MoLRSI: An AI Agent That Improves Its Molecular Discovery Workflows

Introduction

Molecular discovery is difficult not only because the chemical search space is enormous, but also because useful candidates must often satisfy competing requirements. A molecule may need strong activity, low toxicity, structural novelty, and adequate selectivity at the same time. Existing AI agents can automate protocols designed by experts, yet automation alone does not tell the system how to revise those protocols when the task, data, or experimental feedback changes.

MoLRSI, introduced in the paper, addresses this limitation by treating the discovery workflow itself as an optimization target. The agent searches for molecules while also evaluating and refining the models, workflow components, and search policies used to find them. Its experience is stored across tasks, allowing later campaigns to benefit from earlier successes and failures.

Key points

• Workflow optimization is part of the search. MoLRSI uses workflow performance, refinement outcomes, and execution feedback to modify how a discovery campaign is run, rather than following a fixed pipeline.
• Persistent cross-task memory. Strategies that worked in previous campaigns can inform later searches, reducing the need to design every workflow from scratch.
• Multi-objective antibiotic discovery. In the reported antibiotic setting, the agent balanced antibacterial activity, cytotoxicity, and structural novelty. Using the same training data and screening library as a previous expert-led workflow, it reached a 15.9% experimental hit rate, compared with 8.7% for the earlier approach.
• Peptide–MHC binder design. For de novo design, MoLRSI jointly considered target binding and specificity. It generated eight experimentally validated hits against MART-1 and SARS-CoV-1 pMHCs, showing preferential binding over the tested off-targets. The supplied material reports that Claude Science failed in this comparison.

Why it matters

The broader contribution is a shift in how AI agents are used for scientific discovery. Instead of viewing an agent as a system that executes a human-authored recipe, MoLRSI treats it as a system that can improve the recipe itself. Experimental and operational signals become inputs to the next round of workflow design, while cross-task memory gives the agent a way to accumulate procedural knowledge.

This does not establish that the method will work across all drug discovery problems. The evidence described here comes from two application settings, and questions about reproducibility, workflow transparency, screening scale, and performance on new targets remain open. Still, the results illustrate an important direction for AI for Science: progress may depend not only on better predictive models, but also on agents that can diagnose failed searches and improve the strategies used to conduct them.

Source: Hugging Face Daily Papers