Abstract 1467: Mechanism-aware cancer therapy planning with group-relative policy optimization on a multimodal oncology foundation model.

Key Points

• The aim is to enhance cancer treatment planning by explaining mechanisms that drive patient outcomes using a novel framework.
• Trained a multimodal oncology foundation model on extensive patient data including clinical, genomic, and imaging metrics.
• Implemented a mechanism-aware reasoning layer to transform the model into a tool for therapy design.
• Developed group-relative policy optimization to assess and optimize treatment strategies based on patient cohorts.
• GRPO identifies mechanisms leading to poor outcomes and associates them with potential drug interventions.
• Demonstrated improved predicted survival and progression delay for cohorts using GRPO-informed therapies compared to standard care.
• Recovered known relationships between mechanisms and therapies, proposing new treatment hypotheses for targeted patients.