在多模态肿瘤学基础模型上采用组相对策略优化进行机制感知的癌症治疗规划

Modern multimodal oncology foundation models (OFMs) can predict patient trajectories and simulate treatment effects from clinical, genomic, transcriptomic, and histologic data, but they remain largely black boxes: they rarely explain which mechanisms drive risk or how to modulate those mechanisms with feasible interventions. We trained a multimodal OFM on over 1.2 million cancer patients with longitudinal clinical records, tumor DNA and RNA sequencing, and H&E imaging. The model learns a joint latent representation of each patient's evolving disease state and can forecast outcomes and simulate counterfactual treatment trajectories. We then added a mechanism-aware reasoning layer that turns the OFM from a predictor into an engine for mechanism-based therapy design. Our key innovation is group-relative policy optimization (GRPO), a reinforcement-learning framework that links mechanism states inferred by the OFM (e.g., pathway activation patterns and resistance programs), a drug intervention space built from curated drug-target relationships and perturbation signatures, and an explicit clinical reward. Rather than optimizing an abstract objective for a single patient, we define the reward as improvement in predicted outcomes for cohorts of patients who share similar disease state and driver mechanisms. For each disease context, we identify mechanisms associated with poor outcome, link them to candidate drugs and observed outcomes in real-world data, and then use GRPO to evaluate policies (sets of mechanism-level interventions) by asking whether applying a policy to that cohort improves predicted survival or delays progression relative to matched standard-of-care controls. This group-relative reward stabilizes learning, avoids overfitting to idiosyncratic outliers, and aligns the learned policies with how clinicians naturally reason about “patients like these.” To make outputs biologically and clinically interpretable, we map proposed mechanism shifts to existing drugs and combinations, to mechanism-defined patient clusters whose outcomes are driven by similar latent programs, and to de novo mechanism opportunities where the optimal policy improves outcomes but no current drug fully explains the effect, flagging potential targets for discovery. Across multiple indications, GRPO recovers known mechanism-therapy relationships, identifies patient subgroups whose outcomes are improved when therapies they receive align with GRPO-suggested mechanism policies compared with matched patients receiving discordant therapies, and proposes novel mechanism-therapy hypotheses involving combined or sequential modulation of programs that are not jointly targeted today. This turns a large-scale OFM into a mechanism-aware decision engine that reasons about resistance and treatment opportunities in the same latent space used for outcome prediction.