DockQ
v2.1.3Assess docking quality between model and native structures
Assess docking quality between model and native structures
The DockQ score is a number from 0 to 1 that measures how closely each interface in a model complex matches the same interface in a reference structure, usually an experimental one. DockQ online on ProteinIQ runs DockQ v2 (release 2.1.3) on a model and a reference in PDB or mmCIF format. An AlphaFold2 model of the GCN4 leucine zipper scored 0.935 against the 2ZTA crystal structure in the example job below, a High quality interface that recovers 97.9% of the native contacts.
Basu and Wallner introduced DockQ in 2016 to merge the three measures of the CAPRI docking assessment, the fraction of native contacts (fnat), interface RMSD (iRMSD) and ligand RMSD (LRMSD), into one continuous DockQ score. DockQ v2, published by Mirabello and Wallner in 2024, scores every interface of a multi-chain complex with automatic chain mapping and adds protein-nucleic acid and small-molecule interfaces. Each job returns one row per interface with its CAPRI quality class.
A DockQ comparison of one model with one reference costs 5 credits. The exact quote appears before submission.
The price does not depend on structure size, chain count or the number of models in a file, and chain mapping, Small molecule mode, CAPRI peptide mode, Skip alignment and Allowed mismatches do not change it either. Batch and workflow comparisons are quoted separately for each model and reference pair.
| Input | Accepted files | Limits |
|---|---|---|
| Model structure | PDB (.pdb, .ent) or mmCIF (.cif) | Required; 50 MiB per file; at least 2 chains |
| Native/reference structure | PDB (.pdb, .ent) or mmCIF (.cif) | Required; 50 MiB per file; at least 2 chains |
The model is the predicted or docked complex and must contain every chain of the interfaces to be scored. The reference is the experimental or trusted complex. Chain names do not need to match, because DockQ maps chains automatically. Both inputs can fetch an entry from RCSB and trim a PDB structure before submission. An optional Job name labels the run without changing the analysis.
Compressed .gz files must be decompressed before upload. Only the first coordinate model of each file is scored, so an ensemble is not evaluated as a set of models.
Both structures should represent the same or highly similar complexes. DockQ decides chain correspondence from its own sequence alignment. Missing residues and aligned gaps are accepted with the default mismatch tolerance of zero, so complete sequence identity is not required, but larger differences can still end in a homology or interface error.
A small-molecule pose from a docking tool such as AutoDock Vina, GNINA or DiffDock comes as a separate file. For DockQ, the pose must be merged into the receptor file as its own chain, in both the model and the reference, before upload.
| Parameter | Type | Default | Description |
|---|---|---|---|
Chain mapping (chain_mapping) | string | Empty, automatic | A single MODEL:NATIVE mapping; fixed chain IDs must exist in the selected mode. |
Small molecule mode (small_molecule) | boolean | false | Include supported heteroatom chains and use native small-molecule scoring for their interfaces. |
CAPRI peptide mode (capri_peptide) | boolean | false | Use experimental peptide contact and interface definitions described below. |
Skip alignment (no_align) | boolean | false | Use residue numbering instead of sequence alignment; structural superposition remains active. |
Allowed mismatches (allowed_mismatches) | integer | 0 | Allow 0 to 10 aligned substitutions during automatic chain matching. |
By default, DockQ tries chain correspondences between model and reference and keeps the one with the highest total DockQ over the reference interfaces. This also settles homodimers and other complexes with identical chains, where any copy could match.
For explicit control, use the MODEL:NATIVE format:
AB:AB maps model chains A and B to native chains A and BAB:HL maps model chains A and B to native chains H and L, common for antibodiesA*:W* fixes model chain A to native chain W and lets DockQ map the rest:BC or *:BC scores only native chains B and C while optimizing their model correspondenceMapping notation uses one character per chain and exactly one colon. Comma-separated pairs such as A:A,B:B are invalid. For mmCIF inputs, standard mode uses author chain IDs (auth_asym_id), while Small molecule mode uses label chain IDs (label_asym_id). Standard mode excludes HETATM-only chains.
Enable this to score supported small-molecule poses. The ligand must already be present in the uploaded PDB or mmCIF file as a separate chain. For these interfaces DockQ reports a symmetry-corrected LRMSD, an atom mapping, and a DockQ value derived from LRMSD alone, explained under small-molecule scores below.
