DiffDock FAQ: binding affinity and confidence scores

These answers cover what DiffDock is, how many poses to request, how to read scores, how it compares to traditional docking, and where the model's accuracy falls off.

Frequently asked questions

What is DiffDock?

DiffDock is a diffusion generative model for molecular docking. It treats docking as a generative problem rather than a search/scoring problem. The model learns to reverse a diffusion process that gradually adds noise to ligand poses, enabling it to generate diverse, high-quality binding predictions. The .bloom* platform uses DiffDock-L, the latest version with state-of-the-art success rates on standard benchmarks.

Does DiffDock use a binding site?

No. This tool runs blind (whole-protein) docking only — there's no parameter to supply a box, a residue range, or a reference ligand. DiffDock's confidence model ranks poses across the full protein surface.

How many poses should I request?

10 is a good default. Increase if you need diversity; decrease for speed when triaging many ligands.

Are scores comparable across proteins?

No. Scores are most meaningful within the same target and setup. Compare poses for the same protein/box.

Why are scores low?

Often due to poor ligand geometry or missing protein cleanup. Re-check ligand sanitization and chain selection.

Can I reproduce a run?

Not exactly — there's no seed parameter, and diffusion sampling is stochastic, so the same inputs won't produce identical pose coordinates run to run. Request more poses (samples_per_complex) in a single run to cover that variability instead of relying on re-runs to match.

What if my protein is large?

Trim to relevant chains/domains. Oversized inputs slow parsing and can reduce quality.

Does DiffDock report RMSD to a known pose?

No. There's no reference-ligand input on this tool, so no RMSD against a known binding mode is computed — validate a pose against a co-crystallized ligand by visual comparison instead.

How does DiffDock compare to traditional docking?

Traditional docking methods search for optimal poses by optimizing a physics-based scoring function. DiffDock instead learns the distribution of ligand poses from crystallographic data using diffusion models. On standard benchmarks, DiffDock-L significantly outperforms traditional methods, though performance varies by target class.

Does DiffDock predict binding affinity?

No. DiffDock's confidence score predicts geometric accuracy (whether the pose is within 2 Angstroms of the true binding mode), not binding affinity. For affinity predictions, use the generated poses as input to separate scoring methods or free energy calculations.

What are the model's limitations?

DiffDock works best on drug-like small molecules. Performance may decrease for: very large ligands (>100 heavy atoms), highly flexible molecules with many rotatable bonds, covalent binders, metal-coordinating compounds, and proteins with significant induced-fit conformational changes.

Read next