DockQ

What is DockQ?

DockQ is a continuous quality measure for evaluating protein docking models. When you predict how proteins interact using docking tools like AutoDock Vina or DiffDock, you need a way to assess whether your predicted complex matches the real (native) structure. DockQ provides this assessment by combining multiple accuracy metrics into a single score between 0 and 1.

The key advantage of DockQ over traditional CAPRI criteria is its continuous scale. Instead of coarse classifications like "acceptable" or "medium," DockQ produces a precise score that allows you to rank models, train machine learning algorithms, and track incremental improvements in docking quality.

DockQ evaluates all interfaces in your complex simultaneously, making it suitable for multi-chain assemblies, not just binary protein-protein interactions. It also supports protein-nucleic acid and small molecule interfaces.

How does DockQ work?

DockQ combines three established CAPRI quality measures into a single score using the formula:

$\text{DockQ} = \frac{f_{\text{nat}} + \text{scaled}(\text{LRMSD}) + \text{scaled}(\text{iRMSD})}{3}$

Each component captures a different aspect of docking accuracy:

Fraction of native contacts ($f_{\text{nat}}$) measures how many of the true interface contacts your model recovers. A contact exists when any atom from one chain is within 5 of any atom from another chain. This metric directly answers: "Did you find the right binding interface?"

Ligand RMSD (LRMSD) measures the positional displacement of the smaller chain (ligand) after superimposing the larger chain (receptor). This captures how well you placed the ligand in 3D space relative to the receptor.

Interface RMSD (iRMSD) measures the RMSD of interfacial residues only, after superimposing the interface atoms. This metric is stricter than LRMSD because it focuses exclusively on the binding region.

The scaling function

Raw RMSD values can range from 0 to infinity, which makes them difficult to combine with $f_{\text{nat}}$ (bounded 0-1). DockQ addresses this using inverse square scaling:

$\text{scaled}(\text{RMS}, d) = \frac{1}{1 + (\text{RMS}/d)^2}$

The scaling parameters were optimized on a training set of 56,015 docking models:

• $d_1 = 8.5 \text{ }$ for LRMSD
• $d_2 = 1.5 \text{ }$ for iRMSD

This scaling smoothly maps large RMSD values toward zero while preserving sensitivity in the accurate range.

Quality classification thresholds

DockQ scores map to qualitative categories using thresholds optimized for consistency with CAPRI assessments:

These thresholds achieve 91-97% precision at 90% recall across all quality classes, significantly outperforming the earlier IS-score method (66-71% precision).

Understanding the output

DockQ reports metrics for each interface in your complex:

Interface: The chain pair being evaluated (e.g., "A:B" for the interface between chains A and B). Multi-chain complexes produce multiple rows.

DockQ: The combined quality score (0-1). This is your primary metric for ranking models.

Quality: Categorical classification (Incorrect/Acceptable/Medium/High) based on the thresholds above.

iRMSD: Interface RMSD in Angstroms. Values under 1 indicate excellent interface geometry; values above 4 suggest significant structural differences.

LRMSD: Ligand RMSD in Angstroms. Measures global positioning error of the smaller chain. Values under 5 are generally acceptable.

fnat: Fraction of native contacts recovered (0-1). Values above 0.5 indicate the correct binding mode was identified.

fnonnat: Fraction of predicted contacts that are non-native (false positives). Lower is better. High values suggest your model creates artificial contacts.

F1: Harmonic mean of precision and recall for interface contacts. Balances $f_{\text{nat}}$ (recall) against the complementary precision measure.

Clashes: Number of interfacial residue pairs with atoms closer than 2 . Non-zero values indicate steric problems in your model.

When to use DockQ

After docking: Evaluate binding poses from AutoDock Vina, DiffDock, GNINA, or LightDock against a known crystal structure. DockQ helps you select the best pose when multiple candidates exist.

Validating predicted structures: Compare AlphaFold 2 or Boltz-2 complex predictions against experimentally determined structures. DockQ provides more nuanced feedback than simple RMSD calculations.

Benchmarking methods: When developing or comparing docking algorithms, DockQ provides a standardized metric that correlates with CAPRI assessments.

