Modeling cancer dependency with deep graph models

Abstract A fundamental premise for precision oncology is a catalog of diverse actionable targets that could enable personalized treatment. Large scale Genome-wide lost-of-function screens such as cancer dependency map have systematically identified single gene vulnerabilities in numerous cell lines. However, it remains challenging to scale such analyses to many clinical samples and untangle molecular networks underlying observed vulnerabilities. We developed a deep learning framework, DepGPS, combing graph neural networks with transformers to model the network interactions underlying tumor vulnerabilities. Our model demonstrated an improved ability to predict context-specific vulnerabilities over existing models and showed a higher responsiveness in perturbation analysis. Furthermore, perturbation induced dependency changes by our model demonstrated utility to support context-aware identification of synthetic lethal genes. Overall, our model represents a valuable tool to extend tumor vulnerability analyses to broader range of subjects and could help to decipher molecular networks dictating context-specific tumor vulnerabilities.