DockQ

DockQ integrates fraction of native contacts (Fnat), interface RMSD (iRMSD), and ligand RMSD (lRMSD) into a single continuous score from 0 to 1. The formula is:

DockQ = (Fnat + 1/(1+(iRMSD/d0)^2) + 1/(1+(lRMSD/d0)^2)) / 3

• Fnat: Proportion of residue-residue contacts within 5Å that match the native structure
• iRMSD: Backbone RMSD of the interface residues after optimal superposition
• lRMSD: RMSD of the ligand after superposition on the receptor interface
• d0: A scaling factor (typically 1.5Å) that normalizes the RMSD contributions

DockQ maps continuous scores to four discrete quality categories aligned with the CAPRI (Critical Assessment of PRedicted Interactions) evaluation criteria:

• Incorrect: DockQ < 0.23 — No meaningful similarity to the native complex
• Acceptable: 0.23 ≤ DockQ < 0.49 — Correct binding region but inaccurate orientation
• Medium: 0.49 ≤ DockQ < 0.80 — Good overall geometry with minor deviations
• High: DockQ ≥ 0.80 — Near-native accuracy with atomic-level precision

This mapping allows direct comparison with CAPRI blind challenge results.

DockQ addresses critical failure modes of individual scoring metrics:

• Fnat alone can score highly even when the ligand is completely flipped, as long as interface contacts are preserved
• iRMSD alone penalizes flexible loops far from the interface, obscuring docking accuracy
• lRMSD alone is sensitive to receptor size and does not capture interface quality

By combining all three, DockQ requires simultaneous agreement in contact preservation, interface geometry, and relative orientation, preventing inflated scores from models that satisfy only one criterion.

DockQ's accuracy depends critically on the definition of the interface — the set of residues considered to be in contact. Standard practice defines interface residues as those with any heavy atom within 10Å of a residue on the binding partner in the native structure.

• Overly inclusive interfaces dilute the sensitivity of iRMSD
• Overly restrictive interfaces may miss biologically relevant contacts
• Consistent definition across benchmark datasets is essential for fair comparison

The original DockQ implementation uses the 10Å distance cutoff on the experimentally determined complex.

DockQ has become the de facto standard for evaluating protein-protein docking algorithms, used extensively in:

• CAPRI blind prediction rounds for official scoring
• Docking benchmark datasets (e.g., Docking Benchmark 5.5) for method development
• AlphaFold-Multimer evaluation to assess predicted complex accuracy
• Antibody-antigen docking assessment with specialized variants

The continuous nature of DockQ enables statistical comparison between methods using mean scores, success rates, and significance testing, rather than relying solely on categorical binning.

An enhanced variant, DockQ2, extends the original metric to address multi-chain complexes and symmetry:

• Multi-chain support: Evaluates complexes with more than two chains by decomposing into pairwise interfaces
• Symmetry-aware scoring: Correctly handles homomeric complexes where chain assignment is ambiguous
• Biological interface identification: Distinguishes crystal contacts from biologically relevant interfaces
• Improved numerical stability: Handles edge cases where the original DockQ formula becomes undefined

DockQ2 maintains backward compatibility with the original DockQ scale while extending applicability to the full PDB.

Measures the root mean square deviation of backbone atoms at the protein-protein interface after optimal superposition of the native and predicted complexes.

• Calculated only on residues within 10Å of the binding partner
• Sensitive to local binding pose accuracy
• Lower values indicate better predictions
• DockQ normalizes iRMSD using a sigmoidal function to map it to a [0,1] range

Quantifies the proportion of native residue-residue contacts recovered in the predicted complex.

• A contact is defined as two residues within 5Å across the interface
• Ranges from 0 (no native contacts) to 1 (all native contacts preserved)
• Directly captures biological relevance of the predicted interface
• DockQ uses Fnat as a primary component, weighted equally with iRMSD

Measures the global positional deviation of the ligand protein after superimposing the receptor chains of the native and predicted complexes.

• Captures large-scale orientation errors
• Complementary to iRMSD for detecting rigid-body docking failures
• DockQ incorporates lRMSD to penalize models with correct interface contacts but incorrect overall orientation

The Critical Assessment of PRedicted Interactions community benchmark classifies docking predictions into quality tiers.

• High: Fnat ≥ 0.5 and (lRMSD ≤ 1.0Å or iRMSD ≤ 1.0Å)
• Medium: Fnat ≥ 0.3 and (lRMSD ≤ 5.0Å or iRMSD ≤ 2.0Å)
• Acceptable: Fnat ≥ 0.1 and (lRMSD ≤ 10.0Å or iRMSD ≤ 4.0Å)
• Incorrect: Does not meet Acceptable criteria
• DockQ maps continuously to these CAPRI categories, with 0.8+ corresponding to High quality

DockQ combines iRMSD, Fnat, and lRMSD into a single continuous metric from 0 to 1.

• Formula: DockQ = [Fnat × f(iRMSD) × f(lRMSD)]^(1/3)
• f(x) is a sigmoidal function: f(x) = 1 / [1 + (x/d₀)²]
• d₀ parameters calibrated on CAPRI benchmark data
• Score of 0.0 = incorrect prediction; 1.0 = perfect prediction
• Enables direct comparison across docking methods without arbitrary binning

Recent extensions address limitations of the original DockQ for multimeric and symmetric complexes.

• DSSO (DockQ Score for Symmetric Oligomers): Adapts DockQ for cyclic and dihedral symmetry by evaluating interface quality across all symmetry-related chains
• DockQ2: Generalizes to arbitrary stoichiometries and handles cross-interface contacts more robustly
• Both preserve the 0-1 scale while extending applicability beyond heterodimers