Interpretable deep learning models for translational medicine Renewal

Abstract Conventional precision oncology concentrates on applying molecularly targeted drugs based on genomic markers from a tumor. Such a genome-informed approach has limitations, mainly because it covers around 15% of all cancer patients. While chemotherapies remain the backbone of oncology, their application is not personalized. In this study, we proposed to develop an advanced artificial intelligence (AI) framework for predicting cancer cell sensitivity to chemotherapies. We will apply the framework to solve a real-world clinical problem: guiding the precision application of oxaliplatin in treating colon cancer adjuvant therapy. Our framework consists of two phases: A representation-learning phase, which involves developing and applying advanced AI models that can infer and represent the state of cellular signaling systems of cancer cells by mining diverse omics data of cancers. A cell-state- oriented drug sensitivity prediction phase, which involves using advanced classification models to utilize the inferred cancer cell states to predict the drug response of different cancer cells. Importantly, we will address an unmet clinical need in treating colon cancer patients as a real-world use case. With estimated 1.9 million incidences, colorectal cancer (CRC) is the 3rd most common cancer and the 2nd leading cause of cancer death. Patients with resected high-risk stage II/III colon cancer (CC) receive adjuvant chemotherapy to prevent cancer recurrence. A regimen consisting of 5-fluorouracil, leucovorin, and oxaliplatin (FOLFOX) is the standard of care for CC adjuvant therapy. It is estimated that including oxaliplatin in the regimen benefits 5% - 6% of patients but subjects ~90% of patients to life-impairing neurotoxicity. A significant effort in oncology is to identify the patients who would not benefit from oxaliplatin and save them from unnecessary side effects (de-escalation). In this study, we will collect large-scale omics data from colon cancer patients. We propose advanced causal discovery and deep learning to infer the state of signaling systems of cancer cells and further develop advanced individualized treatment effect prediction models to guide oxaliplatin treatment.