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

MolRSI is an AI agent that combines experience-guided workflow search with persistent cross-task memory to recursively improve molecular discovery workflows, and it is experimentally validated in antibiotic screening and pMHC binder design.

Molecular discovery requires searching vast combinatorial spaces for candidates that satisfy competing objectives such as potency, specificity, toxicity, novelty and stability. Computational workflows combine molecular representations, model architectures, training objectives and selection procedures, and their configuration determines which candidates advance to experimental testing. AI agents can automate expert-defined protocols, but the paper states that reported studies do not establish how experience from one campaign improves workflow development on subsequent problems when campaigns remain isolated.

MolRSI uses an Explore Agent to convert papers, code, tools, benchmarks and algorithms into validated executable skills that form a resource library. It develops candidate workflows with experience-guided Monte Carlo tree search over workflow families and parent workflows, proposing modifications, executing and evaluating them, and updating persistent memory. Solution memory stores workflows and development scores; refinement memory stores modifications and their outcomes; execution memory stores resource use, errors and repairs. Historical transfer scores based on task similarity, current-task performance and an exploration bonus guide family and parent selection. Each search updates memory and changes the effective search policy, forming a recursive self-improvement loop within and across tasks. The agent was evaluated on six antibacterial activity tasks and on miniprotein-binder and pMHC binder-design tasks, including prospective experimental screening.

• In antibiotic discovery, the MolRSI workflow achieved an experimental hit rate of 15.9%, versus 8.7% for the previous expert-led workflow using the same training data and screening library, an approximately 82% relative increase. §2.2
• Experimental testing identified 17 of 107 compounds meeting the MSSA hit threshold of 32 or lower, and C3 and C32 also inhibited MRSA with MICs of 1 and 8. §2.2
• On a shared held-out set of 271 S. aureus compounds, MolRSI achieved AUROC 0.828 versus 0.590 and AUPRC 0.300 versus 0.090 for the previous predictor. §2.2
• MolRSI achieved the highest macro EF on the 14-target miniprotein benchmark at both cutoffs: 8.01 at 1% versus 7.62 for the next-best agent, and 7.29 at 2% versus 6.71. §2.3
• On held-out MART-1, MolRSI achieved EF@1% and EF@2% of 3.65 and 5.55; without prior-task experience the same system achieved 0.00 at both cutoffs, and the highest baseline EF@2% was 1.85. §2.3
• Prospective screening yielded five MART-1 hits among 251 evaluable MolRSI designs and three SARS-CoV-1 hits among 266, while Claude Science yielded no hits among 288 and 267 evaluable designs. §2.3

• Biological validation remains necessary to establish whether better computational scores lead to more effective molecular candidates, and the paper states that future systems could incorporate experimental outcomes into policy updates.
• The antibacterial hit-rate comparison reports 95% Wilson confidence intervals but no between-group hypothesis tests, with denominators of 107 and 241 compounds.
• The prospective pMHC comparison rests on five MART-1 hits and three SARS-CoV-1 hits, so statistical power for the between-method comparison is left open.
• The paper notes that variation among experience pools highlights the importance of the contributing tasks, and that future benchmarks could assess how effectively an agent adapts its search policy over successive tasks.

• MolRSI couples workflow search with persistent cross-task memory to improve its search policy for molecular discovery. §2.1
• The Explore Agent builds a resource library of validated executable skills and repairs or removes skills that fail execution validation. §4.1.1
• Solution memory stores workflows, configurations and scores; refinement memory stores modifications and outcomes; execution memory stores runtime, errors and repairs. §4.1.2
• Experience-guided MCTS selects workflow families and parents using current-task performance, an exploration bonus and historical transfer scores. §4.1.3
• In antibiotic discovery, MolRSI selected 107 compounds and found 17 hits at the MSSA threshold, a 15.9% hit rate versus 8.7% for the expert workflow. §2.2
• On a shared 271-compound held-out S. aureus set, MolRSI achieved AUROC 0.828 versus 0.590 and AUPRC 0.300 versus 0.090. §2.2
• On E. coli and N. gonorrhoeae, MolRSI's first evaluated workflow outperformed the final workflows selected by every baseline agent after 20 search iterations. §2.2
• Prospective screening found eight pMHC hits for MolRSI and none for Claude Science under the tested conditions. §2.3