MolClaw: Hierarchical AI for Drug Discovery
- MolClaw is a hierarchical, autonomous AI agent designed for drug molecule evaluation, screening, and optimization through multi-step workflows.
- It employs a three-tier skill architecture with 70 skills that integrate over 30 specialized domain resources, ensuring validated and recoverable scientific processes.
- MolClaw demonstrates state-of-the-art performance on the MolBench benchmark, achieving significant improvements in binding affinity, docking screening, and iterative molecular optimization.
MolClaw is a hierarchical, autonomous AI agent for computational drug discovery designed to handle multi-step, tool-heavy workflows for drug molecule evaluation, screening, and optimization. It unifies over 30 specialized domain resources through a three-tier hierarchical skill architecture with 70 skills in total, and it is introduced together with MolBench, a benchmark spanning molecular screening, molecular optimization, and end-to-end discovery challenges that require 8 to 50+ sequential tool calls. Its central thesis is that computational drug discovery is not merely a tool-access problem but a workflow-orchestration problem, and that current LLM agents underperform chiefly because they cannot reliably compose specialized tools into validated long-horizon scientific pipelines (Zhang et al., 2 Apr 2026).
1. Problem setting and domain scope
The system is framed around early-stage computational drug discovery workflows that are intrinsically heterogeneous and interdependent. The paper enumerates protein preparation, pocket finding, docking, scoring, property filtering, molecular generation, ADMET evaluation, molecular dynamics, and iterative lead optimization as representative stages. A typical workflow may involve ESMFold or AlphaFold for structure prediction, fpocket or P2Rank for pocket detection, docking tools such as AutoDock Vina or DiffDock, RDKit for property calculations, GROMACS or OpenMM for molecular dynamics, and MM-PBSA or affinity predictors for evaluation. These steps require coordination across incompatible formats and parameters, manual format conversion, and quality control, and are described as error-prone (Zhang et al., 2 Apr 2026).
Within this framing, MolClaw is positioned against an observed limitation of existing LLM agents. The stated limitation is not simply that agents lack access to scientific software, but that they can often write code or call single tools while failing to orchestrate structured, multi-stage scientific workflows reliably. This makes workflow composition, rather than raw tool availability, the principal bottleneck. A plausible implication is that improvements in agent performance depend less on exposing additional APIs than on constraining planning, validation, and recovery across long execution horizons.
2. Hierarchical skill system
MolClaw’s defining mechanism is a three-tier skill architecture consisting of 58 tool-level skills, 11 workflow-level skills, and 1 discipline-level skill. The tiers are explicitly differentiated by granularity and function.
| Tier | Count | Role |
|---|---|---|
| L3 | 1 | Discipline-level governance |
| L2 | 11 | Task-specific validated workflows |
| L1 | 58 | Fine-grained tool wrappers |
The L3 layer is a single methodology document, molclaw-drug-discovery-methodology, which encodes 25 formal scientific principles across 7 chapters. It does not prescribe tool calls directly. Instead, it supplies universal scientific governance: how to plan before acting, how to verify outputs, how to handle data provenance, how to manage residue numbering, how to enforce convergence in iterative optimization, and how to report results honestly. The paper characterizes this, in plain language, as the “scientific constitution” of the agent (Zhang et al., 2 Apr 2026).
The L2 layer comprises 11 task-specific workflow documents. Examples listed in the paper include target protein preparation, molecular docking screening, property filtering, generative molecular design, iterative molecular optimization, binding free energy calculation, conformational sampling, post-docking evaluation, peptide/protein binder design, protein sequence design validation, and multi-target selectivity assessment. These workflow-level skills compose L1 tools into validated end-to-end pipelines with quality gates, decision trees, and failure recovery.