Polymer metrics such as iRMSD, fnat, fnonnat, F1 and clashes stay empty for small-molecule interfaces, and no Quality class is assigned. Standard polymer interfaces in the same complex keep their usual metrics.
CAPRI peptide mode is off by default. Enabling it changes the native-contact cutoff from 5 Å to 4 Å and defines interface residues using Cβ distances below 8 Å, with Cα used when Cβ is absent, instead of the standard 10 Å heavy-atom interface cutoff. iRMSD is still calculated from the selected residues' backbone atoms, and the DockQ formula and its scaling constants are unchanged.
The option is experimental. The DockQ 2.1.3 help text says DockQ cannot be trusted with this setting, so the usual quality classes are not a validated peptide benchmark. The option changes polymer-interface scoring; small-molecule interfaces keep their LRMSD-based calculation. See the pinned DockQ 2.1.3 scoring code and distance cutoffs.
Enable this only when residue numbering provides the intended correspondence between the structures. It skips sequence alignment and uses residue numbers directly. DockQ still performs the structural superposition needed for RMSD scoring, so the inputs do not need to share a coordinate frame.
This setting raises the tolerance for sequence differences during chain mapping, which helps when the model and reference differ by point mutations. The default of 0 allows no aligned substitutions; aligned gaps are handled by DockQ and do not need a positive tolerance. The setting accepts integers from 0 to 10.
| Output | Contents |
|---|---|
Results | One row per interface with DockQ, Quality, iRMSD, LRMSD, fnat, fnonnat, F1, clashes and contact counts |
dockq_results.json | DockQ's complete native output: scores at full precision, interface and atom mappings, and the global mapping |
dockq_stdout.log | DockQ's own text report, including the total over all native interfaces |
dockq_stderr.log | DockQ's diagnostic stream, kept for failed and successful runs |
Jobs submitted since the current result format keep full native precision in the results table and workflow rows. Older jobs keep their original rounded table values; the native JSON remains the source for exact values.
Each row is one interface, named by its pair of model chains (A:B for the interface between chains A and B). A dimer gives one row; a complex with more chains gives one row for each pair of chains that are in contact in the reference.
| Column | Meaning |
|---|---|
DockQ | The combined score from 0 to 1, the primary number for ranking models |
Quality | CAPRI class from the DockQ value: Incorrect, Acceptable, Medium or High |
iRMSD (A) | Interface RMSD in Å over the backbone of the reference interface residues |
LRMSD (A) | Ligand RMSD in Å: the error in placing the smaller partner once the larger one is superposed |
fnat | Fraction of reference contacts that the model reproduces, from 0 to 1 |
fnonnat | Fraction of model contacts that are absent from the reference; lower is better |
F1 | Harmonic mean of contact recall (fnat) and precision |
Clashes | Interface residue pairs with atoms closer than 2 Å |
Native contacts recovered, Native contacts total | The counts behind fnat (nat_correct of nat_total) |
Non-native contacts, Model contacts total | The counts behind fnonnat (nonnat_count of model_total) |
Chain 1 residues, Chain 2 residues, Chain 1 role | Chain lengths and which side DockQ treated as the receptor (the larger partner) and as the ligand (len1, len2, class1, class2) |
Small molecule | is_het, true for a small-molecule interface |
A missing metric stays empty; it does not mean zero.
The quality classes come from the CAPRI assessment, which classifies a model from fnat, LRMSD and iRMSD. Basu and Wallner chose the DockQ cutoffs 0.23, 0.49 and 0.80 to reproduce those classes, and on CAPRI models DockQ recovered them with an average positive predictive value of 94% at 90% recall. DockQ v2 prints the same bands, and ProteinIQ assigns Quality from them.
| Class | DockQ | CAPRI criteria |
|---|---|---|
| Incorrect | 0 to below 0.23 | fnat below 0.1, or LRMSD above 10 Å and iRMSD above 4 Å |
| Acceptable | 0.23 to below 0.49 | fnat at least 0.1, and LRMSD up to 10 Å or iRMSD up to 4 Å |
| Medium | 0.49 to below 0.80 | fnat at least 0.3, and LRMSD up to 5 Å or iRMSD up to 2 Å |
| High | 0.80 to 1 | fnat at least 0.5, and LRMSD up to 1 Å or iRMSD up to 1 Å |
A model takes the highest CAPRI class whose conditions it meets. The DockQ bands reproduce that classification closely but not exactly, so a model near 0.23 or 0.49 can fall on the other side of a CAPRI boundary. Most structure prediction benchmarks count a model with DockQ of 0.23 or more as a success.
