Replacing  Drug Discovery  âExpert in The Loopâ with AI Agent
M. Pandey, T. Sajed, E. Manskaia, I. Semenov, F. Ban, J. Smith, A. Cherkasov*
*Department of Urologic Sciences at the University of British Columbia and Vancouver Prostate Centre
The CACHE (Critical Assessment of Computational Hit-finding Experiments) Challenge 7 provides a rigorous benchmark for evaluating computational approaches to small-molecule hit identification. Here, we present an autonomous, AI-powered drug discovery agent developed to address this challenge through the integration of machine learning and molecular docking within an interactive, reasoning-capable framework. Our agent is designed to streamline the structure-based hit-finding pipeline by combining molecular docking with machine learning models to score and prioritize candidate compounds based on predicted binding affinity and interaction quality. A distinguishing feature of the system is its capacity to reason about molecular interactions by interpreting docking results, evaluating binding pose plausibility, and generating structural hypotheses. Crucially, this reasoning is accessible through natural language, enabling researchers to actively guide, question, and refine the discovery process without requiring deep computational expertise. The agent is further designed to explore beyond conventional screening libraries, with the goal of identifying novel scaffolds that may represent underexplored regions of chemical space relevant to the target. By coupling AI-driven reasoning with established docking methodologies, the system aims to accelerate early-stage drug discovery while preserving interpretability and user control.We will present our computational approach, the agent's architecture, and preliminary findings as part of the CACHE 7 evaluation. This work underscores the potential of conversational AI agents as collaborative tools in computational drug discovery, enabling more iterative and hypothesis-driven hit identification campaigns.