Multi-target Parallel Drug Discovery with Multi-agent Orchestration

AI Agent, Fuad Al Abir, Sixue Zhang, Jake Yue Chen

Keywords: Agent, Multi-Agent Systems, AI4Science, Drug Discovery, Alzheimer's Disease

TL;DR: We built a team of AI agents that autonomously identifies novel targets and designs potential drug candidates.

Abstract: The immense cost and high failure rates of traditional drug discovery necessitate a shift towards more systematic, automated approaches. While specialized AI tools have excelled at individual tasks, their integration into a cohesive, end-to-end discovery workflow remains a significant challenge. We introduce a modular, multi-agent framework that autonomously navigates the early-stage drug discovery pipeline, from target identification to the generation of optimized hit candidates. Our system orchestrates specialized agents that synergize AI Agents for literature mining and generative chemistry with robust machine learning classifiers for bioactivity and ADME/Tox prediction. We demonstrate the system's capabilities by applying it to Alzheimer's Disease, identifying and generating novel inhibitors for five protein targets—SGLT2, CGAS, SEH, HDAC, and DYRK1A, and successfully generated novel molecular scaffolds with high predicted potency and favorable drug-like properties for four of the targets. The framework's failure to build a reliable model for the data-scarce target, CGAS, highlights a key limitation: the performance of autonomous systems is fundamentally tethered to the quality and availability of the underlying data. Our work presents a transparent blueprint for an integrated discovery engine and provides a realistic perspective on the current capabilities of AI agents in science, suggesting they operate most effectively within a human-in-the-loop paradigm where expert oversight guides data curation and model validation. The code is publicly available at https://github.com/UAB-SPARC/agentic-drug-discovery.

Submission Number: 333