MolClaw: An Autonomous Agent with Hierarchical Skills for Drug Molecule Evaluation, Screening, and Optimization

Lisheng Zhang; Lilong Wang; Xiangyu Sun; Wei Tang; Haoyang Su; Yuehui Qian; Qikui Yang; Qingsong Li

cs.AIApr 2, 2026

MolClaw: An Autonomous Agent with Hierarchical Skills for Drug Molecule Evaluation, Screening, and Optimization

Authors: Lisheng Zhang, Lilong Wang, Xiangyu Sun, Wei Tang, Haoyang Su, Yuehui Qian, Qikui Yang, Qingsong Li,

Organizations: Peking University Health Science Center, Peking University, Beijing, China · Shanghai AI Laboratory, Shanghai, China · Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China

Abstract

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.

Figures & tables

Figure 1: Technical architecture and key capabilities of MolClaw. ( A ) illustrates the hierarchical skill ecosystem and the MolClaw’s execution process. ( B ) highlights MolClaw’s superior performance across five molecular screening and optimization tasks. ( C , D ) demonstrate MolClaw’s ability to conduct long-term, iterative drug discovery processes. ( E ) showcases autonomous scientific discovery and experimental execution.

Figure 2: Evaluation results of MolClaw, three representative agents, and multiple large language models (LLMs) on the MolBench-MS and MolBench-MO benchmarks. ( A–C ) Performance comparison on the three subtasks of MolBench-MS. ( D , E ) Performance comparison on the two subtasks of MolBench-MO, where both success rate and delta metrics are reported for the second molecular optimization task. ( F–G ) Ablation study of MolClaw compared with vanilla agents, including Claude Code (CC) and OpenClaw (OC), using the default LLM of Claude Sonnet 4.6 and an alternative open-source LLM of Qwen 3.5.

Figure 3: Agent execution traces for the three MolBench-E2E tasks. ( A ) E2E-Q1: coarse-grained conformational sampling. Five tool-level failures (red) were resolved via skill-governed recovery actions (orange), yielding 20 verified all-atom structures. ( B ) E2E-Q2: QED-driven iterative optimization. One tool fallback (F1), two constraint-driven rejections (F2–F3), and five strategy adaptations (D1–D5) were autonomously managed across five rounds, with all 19 reported values verified against source files. ( C ) E2E-Q3: structure-guided lead optimization of Erlotinib. Three failure categories—baseline docking crashes, persistent ProLIF errors, and generative-model collapse—were resolved via pipeline rewriting, score-only SAR, and manual analog design. Inset: optimization trajectory reaching the $-8.9$ kcal/mol target. Blue dashed boxes: agent decisions; purple badges: governing skill layer; green diamonds: verification checkpoints.

Figure 4: Coarse-grained conformational sampling of the EGFR kinase domain by OpenAWSEM and GoCa. ( A ) Superposition of 10 PULCHRA-reconstructed all-atom conformations from the OpenAWSEM ensemble, aligned to the 1M17 crystal structure. ( B ) Corresponding superposition for the GoCa ensemble. ( C ) C $\alpha$ -RMSD to native structure: GoCa $4.54\pm 0.93$ Å versus AWSEM $7.78\pm 1.53$ Å ( $P=7.69\times 10^{-4}$ ). ( D ) Radius of gyration: GoCa $R_{\mathrm{g}}=23.77\pm 0.65$ Å, AWSEM $R_{\mathrm{g}}=19.66\pm 1.18$ Å, native $R_{\mathrm{g}}=22.2$ Å (dashed line; $P=1.83\times 10^{-4}$ ). ( E ) Pairwise C $\alpha$ -RMSD within each ensemble ( $n=45$ pairs): GoCa $4.28\pm 0.88$ Å versus AWSEM $5.67\pm 1.60$ Å ( $P=2.44\times 10^{-5}$ ). ( F ) Per-residue RMSF profiles across 312 residues (Pearson $r=0.685$ ); shaded regions denote the N-lobe $\beta$ -sheet, $\alpha$ C-helix, and C-lobe. Violin plots show individual data points with box-and-whisker summaries; $P$ -values from two-sided Mann–Whitney $U$ tests; ${}^{***}P<0.001$ ; $n=10$ conformations per method.

