DiffDock troubleshooting

Use this page to diagnose the failures you hit while running DiffDock — ligand parsing failures, poor pose quality, stalled jobs, and confidence scores that look off. Most problems trace back to input preparation rather than the model itself.

Common issues and fixes

This guide addresses common issues when running DiffDock through the .bloom* platform. Most issues stem from input preparation rather than the model itself.

Ligand parsing failures

• Sanitize ligands and remove query atoms; malformed valences are the usual culprit.
• Provide 3D coordinates for macrocycles or highly flexible ligands to reduce failures.
• Check for uncommon elements; only common drug-like atom types are supported.

Poor pose quality

• Increase inference_steps or request more poses (samples_per_complex, up to 40) to widen the search.
• Note that some targets are inherently difficult for any docking method. Consider the top 5 poses rather than only the top 1.
• For highly flexible ligands with many rotatable bonds, DiffDock's torsional diffusion may require more sampling; increase samples_per_complex toward the 40-pose maximum.

Slow or stalled jobs

• Keep proteins small and omit unrelated chains; large complexes increase runtime.
• Reduce samples_per_complex when batching many ligands.
• Split very large ligand sets into multiple submissions to avoid timeouts.

Unexpected confidence scores

• DiffDock's confidence model predicts geometric accuracy (RMSD < 2 Angstroms), not binding affinity. Low confidence does not mean weak binding.
• Confidence scores are relative within a run; do not compare scores across different proteins or ligands.
• Very low confidence across all poses often indicates the ligand is incompatible with the protein's surface geometry, or the input structure needs cleanup.

Still stuck?

• Capture the job ID, protein filename, and ligand identifiers.
• Note any custom parameters (samples_per_complex, inference_steps) and share with support.

Read next