The L1 layer contains 58 fine-grained tool wrappers that standardize atomic operations. Examples include QuickVina2 docking, RDKit descriptor computation, ESMFold or Chai-1 structure prediction, fpocket or P2Rank pocket detection, REINVENT4 generation, ProLIF or PLIP interaction analysis, GROMACS or OpenMM molecular dynamics, MM-PBSA, ADMET-AI, and protein-design tools such as ProteinMPNN, EvoBind, and Chroma. The text also highlights databases and resources including UniProt, PDB, AlphaFold DB, PubChem, and ChEMBL. This arrangement indicates that MolClaw does not treat “tool use” as a flat action space; instead, it factorizes execution into scientific doctrine, workflow templates, and atomic tool operations.
3. Runtime execution, validation, and infrastructure
At runtime, MolClaw follows a top-down execution flow rather than exposing a flat list of tools. The sequence is explicitly described as: first, read L3; second, match the task to an L2 workflow; third, execute L1 tools under L2 control; and fourth, verify and recover. In the recovery logic, L1 tool failures propagate upward, L2 workflows trigger fallback logic, and L3 principles enforce sanity checks, provenance collection, and convergence criteria. The paper emphasizes that iterative tasks follow an Assess → Diagnose → Design → Verify pattern with explicit quality gates and self-audits (Zhang et al., 2 Apr 2026).
Several implementation details reinforce this closed-loop design. All tools are accessed through a unified Science Context Protocol infrastructure compatible with MCP. The stated functions of this SCP/MCP-based layer are standardized invocation, GPU cluster scheduling, concurrent jobs, access control, file transfer, and provenance. Skills are not described as prompts alone; they include executable code templates, input/output schemas, parameter defaults, error handling, validation checks, and data handoff contracts.
The agent’s execution is further structured by a six-step pipeline: orchestration/planning, skill discovery and matching, argument construction and validation, skill execution, outcome synthesis and feedback, and final delivery/reporting. A three-phase system prompt governs this behavior. Phase 0 is load skills, plan, self-audit; Phase 1 is stepwise execution with quality gates; Phase 2 is pre-report verification and synthesis.
The methodology layer also encodes explicit anti-hallucination procedures. These include Count-before-report, Three-checkpoint self-audit, and Three-category information labeling separating tool-computed facts, agent interpretations, and literature-derived values. The computation-first hierarchy is ordered as direct computation, alternative computation, approximate computation, and literature last. Additional safeguards target domain-specific failure modes: a dedicated residue-mapper tool addresses residue numbering mismatch across PDB numbering, UniProt numbering, and tool-internal numbering; docking workflows enforce a minimum box size of 25 Å, progressive enlargement on failure, pocket center validation, score sanity checks, and parameter locking across optimization rounds; and iterative optimization rules require each round to answer what is improving, what strategy is used, and how improvement will be measured, together with exploration/exploitation scheduling, explicit convergence criteria, and full per-round records.
4. MolBench benchmark
MolBench is introduced as a benchmark for drug-discovery agents in three tiers: MolBench-MS for molecular screening, MolBench-MO for molecular optimization, and MolBench-E2E for end-to-end discovery. The benchmark is designed around the proposition that many decisive failures emerge only when agents must sustain long-horizon execution through iterative loops and failure recovery (Zhang et al., 2 Apr 2026).
MolBench-MS includes three tasks: property filtering, binding affinity comparison, and docking screening. These probe whether an agent can screen molecules under constraints, rank binding strengths, and perform virtual screening. MolBench-MO includes molecule editing and physicochemical property optimization, with emphasis on modifying molecules while preserving constraints and improving properties such as QED, LogP, and LogS.
MolBench-E2E contains three long-horizon tasks. E2E-Q1 is coarse-grained conformational sampling plus all-atom reconstruction. E2E-Q2 is QED-driven iterative molecular optimization. E2E-Q3 is structure-guided iterative lead optimization against EGFR. The paper states that these end-to-end tasks require 8 to 50+ sequential tool calls and, in practice, involve iterative loops and failure recovery.
This benchmark design is notable because it spans tasks that are trivially scriptable, tasks that require correct multi-step pipelines, and tasks that depend on iterative scientific control. That structure is essential for the paper’s later claim that the key capability bottleneck is workflow orchestration rather than generic code generation.