Figure 5: QED-driven iterative optimization of a triazolo-benzodiazepine scaffold by the AI agent. ( A ) Multi-dimensional property trajectory across five optimization rounds: QED score (target

\geq

0.70), MW, ALogP, Tanimoto similarity (constraint

\geq

0.40), TPSA and rotatable bonds. ( B ) QED desirability decomposition by component and round (R0–R5). ( C ) QED–Tanimoto trade-off for all 182 molecules; red stars, selected best per round; green quadrant, feasible region. ( D ) ADMET profile evolution (top: risk endpoints; bottom: benefit endpoints). ( E ) Generation efficiency: molecules requested, generated and qualifying per round (left); yield and qualification rates (right). ( F ) QED component contribution waterfall (R0

\to

R4). ( G ) Radar chart comparing R0 and R4 desirability profiles. ( H ) Cumulative QED gain (left) and marginal gain per round (right). QED and Tanimoto computed via RDKit (Morgan fingerprint, radius = 2, 2,048 bits); ADMET by ADMET-AI;

n=182

molecules (54 + 28 + 44 + 23 + 33).

Figure 6: Comprehensive evaluation of AI-agent-driven iterative lead optimization of Erlotinib targeting the EGFR kinase domain. ( A ) Optimization trajectory showing best QuickVina docking score per round; blue dashed line: Erlotinib baseline ( $-6.9$ kcal/mol); red dashed line: $-8.9$ kcal/mol target. ( B ) Docking score distributions across Rounds 1–6 (box-and-strip plot, $n=54$ ). ( C ) Tanimoto similarity heatmap between round-best molecules. ( D ) Drug-likeness radar plot comparing Erlotinib with two target-met derivatives. ( E ) Agent performance audit against seven-criterion rubric (100/100). ( F ) Molecule generation sources per round (left) and convergence curve (right). ( G ) Atom-pair contact counts ( $<4.0$ Å) across 15 binding-site residues. ( H ) Contact frequency by interaction type across round-best poses. ( I ) Interaction evolution for six key residues across rounds. ( J ) Protein–ligand interaction fingerprint heatmap for all seven round-best poses.

Figure 7: Schrödinger-style 2D protein–ligand interaction diagrams and 3D pose overlay. ( A ) Erlotinib baseline ( $-6.9$ kcal/mol): two H-bonds (Thr766, Asp831) and eight hydrophobic contacts. ( B ) R1 best ( $-7.4$ ): methoxy shortening + meta-Br; new Met769 H-bond (2.99 Å). ( C ) R2 best ( $-8.0$ ): Br $\to$ F substitution; Met769 maintained. ( D ) R3 best ( $-8.3$ ): F + OH + CH 3 on aniline. ( E ) R4 best ( $-8.9$ , target met): 2,6-diF-4-OH aniline; four simultaneous H-bonds (Met769, Thr766, Thr830, Asp831). ( F ) R5 best ( $-8.4$ ): three H-bonds. ( G ) R6 best ( $-8.9$ , target met): distinct three-H-bond network (Lys721, Thr766, Met769). ( H ) 3D rendering of representative poses within the EGFR ATP-binding pocket (electrostatic potential surface).

Figure 8: Statistical validation, source attribution, and interaction conservation analysis. ( A ) Docking scores of all 54 molecules by source: REINVENT4 (blue, $n=20$ ) and agent-designed (red, $n=34$ ). ( B ) Per-round mean scores $\pm$ s.e.m. tested against baseline (Wilcoxon); R1 n.s., R2–R3 ∗∗ , R4–R6 ∗∗∗ . ( C ) Early (R1–R3) vs. late (R4–R6) violin plot ( $p=1.24\times 10^{-4}$ ). ( D ) Agent vs. REINVENT violin plot ( $p=0.104$ , n.s.). ( E ) H-bond residue count across rounds (Kruskal–Wallis $p=0.89$ , n.s.). ( F ) H-bond frequency for five key residues per round. ( G ) Met769 H-bond present vs. absent ( $p=0.0036$ ). ( H ) Forest plot of per-residue H-bond impact; only Met769 significant ( $p=0.007$ ). ( I ) Score vs. Tanimoto (Spearman $\rho=+0.650$ , $p<10^{-7}$ ). ( J ) Tanimoto across rounds ( $\rho=-0.691$ , $p=7.4\times 10^{-9}$ ). ( K ) 3D rendering of R6 target-met compound in EGFR binding pocket with three-point H-bond network. $n=54$ molecules; ${}^{**}P<0.01$ ; ${}^{***}P<0.001$ .

