DockQ is a score from 0 to 1 for how closely a model of a protein complex reproduces the interface of a reference structure, usually an experimental one. Sankar Basu and Björn Wallner introduced it in 2016 to merge the three measures used in the CAPRI docking assessment, the fraction of native contacts (fnat), the interface RMSD (iRMSD) and the ligand RMSD (LRMSD), into one continuous number. A DockQ of 0.23 or more corresponds to an acceptable model, 0.49 or more to medium quality and 0.80 or more to high quality. You can calculate it for your own model and reference with the DockQ tool on ProteinIQ.[1]
DockQ needs the answer. pDockQ and pDockQ2 estimate it for predicted complexes when no reference exists. pDockQ, from Arne Elofsson's group in 2022, uses the pLDDT of the interface residues and the number of contacts between them. pDockQ2, from the same group in 2023, adds the predicted aligned error across the interface. Both were fitted to DockQ, so they share its scale and its 0.23 line for an acceptable interface.[2][3]
The two kinds of score can disagree in ways that are easy to misread. In the worked example below, an AlphaFold-Multimer model of human hemoglobin α and β scores a DockQ of 0.967 against the crystal structure, while its pDockQ is only 0.57, a gap that comes from how pDockQ was built rather than from any problem with the model.
How DockQ is calculated
DockQ compares one interface at a time: two chains, or two groups of chains, in the model against the same pair in the reference. It starts from three measures that CAPRI standardized for assessing docking models:[1]
- fnat is the fraction of residue contacts in the reference interface that the model also has. Two residues on different chains are in contact when any of their heavy atoms are within 5 Å of each other.
- iRMSD is the backbone RMSD of the interface residues after superimposing them. Interface residues are those with a heavy atom within 10 Å of the partner chain in the reference.
- LRMSD is the backbone RMSD of the smaller chain, called the ligand, after the model is superimposed on the reference using the larger chain, called the receptor.
fnat is already on a 0 to 1 scale, but the two RMSDs have no upper limit. DockQ maps each one to a 0 to 1 value S with the inverse-square function used by the S-score, the same shape as the TM-score term, and averages the three terms:[1]
The constants, 1.5 Å for iRMSD and 8.5 Å for LRMSD, are the RMSD at which each term drops to 0.5. Basu and Wallner chose them by a grid search on 56,015 SwarmDock models of 118 complexes from Docking Benchmark 4.0, keeping the pair that best reproduced the CAPRI quality classes, with an average F1 of 0.91.[1]
| RMSD | iRMSD term (scale 1.5 Å) | LRMSD term (scale 8.5 Å) |
|---|---|---|
| 0.5 Å | 0.90 | 1.00 |
| 1 Å | 0.69 | 0.99 |
| 2 Å | 0.36 | 0.95 |
| 4 Å | 0.12 | 0.82 |
| 8.5 Å | 0.03 | 0.50 |
| 20 Å | 0.01 | 0.15 |
The table shows where DockQ is sensitive. The iRMSD term reacts to errors of 1 to 4 Å, the range that separates a good interface from a roughly placed one, while the LRMSD term only falls off when the ligand is several ångströms out of place. Past about 7 Å of iRMSD, the term is close to zero whatever the exact value, which was the intent: Basu and Wallner wanted models that are equally wrong to score equally low.[1]
DockQ quality classes
CAPRI sorts docking models into four classes using fnat together with LRMSD or iRMSD. The three DockQ cutoffs, 0.23, 0.49 and 0.80, were set to reproduce those classes on the training models, and they stayed within 0.02 of optimal on an independent set of 15,328 models submitted to CAPRI between 2005 and 2014.[1]
| Class | fnat | And either | DockQ |
|---|---|---|---|
| High | At least 0.5 | LRMSD up to 1.0 Å or iRMSD up to 1.0 Å | 0.80 to 1 |
| Medium | At least 0.3 | LRMSD up to 5.0 Å or iRMSD up to 2.0 Å | 0.49 to 0.80 |
| Acceptable | At least 0.1 | LRMSD up to 10.0 Å or iRMSD up to 4.0 Å | 0.23 to 0.49 |
| Incorrect | Below 0.1 | Or neither RMSD within the acceptable limits | Below 0.23 |
On the independent CAPRI set, DockQ reached 95%, 97% and 91% precision at 90% recall for acceptable or better, medium or better, and high-quality models. The earlier IS-score reached 71%, 72% and 66% at the same recall.[1] Most of DockQ's disagreements with CAPRI were models within 0.02 of a cutoff, and the authors present the cutoffs as a way to relate the two systems rather than as a reason to bin models. Two nearly identical models can land on opposite sides of a class boundary, and the continuous score avoids that.[1]
