Skills diffdock

🧬

diffdock

Name: diffdock
Author: K-Dense-AI

Safe ⚙️ External commands🌐 Network access📁 Filesystem access⚡ Contains scripts

Predict molecular binding poses with AI docking

Also available from: davila7

DiffDock uses advanced diffusion models to predict how small molecules bind to proteins in 3D space. Researchers can accelerate drug discovery by generating accurate binding poses with confidence scores for structure-based drug design.

Supports: Claude Codex Code(CC)

🥈 79 Silver

Download the skill ZIP

Upload in Claude

Go to Settings → Capabilities → Skills → Upload skill

Toggle on and start using

Test it

Using "diffdock". Dock aspirin to COX-2 protein

Expected outcome:

Generated 10 binding poses for aspirin-COX-2 complex
Top prediction confidence: 0.85 (High confidence)
Binding site: Active site near residues Arg120 and Tyr355
Review recommended: Visualize top 3 poses for structural plausibility

Using "diffdock". Screen library of 100 fragments against kinase target

Expected outcome:

Processed 100 ligand-protein complexes with 20 samples each
Mean processing time: 45 seconds per complex on GPU
High confidence hits: 12 compounds with score above 0
Top 5 hits exported to screening_hits.csv

Security Audit

Safe

v4 • 1/17/2026

The static analysis flagged 295 potential issues, but ALL are FALSE POSITIVES. The scanner incorrectly identified scientific protein sequences (GFP containing 'SAM') as Windows SAM database references, scientific paper citations as weak cryptographic algorithms, standard Python loops as C2 beacon patterns, and markdown code block syntax as shell execution. This is a legitimate molecular docking research tool with no malicious intent or security vulnerabilities.

Files scanned

2,493

Lines analyzed

findings

Total audits

Risk Factors

⚙️ External commands (4)

SKILL.md:42-470 references/confidence_and_limitations.md:146-166 references/parameters_reference.md:8-159 references/workflows_examples.md:9-377

🌐 Network access (2)

SKILL.md:53-480 scripts/setup_check.py:179-194

📁 Filesystem access (2)

scripts/analyze_results.py:214 references/workflows_examples.md:290

⚡ Contains scripts (1)

scripts/setup_check.py:40

Audited by: claude View Audit History →

Quality Score

Architecture

100

Maintainability

Content

Community

100

Security

Spec Compliance

What You Can Build

Virtual screening campaigns

Screen thousands of compounds against target proteins to identify promising drug candidates for further study

Binding site prediction

Predict where small molecules bind to protein structures to understand mechanisms and guide experiments

Lead optimization

Generate binding poses for compound modifications to improve interactions with target proteins

Try These Prompts

Basic single docking

Dock the ligand COc1ccc(C(=O)Nc2ccccc2)cc1 to the protein in protein.pdb and save results to results/docking/

Batch screening

Create a batch CSV for screening 50 compounds against protein.pdb, then run DiffDock with 20 samples per complex

Analyze results

Analyze the confidence scores from DiffDock results in results/batch/ and show the top 10 predictions

Custom parameters

Create a custom config for docking flexible ligands with increased temperature and 30 inference steps

Best Practices

Always validate environment with setup_check.py before large batch jobs
Use 20-40 samples per complex for important predictions
Combine with scoring functions like GNINA for affinity estimation
Visualize top 3-5 poses to check for structural plausibility

Avoid

Using confidence scores as direct binding affinity measurements
Running large virtual screening without GPU access
Assuming single prediction is correct without examining alternatives
Ignoring protein preparation and missing residue issues

Frequently Asked Questions

What is the difference between confidence and affinity?

Confidence measures how certain the model is about the predicted pose. Affinity measures binding strength. High confidence does not mean strong binding.

How many samples should I generate per complex?

Use 10 for quick screening, 20-40 for important predictions, and 40+ for very flexible or large ligands.

Can I use protein sequences instead of PDB files?

Yes, DiffDock uses ESMFold to predict protein structure from sequence. PDB files typically give better results though.

What ligand formats are supported?

SMILES strings, SDF files, and MOL2 files. SMILES are most convenient for high-throughput screening.

How do I interpret negative confidence scores?

Scores below -1.5 indicate low confidence. Consider more samples, ensemble docking, or experimental validation.

Can DiffDock predict binding affinity?

No, DiffDock only predicts binding poses. Use GNINA, MM/GBSA, or experimental methods for affinity prediction.