Drug Discovery

Drug discovery identifies, designs, and tests molecules that can safely affect disease biology.

See all 164 papers

DrugHarness: An Evolvable Agent Harness for Computational Drug Discovery
Oct 16, 2026Yuxing Chen, Zhisheng Zhang, Cheng Tang +20LLM Agent Harnesses Drug Discovery
Shanghai Artificial Intelligence Laboratory · China Pharmaceutical University · University of the Chinese Academy of Sciences +7
MADDIE: Multi-Agent Drug Discovery Engine
Oct 12, 2026Ye Liu, Reza Averly, Avery Meyer +4Drug Discovery Multi-Agent LLM Systems
Biomedical Informatics, The Ohio State University. · Computer Science and Engineering, The Ohio State University. · Medicinal Chemistry and Pharmacognosy, The Ohio State University. +1
From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery?
Oct 5, 2026Pierre Llompart, Levent Guner, Helen Lai +2AI Agents for Scientific Discovery LLM Agent Orchestration
AMD Silo AI, Stockholm, Sweden · Molecular AI, Discovery Sciences, R&D, AstraZeneca, Cambridge, UK · Molecular AI, Discovery Sciences, R&D, AstraZeneca, Gothenburg, Sweden +1

Oct 16, 2026J/K move · Enter open · S save

DrugHarness: An Evolvable Agent Harness for Computational Drug Discovery

Yuxing Chen, Zhisheng Zhang, Cheng Tang, Wei Li+19

Shanghai Artificial Intelligence Laboratory · China Pharmaceutical University · University of the Chinese Academy of Sciences+7

LLM Agent Harnesses Drug Discovery Self-Evolving Agents LLM Agent Skill Learning

Open paper AskPDF arXiv

Computational drug discovery requires agents to compose scientific operations and adapt to changing evidence and constraints. Here we introduce DrugHarness, an evolvable agent harness coupling composable scientific execution with conditional skill induction. Its environment integrates over 80 scientific operations, biomedical resources and task-specific code, preserving associations among candidates, results and artifacts. From execution trajectories and outcome feedback, DrugHarness induces action hypotheses with observable applicability conditions, recommended actions and supporting evidence. An independent review agent evaluates candidate skills derived from these hypotheses through source verification and, when needed, targeted execution trials. Admitted skills guide subsequent tasks and are revised through feedback. Evaluation on MolBench and SMDD-Bench demonstrates capabilities in molecular screening, optimization and constrained design. Without evolution, DrugHarness exceeds reported MolClaw-CC results by 10.8 percentage points in affinity-comparison accuracy and 55% in mean top-three active hits on MolBench. Retrospective evaluation on skill-development tasks shows that evolution increases Qwen's affinity-comparison accuracy from 62.2% to 81.1% and optimization success from 79.5% to 87.2%. In a prospective HPK1 study, DrugHarness achieves a more favorable median docking score than Biomni among system-specific candidates. A synthesized compound achieves a cellular IC50 of 14.19 nM, a 2.64-fold potency improvement over the best input reference, with activity competitive with two selected compounds designed by experienced medicinal chemists. These results support conditional skill evolution for scientific agents and demonstrate a workflow connecting computational drug design with experimental validation.

Abstract

Drug Discovery

Explore similar work

DrugHarness: An Evolvable Agent Harness for Computational Drug Discovery

MADDIE: Multi-Agent Drug Discovery Engine

From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery?