MoLRSI: A Recursive Self-Improving AI agent for Molecular Discovery
Authors: Yikun Zhang, Xiwei Cheng, Mingze Yuan, Yu Zhang, Brian Shim, Aarti Krishnan, Daniel Rosen, Kunyang Sun, Yingze Wang, Zhenfeng Deng, Botao Yu, Bingxu Liu, David Liu, Tianyu Liu, Yuanqi Du, Nir Hacohen, James J. Collins, Wengong Jin
Organizations: Northeastern University, Boston, MA, USA · Broad Institute of MIT and Harvard, Cambridge, MA, USA · Harvard University, Cambridge, MA, USA · Wyss Institute for Biologically Inspired Engineering, Harvard University, Boston, MA, USA · Massachusetts Institute of Technology, Cambridge, MA, USA · Brigham and Women’s Hospital and Dana-Farber Cancer Institute, Boston, MA, USA · University of California, Berkeley, Berkeley, CA, USA · University of Toronto, Toronto, ON, Canada · The Ohio State University, Columbus, OH, USA · University of Washington, Seattle, WA, USA · Howard Hughes Medical Institute, Cambridge, MA, USA · Yale University, New Haven, CT, USA · Microsoft Research New England, Cambridge, MA, USA · Massachusetts General Hospital, Boston, MA, USA
Molecular discovery requires navigating vast combinatorial search spaces while balancing competing biological objectives. AI agents can automate expert-defined molecular design protocols, but automation alone does not address how to optimize workflows for new tasks or improve their development through experience. Here we introduce MoLRSI, an AI agent that combines systematic optimization of workflows and model architectures with persistent cross-task memory to recursively improve its search policy for molecular discovery. Workflow performance, refinement outcomes, and execution feedback guide subsequent searches, allowing the agent to improve both its workflows and its strategy for developing them. We evaluate MoLRSI in two challenging molecular discovery settings that require balancing competing properties. 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. In de novo peptide-MHC binder design, MoLRSI developed workflows that jointly consider target binding and specificity, producing eight experimentally validated hits against MART-1 and SARS-CoV-1 pMHCs with preferential binding over tested off-targets while Claude Science completely failed. These findings demonstrate how experience-guided workflow optimization can address competing discovery objectives while improving the search strategies used in subsequent campaigns.
Figures & tables
Figure 1: MolRSI motivation and architecture. (a) Molecular discovery involves a combinatorial space of decisions across preparation, generation and prioritization, illustrated for de novo protein binder design. Nodes represent options at each stage. The orange path traces one possible workflow. (b) Scientists accumulate lessons from experimental outcomes across projects, learning to avoid unproductive choices and focus exploration on promising regions. (c) MolRSI couples systematic workflow search with persistent cross-task memory. The Explore agent retrieves relevant resources from scientific literature and code repositories, then converts them into reusable, executable skills through implementation and validation. Workflow evaluations update solution memory (workflows and performance), refinement memory (modifications and their effects) and execution memory (execution strategies and repairs). Accumulated experience guides subsequent workflow selection, refinement and execution, recursively improving the effective search policy from πt to πt+1 . Colored paths indicate evaluated workflows, faded elements denote deprioritized choices, and orange-outlined nodes represent new local options. Projects, outcomes and memory entries are illustrative.
Figure 2: Computational evaluation of MolRSI for antibiotic discovery. (a) Test AUPRC and rankings of MolRSI without RSI and baseline agents across six antibacterial tasks. (b) Resource library. (c) Experience from 25 source tasks, excluding the antibacterial targets. (d) Search trajectories over 20 iterations. (e) Search with and without RSI and zero-search transfer, which retains model fitting and validation. (f) Zero-search transfer performance versus memory-pool size. The line shows test AUPRC averaged analytically across the six held-out antibacterial tasks and all possible subsets of each size from the 25-task memory pool. Darker and lighter shading indicate the interquartile range (IQR; 25th–75th percentiles) and the 5th–95th percentile range, respectively, across 2,000 sampled memory subsets at each pool size. In (a,d,e), means and ±1 sample SD are shown across five data splits. Circles in (a) denote individual splits.
Figure 3: Virtual screening and experimental validation with MolRSI. (a) Screening workflow yielding 107 tested compounds; cytotoxicity-filter exceptions are detailed in Section S3.4 . (b) t-SNE projection [ ? ] of training and tested compounds; stars indicate hits. (c) Activity-score ranking of tested compounds. (d) MSSA and MRSA MICs and HepG2, HSkMC and IMR-90 CC50 values for two selected hits, C3 and C32; dots show individual MIC measurements. (e) Hit rates for 107 MolRSI -selected and 241 expert-selected compounds [ ? ], with 95% Wilson confidence intervals. (f) AUROC and AUPRC on the shared 271-compound held-out S. aureus test set. (g) Chemical structures and dose–response curves of C3 and C32. Symbols show responses; lines indicate descriptive logistic fits. Two response series are shown per curve. Hits are defined by an MSSA MIC of 32 μ g/mL or lower.
Figure 4: Benchmarking and experience-guided optimization of pMHC binder-design workflows. (a) Enrichment Factor (EF) at the top-ranked 1% and 2% cutoffs (EF@1% and EF@2%) for MolRSI and four agent baselines across 14 miniprotein-binder tasks. Heatmaps show individual tasks; bars show the unweighted macro average across tasks. (b) Protein resource library organized into input preparation, generation, sequence design and assessment, structure prediction, scoring and score combination. (c) EF@1% and EF@2% on the held-out MART-1 pMHC target for MolRSI with and without RSI and four agent baselines. (d) Best-so-far search objective, defined as the mean ranking score among the top 5% of candidates, during MART-1 and SARS-CoV-1 design-workflow optimization. (e) On-target versus off-target scores for selected MART-1 and SARS-CoV-1 designs from MolRSI and Claude Science, evaluated with the MolRSI ranking score (left) and Claude Science score (right).
Figure 5: Prospective experimental validation of pMHC binder designs. Upper and lower rows in a–c show MART-1 and SARS-CoV-1, respectively. (a) On-target log2FC and its difference from the maximum available off-target log2FC. Dashed lines mark hit thresholds ( x>4 , y>0 ); legend values indicate hits/evaluable designs. (b) UMAP [ ? ] of predicted minibinder structures based on TM-align [ ? ] similarity; stars mark experimental hits. (c) Designs ranked by MolRSI score, with hits highlighted in green. (d) Predicted complexes of sars1_0101 and mart1_0014. Minibinders, MHC and peptides are blue, orange and pink, respectively; yellow dashed lines indicate candidate polar contacts.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Component
Information provided
Shared index
Skill names, categories, purposes and links to the corresponding instructions.
Skill instructions
Resource description, intended use, installation, supported operations and usage examples.
Implementation
Available source code, model definitions, dependencies and supporting scripts.
Source information
Scientific references, repository location, revision and package information.
Interfaces and examples
Available operations and examples of feature extraction, model fitting or inference, as appropriate to the resource.
Validation information
Validation status, checks performed and their outcomes.
Appendix
Table S1: Contents of the resource library. The shared index identifies available skills. Each skill combines instructions with the implementation, source information and validation results available for that resource. The supporting material depends on the resource.
Category
Dataset
Endpoint
Size
Metric
Absorption
Caco-2
Intestinal permeability in a Caco-2 assay
906
MAE
Absorption
HIA
Human intestinal absorption
578
AUROC
Absorption
Pgp
P-glycoprotein inhibition
1,212
AUROC
Absorption
Bioavailability
Oral bioavailability
640
AUROC
Absorption
Lipophilicity
Octanol–water distribution coefficient
4,200
MAE
Absorption
Solubility
Aqueous solubility
9,982
MAE
Appendix
Table S3: The 25 source tasks used for memory accumulation. The first 22 tasks are from the TDC ADMET benchmark, and the remaining three are mammalian cytotoxicity tasks. TDC sizes are the reported benchmark sizes; these tasks used the official benchmark partitions.
Training
Validation
Test
Dataset
Seed
Total
Positive
Total
Positive
Total
Positive
S. aureus
1
31,512
340
3,939
48
3,939
82
2
31,512
347
3,939
48
3,939
75
3
31,512
343
3,939
55
3,939
72
4
31,512
335
3,939
59
3,939
76
5
31,512
337
3,939
54
3,939
79
Appendix
Table S5: Training, validation and test sizes for the nine datasets. Total gives the number of samples and Positive gives the number with a positive label. Negative counts equal Total minus Positive. Seeds are grouped only when all partition sizes and positive counts are identical; grouped seeds retain different sample assignments.
