Artificial intelligence in target discovery: CURE AI prediction of TIGIT and PD-L1 as immunotherapy targets in NSCLC.

Abstract

e13692 Background: We developed CURE AI (Clinical trials Uncovering Real Efficacy Artificial Intelligence), a new type of machine learning platform that measures the individualized benefit of a therapeutic intervention over standard of care. To build CURE AI, we trained a foundation model utilizing clinical and multi-omics data from > 100,000 oncology patients using a proprietary deep learning architecture and training schema developed by Numenos. CURE AI is designed to understand the interplay between all baseline clinicogenomic information to predict the magnitude of benefit of therapy for each patient enrolled on a clinical trial . The foundation model allows CURE AI to predict the outcome (such as PFS) for a therapy the patient did not receive. Assigning each patient an accurate predicted outcome to both the investigational therapy and standard of care improves the characterization of true responders and true non-responders. Once patients are classified by CURE AI by predicted therapy response, characterization of genes, pathways, and mutations that drive response or resistance to investigational therapies can be determined with high confidence. We sought to determine if TIGIT, which has significant clinical trial investment, would have been considered a high priority target for non-small cell lung cancer (NSCLC) using CURE AI. Methods: We analyzed the phase 2 POPLAR and phase 3 OAK trials, which compared atezolizumab to docetaxel in advanced non-small cell lung cancer patients (combined n = 892 for patients with RNA sequencing). CURE AI calculated the predicted response to both atezolizumab and docetaxel for every patient to order all patients by the best predicted response to atezolizumab to best predicted response to docetaxel. q-values were calculated for gene expression differences that define immunotherapy benefit (comparing atezolizumab-benefitting to docetaxel-benefitting patient groups). Results: CURE AI identified multiple immune-pathway related genes that predict immunotherapy benefit. PD-L1 was identified as a top hit, with a q-value of < 10 -16 ,serving as an internal positive control. TIGIT was outside of the top 1500 significant genes with a q-value of 0.04, suggesting that TIGIT has a weak signal that does not generalize to all patients. Subgroup analysis found positive associations with adenocarcinoma histology and male gender, suggesting improved response to anti-TIGIT therapy in these groups. Conclusions: CURE AI predicted the limited clinical benefit of anti-TIGIT therapy in all-comers and the high clinical benefit for PD-L1 using data from the earliest immunotherapy trials. TIGIT should be further assessed in men with adenocarcinomas. The predictive capability of CURE AI can be used for efficient resource allocation, accelerate the development of effective cancer treatment combinations, and inform clinical trial eligibility selection.