5. Reported performance
MolClaw is reported to achieve state-of-the-art performance across all MolBench dimensions, and the paper states that MolClaw-CC ranked first or tied first on all seven metrics spanning molecular screening and molecular optimization. The reported results are given for a Claude Sonnet 4.6 backend and compared against vanilla Claude Code and OpenClaw vanilla baselines (Zhang et al., 2 Apr 2026).
| Task/metric | MolClaw | Baseline examples |
|---|---|---|
| Property filtering accuracy | 98.0% | Claude Code vanilla: 98.0% |
| Property filtering F1 | 99.3% | Claude Code vanilla: 99.3% |
| Binding affinity comparison accuracy | 81.1% | Claude Code vanilla: 51.4% |
| Docking screening hits@3 | 0.80 | Claude Code vanilla: 0.56 |
| Molecule editing | 100.0% | Claude Code vanilla: 97.4% |
| Physicochemical optimization delta | 1.724 | Claude Code vanilla: 0.866 |
| Optimization success rate | 100.0% | Claude Code vanilla: 92.3% |
On MolBench-MS, property filtering shows no numerical advantage over the strongest baseline on the reported accuracy and F1 values, whereas binding affinity comparison and docking screening show large gains. OpenClaw vanilla is reported at 96.0% accuracy, 98.9% F1, 51.4% binding affinity accuracy, and 0.20 hit count on the corresponding screening metrics.
On MolBench-MO, the strongest reported gain is on physicochemical optimization delta, where MolClaw’s 1.724 is described as roughly doubling the vanilla Claude Code result of 0.866. Molecule editing reaches 100.0%, and optimization success rate is 100.0%. OpenClaw vanilla is reported at 97.4% editing, 1.293 delta, and 100.0% success.
The end-to-end tasks are presented as fully autonomous executions with recovery from failures. E2E-Q1 completed coarse-grained simulations and reconstruction. E2E-Q2 reached QED 0.799 from baseline 0.377 while respecting Tanimoto constraints. E2E-Q3 reached two molecules with docking improvement to kcal/mol, exceeding the target threshold. The paper interprets these examples as evidence of robust long-horizon scientific reasoning and execution.
6. Ablation results, interpretation, and naming distinctions
The ablation study is central to the paper’s argument because the gains are explicitly reported as non-uniform. Property filtering shows no meaningful gain; molecule editing shows marginal gain; binding affinity comparison shows large gain; physicochemical optimization shows large gain; and docking screening shows clear gain. The accompanying interpretation is direct: property filtering can be solved by ad hoc RDKit scripting, so vanilla agents already do well, whereas binding affinity comparison requires a correct multi-step pipeline that vanilla agents do not reliably orchestrate. Biomni is cited in the paper as exhibiting a similar pattern, doing well on simple scripting tasks but failing on docking and binding affinity tasks. The statistical support reported includes a significant Friedman test across methods, a MolClaw-CC average rank of 1.5/12, many large Cohen’s effect sizes, Fisher’s exact tests showing significant superiority on many metrics, and Wilson confidence intervals together with Mann–Whitney tests supporting robustness (Zhang et al., 2 Apr 2026).
These results suggest that MolClaw’s principal contribution is not a generic increase in tool-using ability, but a specialization in validated workflow composition under scientific constraints. The paper’s core conclusion is therefore that tool access is commoditized, while workflow orchestration is the bottleneck.
A recurrent source of confusion is the similarity of the name “MolClaw” to other molecular ML systems. “MolCluster” is a different method: an unsupervised model for coarse-grained molecular mapping that integrates a graph neural network with Leiden community detection and supports predefined group preservation and tunable resolution (Zhong et al., 25 Sep 2025). “MoleCLUEs” is also distinct: a jointly trained E3NN-based latent counterfactual method that optimizes molecular conformers by gradient descent on differentiable epistemic and aleatoric uncertainty estimates (Maser et al., 2023). Neither of those papers uses “MolClaw” as the method name. In the present context, MolClaw refers specifically to the autonomous agent for drug molecule evaluation, screening, and optimization.