← All model news

MolRSI makes research workflows themselves the search target: 15.9% wet-lab hit rate in antibacterial screening vs. 8.7% for an expert non-agent pipeline

Machine Heart reports that MolRSI uses experience-guided tree search to automatically combine molecular screening and protein design workflows. It self-reports a 15.9% hit rate in antibacterial wet-lab experiments versus 8.7% for an expert non-agent pipeline, and 8 versus 0 specific hits for pMHC-binding proteins.

Published Updated

What changed

According to a November 14, 2026 report by Machine Heart (the article labels the paper "MoLRSI: A Recursive Self-Improving AI agent for Molecular Discovery" and gives an arXiv number and a project page), the research team made workflow design itself the object of search in molecular discovery. MolRSI first organizes methods from papers and code and converts them into executable skills, then uses experience-guided Monte Carlo tree search to combine, test and modify workflows: each candidate pipeline is a node, and changes to it form branches; plan memory, improvement memory and execution memory store pipeline performance, modification effects and runtime experience respectively; the inner loop optimizes the current task, and the outer loop carries experience into later tasks. The report says experience was first accumulated on 25 molecular property prediction and cytotoxicity tasks, then applied to 6 antibacterial prediction tasks not in the experience pool, with average test AUPRC improving on all 6. Expanding the source task pool from 2 to 25 raised the average test AUPRC of transferable pipelines from about 0.32 to about 0.37 in retrospective analysis, exceeding the 0.361 reached by a system without prior experience after 20 rounds of systematic search. On the antibacterial task, the final pipeline first filtered for structural novelty, then predicted three types of human cytotoxicity, ranked candidates with two MiniMol schemes and a LANTERN–RADR ensemble weighted 55%, 30% and 15%, and finally sampled by structural cluster to preserve diversity; from about 12.1 million compounds it selected 107 for experiments, with 17 hits counted at MSSA MIC no higher than 32 μg/mL; an expert-designed non-agent pipeline, using the same training data and screening library, tested 241 and found 21 hits, an 8.7% versus 15.9% hit rate. On the protein design side, MolRSI first fixed an evaluation pipeline (ESMFold2-Fast structure prediction plus eight equally weighted scores, with selectivity enforced through paired comparisons of target versus non-target), then used it to search generation schemes; after two rounds of phage display it obtained a total of 8 hits meeting the specificity criterion across the MART-1 and SARS-CoV-1 targets, while the report says a Claude Science control under the same conditions got 0; after removing the ranking step, none of 479 evaluable designs was a hit.

What it means for developers

For developers building agent workflows, the interesting part of this report is not a particular score but that it treats "workflows" as searchable objects and explicitly stores cross-task experience for reuse. That directly affects three engineering decisions. First, the boundary of memory: almost all the gains in the report come from inheriting cross-task experience, while the no-experience group still keeps within-task memory updates, which suggests the benefit is concentrated in "transferring from one task to another." If your agent gets only one or two calls per task type, the experience library has no time to accumulate, and the extra cost of that tree search is hard to recover; plan memory and repair memory only make sense when structurally similar tasks recur in the same domain. Second, the evaluator comes first: in the pMHC part, the scoring pipeline was fixed before optimizing generation, and removing the ranking step drove hits to zero, the strongest ablation signal in the report. This suggests the ceiling on generation is often set by "deciding which candidates are worth sending to the lab," so the ranker and threshold design need separate evaluation rather than being summed up by a single final hit rate. Third, metrics are layered: AUPRC, enrichment factor, wet-lab hit rate and specificity criteria are four different quantities, and the report itself distinguishes "comparison of predictors on the same test set" from "hit rates of experimental candidates selected by two different pipelines"; putting numbers from different sources in one column leads to wrong conclusions. In addition, the report explicitly acknowledges that current strategy improvements come mainly from computational search and that wet-lab feedback has not yet entered the loop, which means it is not yet a truly closed-loop experimental agent, and readers should not extrapolate to "AI doing experiments autonomously."

Integration and evaluation

These are conditional suggestions based on the report, not verified conclusions. First, if your pipeline contains a weighted ensemble of multiple models (such as the 55/30/15 three-way combination in the report), first run a "fixed weights versus searched weights" comparison on the same held-out set, with AUPRC and top-k enrichment as metrics, provided the held-out set has no compounds overlapping the training set; only when the searched combination stably beats fixed weights across multiple held-out folds is it worth putting weights into the search space, otherwise the search just fits noise. Second, for cross-task experience reuse, set a "zero-search" baseline first: transfer the best configuration from the source task directly without extra search, then compare with the full tree search, provided the target and source tasks agree on data modality and label definitions; the retrospective curve in the report, with AUPRC rising from about 0.32 to 0.37, shows that experience pool size changes the conclusion, so you should put experience pool size (2, 10, 25 source tasks) on the x-axis and see whether gains appear at the scale you have; if you see no improvement within 5 source tasks, there is no need to invest engineering cost in an experience library. Third, for selectivity tasks (where target and non-target differ by only a few residues), make "target scores higher than non-target" a hard gate rather than a soft penalty, and use the pass rate of paired comparisons as an independent metric, provided a reliable non-target set exists; in the report, removing ranking drove hits to zero, suggesting this step deserves its own ablation rather than being tuned together with the generative model. Fourth, before integrating, confirm which components you can obtain: the report mentions MiniMol, LANTERN–RADR, ESMFold2-Fast, SolubleMPNN and others; their licenses and runnability need to be verified separately, and the report does not state how open these components are.

Limits and open questions

All the above numbers come from the Machine Heart report and have not been independently reproduced. MolRSI is a research system rather than a publicly released model; the available information does not say whether weights, code or an API are open, and the report only gives a paper and a project page; the current KimiSeek live catalog also contains neither MolRSI nor the components mentioned above, so availability can only be judged by statements from the paper authors and the project. The antibacterial hit rates come from experimental candidates selected separately by two pipelines, with sample sizes of 107 and 241; a percentage difference from a single experiment should not be treated as a stable effect. The cross-task experience gains come partly from retrospective analysis, which the authors themselves frame as retrospective evidence. The pMHC specificity criterion is target enrichment log2FC greater than 4 and higher than the highest measured non-target value; the report explicitly says quantitative affinity and broader cross-reactivity still need follow-up measurement; the 0 hits for Claude Science is also the report's account, and the details of how that control was run are not expanded in this article. In addition, wet-lab feedback has not yet entered the improvement loop; the system's self-improvement is currently limited to computational search, and it cannot be inferred that it can autonomously complete a closed experimental design loop.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

AI自己改科研策略,能过湿实验这一关吗?他们把RSI带入分子发现

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.