The 0.23 line is the one used most. Benchmarks of structure predictors usually define success as DockQ of at least 0.23, as both the pDockQ and pDockQ2 papers do.[2][3]
What changed in DockQ v2
DockQ v2, published by Claudio Mirabello and Björn Wallner in 2024, is a reimplementation in Python that keeps the original formula. On more than 17,000 models from 37 CASP15 targets, its scores differed from the first version by 0.004 on average, so results from the two versions can be compared directly.[4]
The changes are in what it can score. Version 2 scores every interface in a multi-chain reference in one run and reports GlobalDockQ, the mean over all interfaces. It also finds the chain mapping itself, trying every assignment of equivalent chains and keeping the one with the highest GlobalDockQ, which matters for homomers where any copy of a chain could match any other. On CASP15 targets, this mapping always gave the same or a higher GlobalDockQ than a mapping taken from QS-score.[4]
| Interface | What is reported | Notes |
|---|---|---|
| Protein-protein | DockQ, fnat, iRMSD, LRMSD, F1 and clashes | Same definitions as the 2016 version |
| Protein-nucleic acid and nucleic acid-nucleic acid | The same metrics as protein-protein | RMSDs use the phosphate and sugar backbone atoms of DNA and RNA |
| Protein-small molecule | LRMSD only | Heavy-atom RMSD after superimposing the receptor's interface; symmetric ligands are matched by graph isomorphism |
| Peptide (optional mode) | The same metrics as protein-protein | 4 Å contacts and an 8 Å Cβ interface; the developers caution that DockQ cannot be trusted in this mode |
For small molecules, the CAPRI classes do not apply. Read the pocket-aligned LRMSD instead, and make sure the ligand sits in its own chain in both files, which DockQ requires.[5]
Version 2 also counts clashes, defined as residue pairs across the interface with heavy atoms closer than 2 Å. Mirabello and Wallner added it because fnat ignores false contacts: a model that piles atoms into the interface can recover many native contacts, and in the extreme, a model with every atom at the origin would have an fnat of 1.[4][5]
An average over interfaces can hide one misplaced chain. In the repository's three-chain example, the interfaces score 0.994, 0.511 and 0.453, and GlobalDockQ is their mean, 0.653.[5] Zhu and colleagues found that in about one in five heteromeric AlphaFold-Multimer predictions (20.5%), only some interfaces were correct, and they judged each complex by its lowest interface DockQ.[3]
Worked example: two predicted dimers against crystal structures
AlphaFold DB began hosting predicted dimers in 2026. The NVIDIA dataset was generated with OpenFold and AlphaFold-Multimer weights, and each entry stores ipTM, ipSAE, pDockQ2 and LIS.[6] Two of those entries have well-known crystal structures to compare against. The first is human hemoglobin α with β (AF-0000000211046780), compared with the 1.74 Å structure of human deoxyhemoglobin, PDB 2HHB.[7][8] The second is the kinase CDK2 with its regulatory subunit CKS1 (AF-0000000210768563), compared with PDB 1BUH at 2.6 Å.[9][10] AlphaFold DB labels the hemoglobin β chain with the chimpanzee UniProt entry, whose sequence is identical to human β-globin.[11]
We ran DockQ 2.1.3, the version behind ProteinIQ's tool, and recalculated pDockQ and pDockQ2 from the downloaded pLDDT and PAE files with the reference implementation's 8 Å Cβ contact definition. The recalculated pDockQ2 values match the ones AlphaFold DB stores to three decimals.[12] The analysis script and data reproduce every number below.
| Score | Hemoglobin α and β vs 2HHB | CDK2 and CKS1 vs 1BUH |
|---|---|---|
| DockQ | 0.967 | 0.645 |
| CAPRI class | High | Medium |
| fnat | 0.98 (54 of 55 native contacts) | 0.87 (41 of 47 native contacts) |
| iRMSD | 0.36 Å | 3.26 Å |
| LRMSD | 1.36 Å | 3.03 Å |
| Contacts in the model (Cβ within 8 Å) | 62 | 38 |
| Mean interface pLDDT | 98.0 | 93.5 |
| pDockQ | 0.57 | 0.33 |
| pDockQ2 | 0.94 | 0.77 |
| ipTM | 0.92 | 0.88 |
| ipSAE (higher and lower direction) | 0.87 and 0.86 | 0.73 and 0.65 |
The hemoglobin model is close to the crystal structure. It recovers 54 of the 55 native contacts with an interface RMSD of 0.36 Å, and every term of DockQ is near its maximum. Scoring the same model against the other α/β contact in the tetramer, between chains A and D of 2HHB, gives a DockQ of 0.020 with none of that interface's 37 native contacts. The model is the α1β1 dimer, and DockQ scores whichever interface it is pointed at. With a multi-copy reference, check which chains the mapping paired.
