Claimed Search scaffold

A self-improving agent selected antibiotic candidates at nearly twice a prior expert workflow's hit rate

MoLRSI, an agent that rewrites its own molecular-discovery workflows using memory of earlier tasks, built a screening workflow whose 107 selected compounds hit at 15.9% against Staphylococcus aureus, against 8.7% for the same group's earlier expert-designed workflow on the same training data and the same 12-million-compound library, and separately produced eight experimentally validated peptide-MHC binders.

Model
Claude Sonnet 5 as the MoLRSI backbone; Claude Opus 5 for the Claude Science comparator
Field
Biology
Date
2026-10-29
Human collaborators
Yikun Zhang, Xiwei Cheng, Wengong Jin, Nir Hacohen, James J. Collins

What was found

The paper's own framing is an agent architecture that optimizes workflows and model architectures and carries memory across tasks, which on its own would be out of scope here; what brings it in is the wet-lab validation. For antibiotics the agent assembled its own activity and cytotoxicity predictors, screened a library of 12,076,365 compounds drawn from the Mcule purchasable database and the Broad Institute collection, and selected 107 compounds for testing, two of which, C3 from ChemBridge and C32 from Vitas-M, were characterized for minimum inhibitory concentration against methicillin-sensitive and methicillin-resistant S. aureus and for cytotoxicity against HepG2, HSkMC and IMR-90 cells. The comparator is the same laboratory's earlier expert-led workflow, which tested 241 compounds from the same library using the same training data. In the second setting the agent designed de novo binders to HLA-A*02:01 presenting the MART-1 and SARS-CoV-1 peptides, with designs synthesized and displayed on phage, yielding eight hits that preferred their intended target over the off-targets tested, while a comparator protocol run with Claude Science yielded none.

Novelty check

arXiv, 2026-11-09: abs:"agent" AND abs:"antibiotic" AND abs:"workflow" returned exactly 1 record, this paper. The work extends two entries this registry already carries from the Collins group, 2020-02-20-halicin and 2023-12-20-antibiotic-structural-class, rather than duplicating them: the expert workflow it beats is that prior line of work, and the authors state the compound structures, experimental measurements and data splits were reused from it. Whether C3 and C32 are structurally novel scaffolds rather than known antibacterials is asserted through the workflow's own novelty filter and a t-SNE projection against the training set, and was not checked here against a chemical database.

Caveats and known objections

An arXiv preprint, not peer-reviewed, so capped at claimed; the listed version note records that the author list itself was revised. Evidence is in vitro only: minimum inhibitory concentrations, cell-line cytotoxicity and phage-display binding, with no animal or clinical testing, which is why this is filed under biology rather than medicine. The headline comparison is against the authors' own prior workflow rather than an independent one, the two arms tested different numbers of compounds (107 against 241), and the hit-rate difference carries Wilson confidence intervals the abstract does not give. The agent and every baseline were configured, prompted and evaluated by the authors, who also supplied the task inputs and evaluation interfaces, so the finding that a competing agent produced no hits is a result about one protocol as the authors ran it, not a general ranking. The paper names no code or data repository for MoLRSI itself. Autonomy is search-scaffold: humans defined the discovery objectives, the libraries, the assays and the search budget, and the agent searched and rewrote workflows inside that frame.

No independent check is recorded on this entry yet.

Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.

Or on GitHub: submit a check challenge the grade send a correction or send a pull request

Entry history (1 event)
  1. AddedEntered the registry graded Claimed and Search scaffold.

Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.

Graded claimed for verification and search scaffold for autonomy. What these mean.

Cite this entry

Plain text
whataifound.org. (2026). A self-improving agent selected antibiotic candidates at nearly twice a prior expert workflow's hit rate. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-10-29-molrsi-molecular-discovery
BibTeX
@misc{whataifound-northeasternuniversity-2026-discovery,
  title        = {A self-improving agent selected antibiotic candidates at nearly twice a prior expert workflow's hit rate},
  author       = {{whataifound.org}},
  year         = {2026},
  howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
  note         = {Result by Northeastern University / Broad Institute of MIT and Harvard. Verification: Claimed. Autonomy: Search scaffold.},
  url          = {https://whataifound.org/finding/2026-10-29-molrsi-molecular-discovery}
}

Related findings

← All biology findings in the registry