---
title: DiffDock outputs and ranking
description: 'Ranked pose SDF and PDB files plus results.json with poseId, rank and score — why that score is pose accuracy within 2 Å, not affinity, and which to keep.'
---

DiffDock returns ranked binding poses with a confidence score for each — the confidence tells you *where* the ligand binds, not how strongly. Use it to pick the top few poses worth minimizing with [MM-GBSA](/docs/mmgbsa). Note that MM-GBSA here returns no binding free energy, so neither tool answers how strongly the ligand binds.

## Output artifacts

DiffDock generates multiple binding poses per ligand by sampling from a learned diffusion process. Each pose represents a complete ligand placement including position, orientation, and torsion angles. The poses are ranked by a confidence model that predicts the likelihood of each pose being close to the true binding mode.

- **Pose files:** SDF for ligands and PDB for complexes; one record per pose.
- **`results.json`:** structured metadata with poses and scores.
- **Visualization previews:** quick-look images for the top-ranked poses (when
  enabled).

## Key fields in `results.json`

- **`poseId`**: unique ID per pose.
- **`ligandId`**: your submitted ligand identifier (if provided).
- **`rank`**: 1-based ordering by model confidence.
- **`score`**: DiffDock confidence score (higher is better).

There's no reference-ligand input on this tool, so no RMSD-to-a-known-pose is computed or returned — validate a pose against a known binding mode by inspection instead.

## Interpreting rankings

DiffDock's confidence model is trained to predict whether a pose is within 2 Angstroms RMSD of the true binding mode. How often the top-ranked pose is actually correct depends heavily on how close your target is to the structures the model was trained on — reported success rates range from 43% on standard benchmarks down to 12% on complexes no method had seen before. See [how DiffDock results compare to the bench](/docs/diffdock/translatability) before treating a top pose as correct.

- Focus on the top 3–5 poses for downstream inspection; quality typically drops
  after that.
- Low scores across all poses often signal poor ligand geometry or an
  unsuitable protein preparation.
- The confidence score reflects the model's estimate of pose accuracy, not binding affinity. DiffDock does not compute a binding energy, and no tool on this platform supplies one — [MM-GBSA](/docs/mmgbsa) minimizes the complex and reports its energy but returns no binding free energy. A high-confidence pose is a geometric hypothesis; affinity needs an assay.

## Filtering tips

- Discard poses with severe clashes on quick visual inspection before scoring
  downstream.
- If multiple poses share similar scores, prioritize those with better
  hydrogen-bonding networks and fewer strained torsions.
- Consider running a physics-based minimization on the top poses to resolve
  minor clashes.

## Canvas Visualization

When a DiffDock job completes, a result card appears on the canvas.

- **Card size**: 400×450px
- **Border color**: Purple (#9C27B0)
- **Visualization**: Interactive 3D Mol* viewer with protein surface and ligand overlay

### What You See

- Confidence score (%) — model's estimate of pose correctness
- Number of poses generated
- Pose gallery to cycle through predicted binding modes

### Interpreting Canvas Metrics

| Metric | Meaning |
|--------|---------|
| High confidence | Pose is geometrically plausible — likely close to the true binding mode |
| Low confidence across all poses | Check ligand geometry or protein preparation; consider [MM-GBSA](/docs/mmgbsa) to rescore |

DiffDock does not report a binding affinity or energy value on the canvas, and no downstream tool here supplies one — [MM-GBSA](/docs/mmgbsa) returns complex energies, not a binding free energy.

> Confidence measures pose accuracy, not binding strength. A high-confidence pose tells you *where* the molecule binds, not *how strongly*.

## Read next