MolRSI

Recursive self-improvement · Molecular discovery

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

Research experience should change what an AI tries next.
MolRSI turns that experience into a better search policy.

Read the paper Watch 60-second demoBibTeX
Yikun Zhang1,2,*Xiwei Cheng1,2,*Mingze Yuan3Yu Zhang2,4,5Brian Shim2,3Aarti Krishnan2,4,5Daniel Rosen2,14Kunyang Sun6Yingze Wang6Zhenfeng Deng7Botao Yu8Bingxu Liu9David Liu2,3,10Quanzheng Li13Tianyu Liu11,†Yuanqi Du12,†Nir Hacohen2,3,13,†James J. Collins2,4,5,†Wengong Jin1,2,†

* Equal contribution   ·   † Co-corresponding authors   ·   Affiliations

Antibacterial discovery
15.9%vs8.7%

Experimental hit rate vs an expert-designed,
non-agentic workflow.

 De novo pMHC binder design
8vs0

Selective binder hits across two targets,
compared with Claude Science.

See the loop in action.

60 seconds · Silent demo

01 / The method

Two loops.
A better search policy.

Human scientists draw on past successes and failures to decide what to try next. MolRSI connects the optimization of today’s workflow with the way it searches on the next task.

01

Inner loop: improve the workflow

Select, evaluate and refine workflows for the current task. Each result provides feedback about what works and what needs to change.

02

Outer loop: improve how to search

Carry experience across tasks. Persistent memory guides future workflow selection, refinement, resource allocation and repair.

Both loops update shared memory

Solutions. Refinements. Execution feedback.
Experience becomes guidance for the next search.

Shared memory connects optimization within a task to learning across tasks. Mechanism illustration.
Persistent experience changes the effective search policy. The backbone language model is not retrained.

02 / Learning from experience

Does prior experience help?

Evaluate how memory changes search, then test how the available experience pool changes transfer.

Antibacterial search

A stronger first workflow.

On E. coli and N. gonorrhoeae, MolRSI’s first workflow with prior memory exceeds all five agents’ final workflows after 20 evaluations in mean held-out test AUPRC.

pMHC ranking

Memory improves enrichment.

3.65 / 5.55vs 0 / 0 without prior memory

EF@1% / EF@2% on held-out MART-1 ranking, after experience from 14 miniprotein tasks.

Explore the antibacterial benchmarks

MolRSI and five agent baselines

Experience scaling

More experience makes better workflows available for transfer.

2 → 25Source tasks
~0.32 → ~0.37Mean test AUPRC

Across six antibacterial tasks, a larger source-task pool makes stronger workflows available for transfer. This is a retrospective transfer analysis.

Retrospective workflow transfer across six antibacterial benchmarks.

03 / Experimental discovery

Search policies.
Real molecular outcomes.

Task 1

Antibacterial discovery

MolRSI develops a screening workflow that balances antibacterial activity, cytotoxicity and structural novelty.

  1. Structural filters
  2. Cytotoxicity filtering
  3. Activity ranking
  4. Diversity-aware selection
  5. Experimental assays

Same data · Same 12.1M-compound library

Compared with an expert-designed, non-agentic workflow.

Hit criterion: MSSA MIC ≤32 μg/mL. Different selected screening cohorts; the comparison uses the same training data and screening library.

From workflow to validation.

Shared-test AUPRC
0.300 vs 0.090

Same held-out set of 271 compounds.

Selected hit C3
1 μg/mL MRSA MIC
Selected hit C32
8 μg/mL MRSA MIC

MRSA follow-up is shown for these two selected hits; the cohort hit-rate criterion is MSSA.

Dose–response follow-up for two selected hits alongside bacterial growth and human-cell viability assays.
Task 2

De novo binder design

Design pMHC binders for target binding and selectivity. Learn a ranking workflow, hold it fixed, then optimize generation.

  1. Learn to rank
  2. Fix the ranking workflow
  3. Optimize generation
  4. Select designs
  5. Phage display + sequencing

Two rounds of selection and sequencing

8 selective binder hits.
0 for Claude Science.

Hits show preferential enrichment for their intended pMHC over the tested off-targets.

Selective hits / evaluable designs
TargetMolRSIClaude Science
MART-15 / 2510 / 288
SARS-CoV-13 / 2660 / 267

Same hit criteria: on-target log₂FC >4 and above every measurable off-target enrichment. Selectivity is relative to the tested off-target panel.

A closer look

From the complex to the interface.

Compare predicted complexes for the same binder with on-target and peptide-swapped off-target pMHCs, then zoom in on the interfaces.

Predicted structural models. The experimental selectivity evidence comes from phage display and sequencing.

Predicted structures · on-target and off-target comparison

04 / Read & cite

Explore the full study.

The methods, benchmarks, experimental validation and supplementary analyses are available in the paper.

Paper on arXiv Read PDF
BibTeX
@misc{zhang2026molrsi,
  title = {{MolRSI}: A Recursive Self-Improving {AI} agent for Molecular Discovery},
  author = {Yikun Zhang and Xiwei Cheng and Mingze Yuan and Yu Zhang and Brian Shim and Aarti Krishnan and Daniel Rosen and Kunyang Sun and Yingze Wang and Zhenfeng Deng and Botao Yu and Bingxu Liu and David Liu and Quanzheng Li and Tianyu Liu and Yuanqi Du and Nir Hacohen and James J. Collins and Wengong Jin},
  year = {2026},
  eprint = {2610.39714},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  doi = {10.48550/arXiv.2610.39714},
  url = {https://arxiv.org/abs/2610.39714}
}
Authors & affiliations
  1. Northeastern University
  2. Broad Institute of MIT and Harvard
  3. Harvard University
  4. Wyss Institute for Biologically Inspired Engineering, Harvard University
  5. Massachusetts Institute of Technology
  6. University of California, Berkeley
  7. University of Toronto
  8. The Ohio State University
  9. University of Washington
  10. Howard Hughes Medical Institute
  11. Yale University
  12. Microsoft Research New England
  13. Massachusetts General Hospital
  14. Brigham and Women’s Hospital and Dana-Farber Cancer Institute

Equal contribution: Yikun Zhang and Xiwei Cheng.
Co-corresponding authors: Tianyu Liu, Yuanqi Du, Nir Hacohen, James J. Collins and Wengong Jin.