0 Rsq=23.9% -0.2 -0.1 0.0 0.1 Difference in Objective Response Rate (%) 48 0.2 Log Hazard Ratio (Overall Survival) The power of the research: Examples PFS in NSCLC is not so good 0.4 0.2 0.0 Rsq=42.6% -0.2 -0.4 -0.2 0.0 Log Hazard Ratio (Progression Free Survival) 49 0.2 0.4 0.0 0.5 PFS in CRC seems good Rsq=65.5% -0.5 log Hazard Ratio (Overall Survival) 1.0 A good one to finish on -0.8 50 -0.6 -0.4 -0.2 0.0 0.2 Log Hazard Ratio (Progression Free Survival) 0.4 0.6