ATHENA-R1: AI Agent for Biomedical Treatment Reasoning

A new AI agent named ATHENA-R1 has been developed to perform complex treatment reasoning by integrating vast biomedical knowledge and iteratively using a universe of 212 specialized tools. Treatment reasoning, which involves weighing disease context, comorbidities, medications, and contraindications, has historically been challenging for AI due to its iterative and evidence-grounded nature. ATHENA-R1 addresses this by identifying missing information, selecting and executing relevant tools, and incorporating evidence at each step. The agent was trained using a novel two-level self-learning framework: multi-agent systems generated tools, tasks, and reasoning trajectories for supervised fine-tuning, followed by reinforcement learning with scientific feedback to reward reasoning quality. Across five benchmarks, including 3,168 drug reasoning tasks and 456 patient treatment cases, ATHENA-R1 achieved 94.7% accuracy on open-ended drug reasoning and 82.9% on treatment reasoning, significantly outperforming existing language models and tool-use systems. Blinded expert evaluations from rare disease organizations favored ATHENA-R1, and physicians rated it highly for complex cardiovascular and infectious-disease cases. Furthermore, adverse-event hypotheses generated by ATHENA-R1 were validated in electronic health records, showing statistically significant odds ratios. This work demonstrates that treatment reasoning can be reframed as a learnable process of iterative evidence gathering, effectively trainable through reinforcement learning.