Developing an AI Co-Scientist for Autonomous Scientific Reasoning in Biomedicine
Biomedical research faces mounting challenges from its inherent complexity and the exponential expansion of scientific literature, making it increasingly difficult for researchers to generate novel hypotheses and achieve breakthrough discoveries through traditional manual approaches that rely heavily on expert curation and domain knowledge. To revolutionize this process, we develop OmniCellAgent, an AI co-scientist that autonomously conducts scientific reasoning by leveraging BioMedGraphica and Graph-Language Foundation Models (GLFMs) in a sophisticated closed-loop architecture that seamlessly integrates data querying, model inference, and explanation generation to produce interpretable hypotheses for critical applications such as target prioritization and drug-combination discovery. Our approach enhances the agent orchestrator through reinforcement learning with reward modeling to optimize task planning while establishing comprehensive evaluation benchmarks that incorporate expert knowledge, literature alignment, and human feedback to assess scientific reasoning performance in biomedical AI systems. See the full paper for detailed methodology and results.