fnat counts contacts. A contact is a pair of residues from the two chains with heavy atoms within 5 Å in the reference, and fnat is the share of those pairs that are also in contact in the model. It says whether the right residues touch, regardless of exact positions.
iRMSD measures local geometry. Residues with heavy atoms within 10 Å of the partner in the reference define the interface, and iRMSD is the backbone RMSD of those residues after superposing them. Values around 1 Å mean the interface is close to atomic accuracy; above 4 Å, the interface is only roughly placed.
LRMSD measures global placement. DockQ superposes the larger partner of the model on the reference (the receptor) and measures the backbone RMSD of the smaller partner (the ligand) without moving it again. A small rotation of the ligand chain around a correct contact patch shows up mostly here.
DockQ averages fnat with the two RMSD values converted to scores between 0 and 1:
An iRMSD of 1.5 Å or an LRMSD of 8.5 Å each earns half of its term. Because the three terms are averaged, the same DockQ can come from different mixtures of errors, so read fnat, iRMSD and LRMSD alongside it. The RMSD guide covers superposition and RMSD in general.
fnonnat and F1 add the other direction. A high fnonnat with a reasonable DockQ means the model forms extra contacts that the reference does not show, often from a slightly rotated or shifted partner. Clashes above 0 point to overlapping atoms at the interface, which energy minimization or a different pose can fix.
For complexes with more than two chains, DockQ scores every interface that exists in the reference and reports GlobalDockQ, the mean over those interfaces, in the native report. The native best_dockq field is the sum of the interface DockQ values for the chosen mapping, not a maximum. ProteinIQ's best-interface and average summaries are separate from both.
GlobalDockQ can hide one wrong interface among several good ones. In a four-chain transcription factor and DNA complex, check the lowest interface score as well as the mean.
For a small-molecule interface, DockQ v2 sets the fnat term to 0 and the iRMSD term to 1, so the score depends on LRMSD alone and runs from about 0.33 to 0.67 instead of 0 to 1. A ligand placed exactly scores 0.667, and one at 2 Å LRMSD, the usual threshold for a correct docking pose, scores 0.649.
These values cannot be read against the polymer bands, which is why ProteinIQ leaves Quality empty for them. Read the LRMSD itself instead. The symmetry-corrected LRMSD in the native JSON tries every graph-equivalent atom mapping, so a symmetric ring placed flipped is not penalized.
DockQ could not find any dockable interface in the reference structure for the selected chains means no pair of the selected reference chains is in contact. Check that the reference contains the complex, not a single chain, or set a chain mapping to the interface you want.
DockQ could not identify homologous chains between the model and reference structures means the sequences differ too much for automatic matching. Raise Allowed mismatches for point mutations, or fix the mapping for the homologous chains.
A low DockQ on a model that looks right usually comes from chain mapping in complexes with similar chains, or from a model that is correct locally but whose partner is rotated, which shows as a high LRMSD with a reasonable fnat.
DockQ needs a reference structure, because it measures error against known coordinates. pDockQ, pDockQ2 and ipSAE are reference-free: they estimate interface quality from a prediction's own confidence, for complexes where no experimental structure exists.
pDockQ (Bryant et al. 2022) combines interface pLDDT with the number of interface contacts and was fitted to DockQ, so 0.23 is again the usual cutoff. It cannot exceed 0.742 and ignores PAE. pDockQ2 (Zhu et al. 2023) replaces the contact count with the PAE between contacting residues and spans almost the full 0 to 1 range. ipSAE (Dunbrack 2025) is a TM-score-like interface confidence built from low-PAE residue pairs and is read in 0.6 to 0.8 bands.
Use DockQ to benchmark a predictor or docking method on complexes with known structures. Without a reference, the ipSAE tool calculates pDockQ, pDockQ2, ipSAE, ipTM and LIS from an AlphaFold2, AlphaFold3 or Boltz prediction, and the ipSAE guide explains how to read them. None of the three predicted scores is defined for protein-ligand complexes.