CDK2 with CKS1 is the more instructive case. The model recovers 41 of 47 native contacts and places CKS1 within 3.0 Å, yet lands in the medium class. The reason is iRMSD: at 3.26 Å, its term contributes 0.18, against 0.87 from fnat and 0.89 from LRMSD. Most of that error comes from two residues of CDK2's activation segment, Val156 and Arg157, which lie 17 to 19 Å from their crystal positions after the interface superposition. The model gives residues 154 to 164 a pLDDT between 38 and 53, so it was not confident about that loop either. The residue check script lists the deviations.
Both crystal structures were public long before AlphaFold-Multimer was trained, so these are not blind tests. They show how the scores behave on the same models, not how often predictions are right.
How pDockQ estimates DockQ
pDockQ condenses the interface into one number, x, the mean pLDDT of the interface residues multiplied by the base-10 logarithm of the number of interface contacts, and passes it through a sigmoid fitted to DockQ:[2][12]
A contact is a pair of residues on different chains whose Cβ atoms are within 8 Å, and the interface residues, marked int in the formula, are all residues in at least one contact. Bryant and colleagues found that the number of contacts and the interface pLDDT each separated acceptable from incorrect models, with areas under the ROC curve of 0.92 and 0.88, and that their product with a logarithm did better, at 0.95. The average pLDDT of the whole complex managed only 0.66, because two well-folded chains can still be docked wrongly.[2]
The fit was made on FoldDock models, AlphaFold2 runs on concatenated heterodimers with paired alignments, and pDockQ predicted DockQ with an average error of 0.11 on the 1,481-complex test set. It also separated interacting from non-interacting protein pairs with an area under the curve of 0.87, finding 51% of true interactions at a 1% false positive rate.[2]
The coefficients set hard limits. pDockQ cannot rise above 0.742, the sum of 0.724 and 0.018, so it can never report a high-quality interface. It reaches 0.23 at x of about 136: with an interface pLDDT of 90, that takes about 32 contacts; with a pLDDT of 70, about 87. A small but confident interface therefore scores low. CDK2 with CKS1 has a mean interface pLDDT of 93.5 but only 38 contacts, which gives x = 147.7 and a pDockQ of 0.33 for a model whose DockQ is 0.645. Hemoglobin, with 62 contacts at a pLDDT of 98.0, reaches x = 175.6 and 0.57.
The opposite error is the more serious one. pDockQ ignores PAE, so it cannot tell whether a large, confidently folded interface is docked in the right place. When Zhu and colleagues applied it to AlphaFold-Multimer models of homomers and larger complexes, more than 10% of chains in each set had pDockQ above 0.5 and DockQ below 0.23.[3]
How pDockQ2 estimates DockQ
pDockQ2 keeps the interface pLDDT and replaces the contact count with the PAE between contacting residues, converted to a 0 to 1 confidence T with the same inverse-square form DockQ uses and a fixed 10 Å scale, then averaged over the contacts:[3][12]
Zhu and colleagues fitted it to AlphaFold-Multimer v2.2 models of 1,928 complexes with two to six chains. In their definition, the interface is everything one chain touches, and pDockQ2 for a chain predicts the DockQ of that chain against all the others combined.[3] The ipSAE script and AlphaFold DB compute it for each pair of chains and each direction instead, aligning on one chain and scoring the other, and AlphaFold DB displays the larger of the two directions.[12][6]
Unlike pDockQ, pDockQ2 spans almost the full range, from about 0.01 to 1.0. It reaches 0.23 at x of about 64, for example an interface pLDDT of 80 with interface PAE of about 5 Å, which gives a PAE term of 0.8. For hemoglobin, the interface PAE averages 1.1 Å, the PAE term is 0.987 and pDockQ2 is 0.94, close to the measured 0.967. For CDK2 with CKS1, the interface PAE averages 2.1 Å and pDockQ2 is 0.77, above the measured 0.645. The poorly placed activation loop barely enters the estimate: only one of its residues, His161, is among the model's 16 contacting CDK2 residues, so the loop's low pLDDT hardly moves the interface average.
