Document resource
Background Numenos developed CURE AI, a large clinicogenomic foundation model (LCGM) that predicts treatment outcomes. Applied to a clinical trial, CURE AI results in a continuous benefit prediction of patients based on the most important and predictive clinical and genomic factors, which can be utilized to inform biomarker and target discovery and improve clinical trial eligibility criteria.Methods We finetuned the CURE AI foundation model on phase 2 and 3 lung cancer immunotherapy clinical trials (OAK and POPLAR, comparing atezolizumab to docetaxel in non-small cell lung cancer) as well as a separate model finetuning on the phase 3 JAVELIN Renal 101 clinical trial (comparing avelumab plus axitinib to sunitinib in clear cell renal cell carcinoma) to create CURE Lung Cancer and CURE Renal Cancer models, respectively. Using CURE Lung Cancer, we predicted trial outcomes in renal cancer, and using the CURE Renal Cancer, we predicted outcomes in lung cancer using only clinical and genomic information available at trial enrollment.Results We have previously demonstrated that CURE AI, finetuned on phase 2 data, predicts phase 3 outcomes. Here, we applied CURE Lung Cancer to a clear cell renal cancer clinical trial (JAVELIN Renal 101), which improved trial significance >25x and improving time to trial significance by > 6 months while retaining 80% of the total trial population. Furthermore, CURE AI, finetuned on ccRCC immunotherapy data (CURE Renal Cancer) predicts immunotherapy response in lung cancer, converting the POPLAR trial from a negative PFS endpoint (p = 0.21) to a positive endpoint (p < 0.01) while retaining 65% of the total trial population. We then utilized the CURE AI models, finetuned on lung or renal immunotherapy data, to inform stratification of real world data. In a pan-cancer analysis, CURE AI successfully stratified patients by immunotherapy response (poor predicted immunotherapy response: low grade glioma, prostate, pancreas; high predicted immunotherapy response and high: melanoma, hepatocellular, leukemia and lymphoma) enabling informed indication expansion decisions. We further show how CURE AI informs combination therapy selection by showcasing how we identify TIGIT as a weak immunotherapy target.Conclusions CURE AI-refined eligibility criteria allows for cross-cancer immunotherapy response predictions, allowing for biologically-informed clinical trial inclusion criteria. CURE AI allows for >20-50% trial size reduction while accelerating the time to trial significance by 6-18 months. CURE AI has significant potential to lead to advancements in clinical trial implementation as well as new target selection for combination therapies with an IO backbone.Trial Registration NCT02684006 NCT02008227 NCT01903993 All data are released.