Target
Data source
Source designs
Retained positives
Sampled negatives
EGFR
Cao et al. [ ? ]
99,803
15
150
FGFR2
Cao et al. [ ? ]
59,999
601
6,010
H3
Cao et al. [ ? ]
59,941
49
490
IL-7R α
Cao et al. [ ? ]
14,999
22
220
InsulinR
Cao et al. [ ? ]
59,973
256
2,560
PDGFR
Cao et al. [ ? ]
99,992
284
2,840
Appendix
Table S9: Composition of the 14-task miniprotein-binder benchmark. Each row identifies the held-out test target for one task; workflow search and selection used the other 13 targets. Labels use the source-provided binary binding endpoint at 4,000 nM. All positives were retained, and up to ten negatives per positive were sampled within each target.
Antigen
Peptide
Role
Candidates
Positives
gp100
YLEPGPVTA
Search
2,611
14
gp100
IMDQVPFSV
Search
3,271
4
MART-1
ELAGIGILTV
Test
3,845
27
Appendix
Table S10: Dataset roles within the single pMHC binder-design task. Both complete gp100 collections were used for workflow search; MART-1 was reserved for test evaluation.
Method
EF@1 %
EF@2 %
EF@5 %
EF@10 %
AUPRC
FGFR2
MolRSI
5.58
4.47
3.16
2.61
0.2041
Claude Science
7.06
5.87
4.22
3.22
0.2871
Biomni
6.07
5.21
3.79
2.69
0.2212
STELLA
4.93
4.71
3.82
2.82
0.2193
Codex
4.27
3.14
2.69
2.09
0.1554
Appendix
Table S11: Target-level results for the miniprotein-binder benchmark. Each target was held out during the corresponding agent workflow search; no cross-task memory was used. Fixed objectives were evaluated on the same candidates. The target-specific rows report EF at four cutoffs and AUPRC.
Method
EF@1 %
EF@2 %
EF@5 %
EF@10 %
AUPRC
MolRSI
3.65
5.55
2.21
2.22
0.0172
Claude Science
0.00
1.85
0.74
1.48
0.0089
Biomni
0.00
1.85
0.74
1.11
0.0087
STELLA
0.00
1.85
1.48
1.48
0.0095
Codex
0.00
0.00
1.48
0.74
0.0088
BoltzGen
0.00
1.85
0.74
0.74
0.0067
Appendix
Table S12: Held-out pMHC test performance. Agent workflows were selected using gp100 data; fixed expert-defined objectives used the same test candidates. EF values are shown at four cutoffs, alongside AUPRC. Values retain the precision of the source export.
Condition
EF@1%
EF@2%
EF@5%
EF@10%
AUPRC
With RSI
3.65
5.55
2.21
2.22
0.0172
Without RSI
0.00
0.00
1.48
1.48
0.0125
Appendix
Table S13: Effect of source task memory on MART-1 test performance. With and without RSI denote searches with and without prior miniprotein-binder experience.
Input
MART-1
SARS-CoV-1
HLA allele
HLA-A*02:01
HLA-A*02:01
On-target peptide
ELAGIGILTV
GLMWLSYFV
Off-target peptide
GLMGAGIAQV
GLIFSSYFV
Structure provenance
PDB 5NHT
PDB 3I6G
Source PDB chains (HLA heavy chain / β2 m / peptide)
H/L/P
A/B/C
Binder-length range
80–120 residues
80–120 residues
Appendix
Table S14: Inputs and target-specific constraints for the two prospective pMHC binder-design tasks. Source chains give the HLA heavy-chain, β2 -microglobulin, and peptide chain identifiers in the corresponding PDB entry. Peptide positions are one-based. Numbers in parentheses are zero-based indices in the concatenated target sequence.
