Designing Agentic AI Systems for Human-Guided Scientific Discovery in Computational Drug Discovery

The integration of biomedical foundation models with expert human insight offers great potential for computational drug discovery, yet researchers continue to face challenges in composing usable, end-to-end workflows from diverse public databases and open-source tools. We present an agentic AI system design that streamlines scientific exploration through intelligent workflow composition, dynamic agent coordination, and adaptive user interaction. Rather than retrofitting existing tools into static pipelines, our system architecture enables flexible composition and adaptation of workflows tailored for users’ guidance. The system integrates AI models, open-source tools, interactive visualizations, and a chat interface to support iterative refinement and hypothesis-driven exploration. We highlight both the architectural design of the system and key lessons learned throughout its development, offering insights to inform future agentic AI systems aimed at enhancing human-AI collaboration in scientific discovery.