Abstract 5359: Translating state-of-the-art deep learning predictions of treatment efficacy to clinical practice

Abstract Despite the clinical success of immune checkpoint blockade therapies, in many patients the disease becomes unresponsive or resistant to treatment. Attempts to predict treatment efficacy and identify determinants of response have suffered from limited accuracy. Here, we introduce a state-of-the-art deep learning method that better captures the complex relationships among the host, the immune system, and tumor biology. Our method, an explainable deep learning framework, is based on transformer architecture that combines data with different feature sets and clinical endpoints for survival prediction or classification. The framework includes(1) a new loss function based on Harrell’s concordance-index, (2) an explainability module providing importance score between features based on mutual contribution to predictions, and (3) a transfer learning strategy that enables leveraging diverse clinical data sets in the public or private domain. Using seven data sets comprising more than 140,000 patients from immunooncologic (IO), targeted, and chemotherapy treatments, our method consistently outperformed other methods previously described in the literature, including CoxPH and random survival forest (RF). For example, our framework achieved a concordance index of 0.66 (±0.04) vs. 0.60 (±0.04) of the second-best method (RF) on MYSTIC IO arms using only clinical data. Utilizing our explainability module, we identified key features driving response prediction that are consistent with those in previous publications. For example, in the Chowell et al. data set, we identified albumin, neutrophil-to-lymphocyte ratio, prior-chemo-treatment, and tumor mutation burden as the most important features. We also identified in sparse mutation calls of 469 genes from the Samstein et al. data set, functional modules of several genes, each with a strong predictive power. For example, the functional module with the highest hazard ratio (HR) (1.43) was the gene pair KEAP1 and STK11, which are well-established drivers of resistance to multiple therapies. We further validated these results across multiple data sets: The Cancer Genome Atlas HR = 1.23), GENIE/anti-PD-L1 (HR = 1.06), and MYSTIC/anti-PD-L1/CTLA-4 (HR = 1.39). Another example is a functional module related to adaptive immunity, comprising the genes AKT2, BTK, CDC73, HLA-B, IKBKE, INPPL1, RFWD2, TRAF2, and WHSC1. This module stratified patients with an HR of 0.58 within the Samstein et al. data set and 0.42 within the independent validation data set. We describe a new deep learning method with state-of-the-art performance in survival prediction and the potential to uncover biological and clinical insights related to disease response and resistance. Our framework simplifies the process of translating complex AI models to clinical practice and may accelerate the identification of targetable drivers of resistance and response. Citation Format: Gustavo Arango-Argoty, Etai Jacob. Translating state-of-the-art deep learning predictions of treatment efficacy to clinical practice. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5359.