An AlphaFold2 model of the GCN4 leucine zipper homodimer, two chains of 33 residues, compared with the 1.8 Å crystal structure 2ZTA with default settings. The interface scores DockQ 0.935, High quality, with fnat 0.979. Automatic chain mapping chooses which model chain corresponds to which identical crystal chain.
A Boltz-2 model of the 97-residue U1A RNA-binding domain with a 21-nucleotide RNA hairpin, compared with 1URN, a 1.92 Å crystal structure that holds three copies of the complex. The protein-RNA interface scores DockQ 0.916, High quality, which shows the polymer classes applied to a protein-nucleic acid interface.
A Boltz-2 model of a bHLH transcription factor dimer bound to a DNA duplex, four chains in all, compared with the 2.7 Å crystal structure 5GNJ. DockQ scores six interfaces, protein-protein, protein-DNA and DNA-DNA, and reports a GlobalDockQ of 0.837 over them.
DockQ is distributed by the Wallner lab as an open-source Python package: pip install DockQ installs the command-line tool, and the source is on GitHub. The latest release on PyPI is 2.1.3, the version ProteinIQ runs. The authors do not document an upload server; their wallnerlab.org/DockQ address redirects to the repository.
On the command line, DockQ model.pdb native.pdb scores a pair, --mapping fixes chains, --small_molecule, --capri_peptide, --no_align and --allowed_mismatches match the settings above, and --json writes the full result. ProteinIQ runs that same command with --json and maps each setting to its flag, so the scores are those a local run of 2.1.3 gives on the same files.
The hosted tool removes the Python setup and shows the per-interface table in the browser. It also adds RCSB fetching and trimming of the inputs, job history with inputs, settings and files, and workflows that score every model against every reference. Paid plans can submit the same jobs through the ProteinIQ API. Locally, DockQ can also use several CPU cores for chain mapping (--n_cpu), which matters for very large multimers.
DockQ compares static structures, and only the first model of each file. It measures agreement with one reference, so a model of a different but real binding mode, or of an induced-fit conformation absent from the reference, scores low. On ProteinIQ, DockQ runs on one CPU and stops after 4 minutes, so for large complexes with many identical chains an explicit chain mapping keeps the search small.
For whole-structure comparison without interface weighting, use the RMSD calculator or the TM-score from US-align. For the geometry of a single model, including clashes beyond the interface, use MolProbity. Models for DockQ commonly come from protein-protein docking tools such as HADDOCK3, LightDock and DiffDock-PP, or from complex predictors such as AlphaFold2 and Boltz-2.
For polymer interfaces, 0.23 or more is Acceptable, 0.49 or more is Medium and 0.80 or more is High quality. Higher scores mean closer agreement with the chosen reference, but a DockQ score alone does not establish binding affinity, biological activity or suitability for lead optimization. CAPRI peptide and small-molecule results need the mode-specific reading described above.
Yes. DockQ v2.1.3 supports protein-nucleic acid and nucleic acid-nucleic acid interfaces, and the results table applies the same thresholds to these polymer interfaces. Interpretation still depends on the biology of the complex.
Yes, with a caveat. DockQ compares static structures, so if the docking protocol included receptor flexibility, the reference should ideally match the induced-fit conformation rather than the unbound receptor.
DockQ tries chain mappings and keeps the one with the highest total score, so for homodimers and other symmetric complexes it finds the subunit correspondence automatically.
No. DockQ requires a reference structure. Without one, use pDockQ2 and ipSAE for predicted complexes, or the scoring function of the docking tool, such as the Vina affinity, for docking poses.
Validate generated and docked ligand poses with structure, chemistry, and reference-pose checks.
Validate protein structures with clashscore, Ramachandran, rotamer, and geometry checks.
Score every protein interface with ESM-2 and a graph neural network
Score interprotein interactions in AlphaFold and Boltz predictions
Generate structural summary reports and diagrams for a single protein PDB structure.
Pairwise RMSD matrices for PDB structure ensembles
Calculate RMSD between protein structures with automatic alignment
Profile protein-ligand interactions from a PDB complex structure.
Calculate protein-ligand interaction fingerprints from 3D structures.
Analyze independent DNA oligos or screen a bounded set for primer interactions with Primer3 thermodynamic conditions.