Learning to Adapt - Deep Reinforcement Learning in Treatment-Resistant Prostate Cancer

Abstract

Standard-of-care treatment regimes have long been designed to for maximal cell kill, yet these strategies often fail when applied to treatment-resistant tumors, resulting in patient relapse. Adaptive treatment strategies have been developed as an alternative approach, harnessing intra-tumoral competition to suppress the growth of treatment resistant populations, to delay or even prevent tumor progression. Following recent clinical implementations of adaptive therapy, it is of significant interest to optimise adaptive treatment protocols. We propose the application of deep reinforcement learning models to provide generalised solutions within adaptive drug scheduling, and demonstrate this framework can outperform the current adaptive protocols, extending time to progression by up to a quarter. This strategy is robust to varying model parameterisations, and the underlying tumor model. We demonstrate the deep learning framework can produce interpretable, adaptive strategies based on a single tumor burden threshold, replicating and informing a novel, analytically-derived optimal treatment strategy with no knowledge of the underlying mathematical tumor model. This approach is highly relevant beyond the simple, analytically-tractable tumor model considered here, demonstrating the capability of deep learning frameworks to help inform and develop treatment strategies in complex settings. Finally, we propose a pathway to integrate mechanistic modelling with DRL to tailor generalist treatment strategies to individual patients in the clinic, generating personalised treatment schedules that consistently outperform clinical standard-of-care protocols.