An agent carries search experience from one campaign to the next: MolRSI screens 107 compounds at a 15.9% hit rate
Synopsis
The molecular-discovery agent MolRSI stores the outcome of each workflow search in cross-task memory, so later tasks start from accumulated experience instead of trial and error from scratch. In virtual antibiotic screening it selected 107 compounds from the same training data and screening library as an earlier expert-led workflow and reached an experimental hit rate of 15.9%, versus 8.7% for that workflow; in peptide–MHC binder design it produced eight experimentally validated hits against MART-1 and SARS-CoV-1, where Claude Science produced none under the same conditions.
Interpretation
Molecular-discovery agents are beginning to carry search experience from one project into the next, rather than only executing expert-written design protocols automatically. Earlier scientific agents could retrieve tools and compose them into workflows, but campaigns stayed isolated, so which modifications worked and which executions failed in one project did not shape the search policy for the next. Across six bacterial-species antibacterial activity prediction tasks, MolRSI reached a higher mean test AUPRC than Biomni, Claude Science, STELLA, Codex and Claude Code under the same search budget, ranking first on four tasks and second on the remaining two.
MolRSI turns scientific methods and their code into executable, run-validated skills, then assembles and revises candidate workflows from them. The search is run by experience-guided Monte Carlo tree search, which decides which workflow family to explore, which parent workflow to modify, and how to allocate compute and repair execution errors. The small-molecule resource library holds 110 reusable skills, comprising 55 models, 48 methods and seven tools or resources; the antibacterial experience came from independent searches on 25 source tasks that excluded the six antibacterial benchmarks.
Experience is kept in three memories: workflows with their scores, proposed modifications with their outcomes, and runtime records with error repairs. Every attempt updates these memories, so later iterations within a task and later tasks both draw on earlier results, closing a recursive self-improvement loop. On all six antibacterial tasks, search with cross-task experience improved mean test AUPRC; raising the number of source tasks from 2 to 25 lifted transferred-workflow mean test AUPRC from about 0.32 to about 0.37, above the 0.361 reached by search without experience after 20 iterations.
Perspective
The demonstrated scope is two task types: virtual antibacterial screening, where 107 compounds were tested from the same training data and roughly 12.1-million-compound screening library as the earlier expert-led workflow, and peptide–MHC binder design, screened by phage display against MART-1 and SARS-CoV-1. The intended readers are researchers and engineers working on scientific agents and molecular design pipelines.
The authors note that whether policy updates stay useful as biological targets and evaluation criteria change remains to be tested, and that which tasks enter the experience pool affects results; they also state that biological validation remains necessary to establish whether better computational scores yield more effective candidates, and propose incorporating experimental outcomes into future policy updates.
