rsi.watch

Yikun Zhang et al. (18 authors) · 2026-10-29

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

“In antibiotic discovery, MolRSI developed a screening workflow balancing antibacterial activity, cytotoxicity and structural novelty, achieving an experimental hit rate of 15.9%, compared with 8.7% for a previous expert-led workflow using the same training data and screening library.”
Who
Yikun Zhang et al. (18 authors)
Date
2026-10-29
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ risk of bias. The 15.9% and the 8.7% come from two separate campaigns by the same group, with 107 and 241 compounds chosen by different selection procedures, and the authors state that "no between-group hypothesis tests were reported". The authors also ran every agent baseline themselves, and Biomni's "outputs were not independently reproduced outside that platform". “Predictive discrimination was evaluated against shared labels, whereas experimental hit proportions were calculated within each campaign’s selected candidates.”
  • ↓ indirect measure. The headline hit rate measures a complete screening workflow against an earlier expert workflow that differs in models, filter order, thresholds and compound selection, so it does not separate the contribution of cross-task experience. The comparisons that do isolate RSI are computational: mean test AUPRC on the six antibacterial tasks and enrichment factors on the ranking benchmarks. “The workflows differed in their predictive models, filter order and thresholds, and compound-selection procedures (Figure S2).”
  • ↓ imprecise. The wet-lab endpoints rest on single experiments: MICs in two or three technical replicates, and cytotoxicity where "Each compound and concentration was tested in two technical wells in a single experiment". The hit rate is 17 of 107 compounds, reported with 95% Wilson binomial confidence intervals and no hypothesis test. The computational results use five data splits per antibacterial task, and for the pool-size curve the authors write that "The bands describe variation in pool composition, not independent search replicates or confidence intervals for the routing method." “MICs were determined in technical replicate (n=2-3) in a single experiment, and the higher replicate value was reported.”
how we rate evidence
Replication
No rerun found. The paper links no code for MolRSI. The datasets used for workflow development are public except the pMHC binder datasets, which are available from the corresponding authors upon reasonable request. We found no rerun by another group. Searched up to 2026-10-30. all claims
Checked
2026-10-30, against the source page
Links
Source Paper

Also on the page

  • The abstract page titles the paper MoLRSI; the paper itself writes MolRSI throughout. An Explore Agent turns scientific papers, code, tools and benchmarks into validated executable skills, and MolRSI assembles workflows from those skills with Monte Carlo tree search. Three memories carry experience: solution memory holds workflows, configurations and development scores, refinement memory holds modifications and their outcomes, execution memory holds runtimes, failures and repairs. Selection rules read those memories, and the paper records no change to the rules themselves. The backbone is Claude Sonnet 5, accessed through the Anthropic API, and the paper records no change to it.
  • The small-molecule resource library holds 110 reusable skills: 55 models, 48 methods and seven tools or resources. Experience came from 25 source tasks, 22 from the Therapeutics Data Commons plus three cytotoxicity tasks, and the six antibacterial tasks were excluded from that pool. With the same library, backbone and 20-attempt budget, RSI improves mean test AUPRC on all six antibacterial tasks, and zero-search transfer beats search without RSI on three of the six. In the pool-size analysis the transferred workflows go from about 0.32 with two source tasks to about 0.37 with all 25, against 0.361 for search without RSI after 20 iterations, and those transfer scores were picked after the fact on test performance: "we retrospectively selected the transfer test AUPRC available within that subset separately for each target".
  • Experimental validation found 17 of 107 tested compounds with an MSSA MIC of 32 micrograms per mL or lower, a hit rate of 15.9%, against 21 of 241 compounds, 8.7%, in the same group's previous expert-driven campaign at the same threshold. Two hits, C3 and C32, inhibit MRSA at 1 and 8 micrograms per mL. On a held-out set of 271 compounds from the previous study (241 candidate compounds that campaign selected and 30 low predicted activity controls; the 12 known antibiotic class positive controls were excluded) the MolRSI activity predictor reaches AUROC 0.828 and AUPRC 0.300, against 0.590 and 0.090 for the previous predictor. Hit rates are reported with 95% Wilson binomial confidence intervals, and the paper adds that "no between-group hypothesis tests were reported" and that the MIC threshold "was aligned across studies, without assuming identical experimental protocols".
  • On 14 miniprotein-binder tasks, without prior-task memory, MolRSI reaches macro EF@1% of 8.01 against 7.62 for the next-best agent, and EF@2% of 7.29 against 6.71. On the held-out MART-1 pMHC target it reaches EF@1% 3.65 and EF@2% 5.55, while the same system without the miniprotein experience reaches 0.00 at both cutoffs and the highest baseline EF@2% is 1.85. In phage-display screening, MolRSI designs gave five MART-1 hits among 251 evaluable designs and three SARS-CoV-1 hits among 266, Claude Science gave none among 288 and 267, and removing the learned ranking gave none among 479. The authors ran the agent baselines themselves; Biomni's "outputs were not independently reproduced outside that platform", in the antibacterial prediction and protein ranking benchmarks every agent except Codex used Claude Sonnet 5, Codex used GPT-5.5, and Claude Science used Claude Opus 5 for prospective design.
  • In the prospective design comparison both systems got 1,000 A100 GPU-hours per target and the same target, off-target and binder-length specifications, while each kept its own generation, refinement and ranking procedure: "The 30-round allowance applied to MolRSI's workflow search; Claude Science instead controlled job scheduling within the shared GPU-hour budget." MolRSI's own search objective reached 6.76 for MART-1 and 7.56 for SARS-CoV-1, 66% and 28% above the initial configurations.
  • The authors' limits: "Biological validation remains necessary to establish whether better computational scores lead to more effective molecular candidates." On what the search returns: "This is the best evaluated workflow under the stated criterion and budget, rather than a global optimum." On the transfer scores: "The bounded transfer scores serve as empirical selection priors; their boundedness alone does not establish a standard target-task or node-level regret guarantee." On the pool-size bands: "The bands describe variation in pool composition, not independent search replicates or confidence intervals for the routing method." In their own pMHC task instructions: "Positives are scarce: 18 across the two search collections, so a handful of candidates moves the metric and a difference of a few points of AUROC is not a reliable ordering of two workflows."

In the self-improving AI catalog

  • MolRSI: what it changes, whether the change lasts, and how it was tested

← Ledger Report an error