Computational drug discovery, particularly the complex workflows of drug molecule screening and optimization, requires orchestrating dozens of specialized tools in multi-step workflows, yet current AI agents struggle to maintain robust performance and consistently underperform in these high-complexity scenarios. Here we present MolClaw, an autonomous agent that leads drug molecule evaluation, screening, and optimization. It unifies over 30 specialized domain resources through a three-tier hierarchical skill architecture (70 skills in total) that facilitates agent long-term interaction at runtime: tool-level skills standardize atomic operations, workflow-level skills compose them into validated pipelines with quality check and reflection, and a discipline-level skill supplies scientific principles governing planning and verification across all scenarios in the field. Additionally, we introduce MolBench, a benchmark comprising molecular screening, optimization, and end-to-end discovery challenges spanning 8 to 50+ sequential tool calls. MolClaw achieves state-of-the-art performance across all metrics, and ablation studies confirm that gains concentrate on tasks that demand structured workflows while vanishing on those solvable with ad hoc scripting, establishing workflow orchestration competence as the primary capability bottleneck for AI-driven drug discovery.
Lisheng Zhang, Lilong Wang, Xiangyu Sun +14
Peking University Health Science Center, Peking University, Beijing, China · Shanghai AI Laboratory, Shanghai, China · Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China
Despite breakthroughs in AI-driven molecular generation, existing discovery pipelines operate largely in silico, with wet-lab feedback reserved for validation rather than iterative improvement. Whether large language model (LLM)-based reasoning agents can accelerate de novo discovery by iterating with lab data remains unproven. Here we use protein-binding aptamer design as a sandbox (1-2 week cycles, $1-$3 per candidate) to demonstrate a wet lab in a closed loop with AI systems. We tasked agentic pipelines with developing aptamer binders against α-thrombin and myoglobin across three design-test-improve cycles (∼1,000 candidates per round; ∼4,400 constructs). Statistically significant, bead-corrected on-target hit rates among new candidates rose across rounds from 0.0% in Round 1 to 26.0% in Round 2 and 42.1% in Round 3 on α-thrombin (37.1% excluding literature motif descendants), and from 10.2% in Round 1 to 3.7% in Round 2 and 33.5% in Round 3 on myoglobin. Single-clone bead-binding titrations of eight full-length α-thrombin candidates confirmed concordance with pooled sequencing enrichment (Pearson r=0.90). In these assays, literature-derived motifs engineered by the agents to reduce off-target reactivity in the pooled screen reached apparent affinities of KD=24.6--133 nM, whereas de novo scaffolds showed low-micromolar binding (KD>1μM). We also report how our pipelines used myoglobin binding data to exploit the G-quadruplex binder space, diagnose cross-reactivity as on-target specificity dropped from 75.0% to 19.1%, and recover by developing strategies that restored specificity to 72.4%. We demonstrate that empirical lab data is more valuable than existing computational proxies, showing how iterative experimental feedback enables AI agents to diagnose assay artifacts and refine molecular designs across rounds.
Zachary Cutts, John Coller, Rachel Chen +9
Google Research · Department of Genetics, Stanford University, Stanford, CA, USA · Google Cloud +1
Early-stage molecular design is an iterative process, not just a task of generating molecules. Researchers turn broad goals into design strategies, refine candidates, assess many properties, and gather evidence before synthesis and tests. AI methods can generate molecules, optimize several goals, predict properties, dock compounds, and account for synthesis. Yet these functions are spread across specialized tools. Experts must still coordinate each step, judge interim results, and integrate evidence. The central challenge is thus to turn research intent into adaptive, traceable runs grounded in scientific tools. We cast this challenge as intent-to-evidence molecular design workflow execution and present CAi Copilot, an expert-oriented agent with three linked layers. The Research Interface Layer turns intent into an executable plan. The Agent Reasoning Layer uses interim results to guide each run. The Execution Substrate supplies molecular tools, metrics, reusable utilities, and backend services. Across 45 tasks, CAi achieves the strongest overall performance, with an outcome score of 84.59, exceeding the next-best result by 18.07 points. Additional benchmarks test how CAi coordinates generation, screening, and multi-criteria evaluation, while exposing limits in long-horizon execution. These results show that CAi turns broad molecular-design intent into transparent, traceable workflows that connect interim decisions to candidate-level evidence.
Zhu Wang, Jiangyu Chen, Yingjun Shang +4
Shanghai Artificial Intelligence Laboratory, Shanghai, China · Nanjing University, Nanjing, China