Molclaw Docking Screening

Large-Scale Docking Screening Skill

Note:

• Local files are not directly accessible by the server. Please upload them to the server using molclaw-file-transfer before execution.
• For PDB file inputs, it is recommended to preprocess them using molclaw-pdbfixer before execution.
• Please refer to skill molclaw-scp-server to complete tool invocation.

Name

molclaw-docking-screening

Description

This skill performs autonomous, large-scale virtual screening for a protein target using a soft pipeline:

1. Drug-likeness filtering (QED and Lipinski)
2. QuickVina2
3. EquiScore
4. Consensus ranking via rank aggregation

It is designed for 10+ molecules and should adapt strategy to input size, target quality, and tool outcomes.

Use this skill when:

• The task is virtual screening for 10+ ligands.
• The user asks for ranking, prioritization, or top-hit selection.
• You need balanced use of physics-based docking and ML rescoring.

Workflow Steps

Stage 0. Input Validation and Setup

• Validate SMILES list is non-empty and count >= 10 for this skill. If = 0.2 and Lipinski violations 200 molecules: apply stronger triage and keep a broad but manageable subset for docking.
• If survivors 0 (and/or summary stats like max/min/mean/median), treat rescoring as successful; if predictionspath is not directly readable, fetch it via molclaw-file-transfer (serverfiletobase64`) and continue.

Stage 5. Consensus Ranking

Do not directly add raw QuickVina and EquiScore scores.

1. Rank all successfully docked molecules by Vina affinity ascending (more negative is better): rank_vina.
2. Rank all available EquiScore predictions descending (larger is better): rank_equiscore.
3. For molecules with missing EquiScore, assign worst EquiScore rank plus penalty.
4. Compute fused rank sum:

• fusedrank = w1 rankvina + w2 rank_equiscore

1. Sort fused_rank ascending (smaller is better).

Recommended default weights:

• w1 = 1.0
• w2 = 1.0

Notes

• Follow question-specific output requirements first.
• Keep score direction consistent: QuickVina lower is better, EquiScore higher is better.
• Use rank aggregation or z-score fusion; avoid directly summing raw scores.
• Some of steps depend on other skills, Please refer carefully.
• Save all intermediate scripts, and save the results (the complete Claude Code reasoning process and output for this question) as result.md.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: InternScience
• Source: InternScience/MolClaw
• License: MIT

Install and usage instructions live in the source repository linked above.