Charles：一个用于癌症的自我批判式智能体AI药物发现分析师

Background: Agentic AI will be members of scientific teams of the future. AI hallucinations and unreliable and untraceable public data cast shadows on such a role. We developed what is to our knowledge the first self-critical, self-correcting agentic-AI drug discovery analyst: Charles. It supports information synthesis and hypothesis generation for cancer drug discovery while maintaining scientific robustness. Methods: Charles is an agentic GPT-based LLM. Building Charles consisted of fact-constrained, imitation-based teaching exercise led by the human PI. A human supervisor carefully optimized the environment to constrain Charles to reliable sources and prevent retrieving external knowledge. An adversarial AI agent acts as the self-critical ‘conscience' to test the veracity of Charles's answers against the true data. We further evaluated this by injecting decoy raw data, allowing quantitative evaluation of the models' veracity. The trained model was used to synthesize factual target summaries. We then developed a multi-agent LLM that leverages these data-grounded summaries to answer real-word drug discovery questions. A planner agent digests the user's questions and orchestrates a coordinated plan of action to direct the specialist agents (bio-/chemo-/systems etc). Finally, a critical AI agent fact-checks the responses from specialist agents, reporting them to the planner agent for further refinement of actions. Together, these modules form Charles. Results: We demonstrate that Charles' responses are strictly grounded in the curated data. Initial evaluations show Charles reproduces all adversarial decoys indicating no outside leakage. To further evaluate the reasoning and factual accuracy of Charles, we utilized a question-answer dataset that Charles answers through interpretation of the >1000 protein summaries. Using the self-critical agent, we evaluated both the summaries and the answers to questions. The initial evaluation of the multi-agent framework shows 99% accuracy in Charles's responses. The planner successfully recruited the appropriate specialized agents while the critical agent showed promising behavior in detecting semantic and factual inconsistencies to the planner. The output summaries are publicly available via canSAR.ai. Conclusions: Charles represents a self-critical, drug discovery-trained, AI analyst capable of synthesizing multimodal information, evaluating its own outputs, and refining the next steps based on that. By combining the reasoning power of LLM-based interactions with the assurance of verified curated data, Charles provides reliable access to complex oncological knowledge while minimizing hallucinations and enabling researchers to make data-driven decisions with more confidence in cancer research.