MoLRSI: A Recursive Self-Improving AI agent for Molecular Discovery

MoLRSI. A Recursive Self-Improving AI agent for Molecular Discovery. By Yikun Zhang and seventeen others, from Northeastern University, the Broad Institute of MIT and Harvard, and twelve other institutions. Published on the thirtieth of October.

Molecular discovery means navigating vast combinatorial search spaces while balancing biological objectives that pull against each other. AI agents can already automate molecular design protocols that experts define. But automation alone leaves two questions open: how to optimise a workflow for a new task, and how to get better at developing workflows through experience. Automating a protocol an expert wrote is one thing. Improving that protocol, and the way protocols get developed, is another.

MoLRSI is an AI agent built to address both. It combines systematic optimisation of workflows and model architectures with a persistent memory that carries across tasks, and it uses them together to recursively improve its own search policy for molecular discovery. The recursion is the point. The agent improves not only the workflows it builds, but also its strategy for building them.

Three kinds of signal guide each subsequent search: how a workflow performs, the outcomes of refinements, and feedback from execution. Because the memory persists across tasks, what was learned in one campaign shapes the search in the next. The optimisation covers model architectures as well as workflows, and the same feedback lets the agent refine two things at once: the workflows themselves, and its strategy for developing them.

MoLRSI is evaluated in two challenging settings, each requiring a balance of competing properties. The first is antibiotic discovery. There the agent developed a screening workflow that balances antibacterial activity, cytotoxicity and structural novelty. It achieved an experimental hit rate of fifteen point nine percent, against eight point seven percent for a previous expert-led workflow that used the same training data and the same screening library. With data and library held the same, the comparison is between workflows.

The second setting is de novo design of peptide-MHC binders. MoLRSI developed workflows that consider target binding and specificity jointly, and produced eight experimentally validated hits against MART-1 and SARS-CoV-1 pMHCs, with preferential binding over the off-targets tested. In the same setting, the abstract reports, Claude Science completely failed.

Both results are experimental, and both involve objectives that trade off against each other. What MoLRSI contributes is a way of treating workflow design itself as something to optimise and remember, so that each discovery campaign leaves the agent better equipped for the one that follows. The authors present this as experience-guided workflow optimisation, addressing competing objectives while improving the search strategies used in later campaigns.