Quality control: Verify that homology models of protein complexes preserve the interface geometry of their templates.

When to use alternatives

For overall structure comparison (not docking-specific), use our RMSD calculator with Kabsch alignment. For comprehensive single-structure validation including geometry checks, use MolProbity.

Input requirements

DockQ requires two structures in PDB or mmCIF format:

Model structure: Your predicted or docked complex. Must contain all chains involved in the interfaces you want to evaluate.

Native structure: The reference (ground truth) complex. Chain naming doesn't need to match the model—DockQ performs automatic chain mapping.

Both structures should represent the same or highly similar proteins. Sequences don't need to be identical, but significant length differences may affect alignment.

Advanced options

Chain mapping

By default, DockQ automatically determines the optimal correspondence between model and native chains by maximizing the average DockQ score. This works well when chains are similar length.

For explicit control, use the chain mapping option with format MODEL:NATIVE:

• AB:AB - Map model chains A,B to native chains A,B
• AB:HL - Map model chains A,B to native chains H,L (common for antibodies)
• A*:W* - Wildcard mapping when chain names differ

Small molecule mode

Enable this for evaluating protein-ligand docking poses. Small molecule mode uses LRMSD as the primary metric since interface contacts are less meaningful for non-protein ligands. Requires mmCIF format for proper chain identification.

Skip alignment

Only enable this if your structures are already superimposed in the same coordinate frame. By default, DockQ performs sequence alignment to establish residue correspondence. Skipping alignment uses residue numbering directly, which can fail if numbering differs between structures.

Allowed mismatches

Increases tolerance for sequence differences during chain mapping. Useful when comparing structures with point mutations or slight sequence variations. The default (0) requires exact matches.

Example workflow

Consider evaluating a docking prediction for the barnase-barstar complex:

1. Download the crystal structure (PDB: 1BRS) as your native reference
2. Generate docking poses using AutoDock Vina or similar
3. Upload your best docked model and the 1BRS structure to DockQ
4. Interpret results:
   • DockQ > 0.8 with fnat > 0.8: Excellent pose, likely correct binding mode
   • DockQ 0.5-0.8 with iRMSD < 2: Good interface but some positioning error
   • DockQ < 0.23: Incorrect binding mode—try alternative poses

Common issues and solutions

"No valid interfaces found": Both structures must have at least two chains in contact. Single-chain structures cannot form interfaces for DockQ to evaluate.

Low DockQ despite visual similarity: Check chain mapping. Automatic mapping may assign chains incorrectly when multiple similar chains exist. Use explicit mapping.

High fnonnat with reasonable DockQ: Your model creates additional contacts not present in the native structure. This often indicates slight rotation or translation of the ligand.

Clashes > 0: Steric clashes indicate physically unrealistic overlap. Consider energy minimization or selecting alternative poses.

FAQ

Is DockQ suitable for protein-DNA complexes?

Yes. DockQ v2 (used here) natively supports protein-nucleic acid interfaces. The same quality thresholds apply, though interpretation may require domain knowledge about typical DNA-protein binding modes.

Can I evaluate flexible docking results?

Yes, but with caveats. DockQ compares static structures. If your docking protocol included receptor flexibility, the native structure should ideally match the induced-fit conformation, not the unbound receptor.

How does DockQ handle symmetric complexes?

DockQ automatically tries different chain mappings and reports the one that maximizes the average score. For homodimers, this means it finds the correct subunit correspondence automatically.

What DockQ score should I aim for?

For drug discovery applications, aim for DockQ > 0.5 (Medium quality). This typically indicates the correct binding site and approximate pose, sufficient for lead optimization. For structural biology applications requiring atomic-level accuracy, aim for DockQ > 0.8.

Can I compare models without a native structure?

No. DockQ requires a reference structure. For ranking docking poses without a native structure, rely on the scoring function from your docking tool (e.g., Vina affinity scores) or consensus approaches.

References

Basu S, Wallner B. (2016). DockQ: A Quality Measure for Protein-Protein Docking Models. PLoS ONE 11(8): e0161879. DOI: 10.1371/journal.pone.0161879