pDockQ2 has failure modes of its own. In the AlphaFold-Multimer benchmark, 72 interfaces scored above 0.8 but had DockQ below 0.1. In most of them the individual chains were right and the docking was wrong, and in some the crystal reference showed a different interface from the one the model predicted, for example a dimer cut from a pantoate kinase crystal in which the model paired the subunits another way. In the other direction, 55 interfaces scored below 0.1 with DockQ above 0.6, mostly complexes where large parts of the modeled sequence were missing from the experimental structure and the prediction was uncertain overall.[3]
How pDockQ relates to ipTM and ipSAE
Of the interface scores that AlphaFold DB and ProteinIQ report, pDockQ and pDockQ2 are the two fitted directly to DockQ, so their values can be read against the CAPRI classes. ipTM, introduced with AlphaFold-Multimer, is a predicted TM-score over pairs of residues in different chains.[13] ipSAE, introduced by Roland Dunbrack in 2025, rebuilds that TM-score term from PAE values but counts only pairs below a PAE cutoff and scales by how many pairs pass.[14] The ipTM and pTM guide and the ipSAE guide cover both calculations.
| Score | Built from | Estimates | Common threshold |
|---|---|---|---|
| DockQ | Model and reference coordinates | Nothing; it is the measurement | 0.23, 0.49 and 0.80 for the CAPRI classes |
| pDockQ | Interface pLDDT and number of contacts | DockQ of a heterodimer | 0.23 |
| pDockQ2 | Interface pLDDT and PAE between contacting residues | DockQ of each interface in a complex | 0.23 |
| ipTM | Error distributions over all interchain residue pairs | TM-score of the interface | Above 0.8 confident, below 0.6 likely incorrect |
| ipSAE | Interchain PAE below a cutoff, scaled by the pairs that pass | A TM-score-like interface confidence | 0.6, 0.7 and 0.8 bands |
The scores also define the interface differently. pDockQ and pDockQ2 start from residues in contact in the predicted structure, so they only score the interface the model drew. ipTM uses every interchain pair, so disordered tails lower it even when the interface is good. ipSAE starts from pairs with low PAE, which AlphaFold DB further restricts to pairs within 15 Å.[6][14] In the worked example, all of them rank hemoglobin above CDK2 with CKS1, but only pDockQ2 lands in the same CAPRI class as the measured DockQ for both models, and ipTM is above 0.8 for both.
AlphaFold DB uses two of these scores together. It shows a predicted dimer on its entry pages only when pDockQ2 is at least 0.23 and ipSAE is at least 0.6, which left 2,158,419 homodimers and 79,156 heterodimers from the NVIDIA dataset.[6] The AlphaFold DB statistics put those numbers in context. Requiring two scores built from different inputs is a reasonable habit for your own predictions as well.
Which score to use
When an experimental structure of the same complex exists, measure DockQ. Report fnat, iRMSD and LRMSD alongside it, since the same DockQ can come from different mixtures of errors, and for complexes with more than two chains look at the lowest interface score as well as GlobalDockQ. This is the usual setting for benchmarking a predictor or a protein-protein docking method on known complexes.
Without a reference, read pDockQ2 and ipSAE together, and check pLDDT and PAE at the interface itself. pDockQ is still useful for comparing with studies that used it, but its 0.742 ceiling and its blindness to PAE make it a weak choice for new work. None of the three predicted scores is defined for protein-ligand complexes.
Calculating DockQ and pDockQ2 on ProteinIQ
The DockQ tool takes a model and a reference in PDB or mmCIF format and runs DockQ 2.1.3. It maps chains automatically or with a fixed mapping, scores every interface in the reference, and reports DockQ, the CAPRI class, fnat, iRMSD, LRMSD, F1 and clashes for each one. Small molecule mode adds ligand LRMSD for ligands stored as separate chains.
The ipSAE tool calculates pDockQ, pDockQ2, ipSAE, ipTM and LIS from the PAE file and structure of an AlphaFold2, AlphaFold3 or Boltz prediction, using the same reference script as the worked example.
A typical workflow for a predicted complex:
- Predict the complex with AlphaFold2 or Boltz-2, with PAE output enabled. Boltz-2 only writes the PAE matrix when Save PAE matrix is turned on.
- Score the prediction with ipSAE and read pDockQ2 and ipSAE for each chain pair.
- If an experimental structure of the same complex exists, upload the model and the structure to DockQ to measure accuracy directly.
- For an interface that scores well, look at which residues make contacts before using them for mutation analysis or follow-up docking.
Tools for interface accuracy on ProteinIQ
- DockQ measures interface accuracy against a reference for protein, nucleic acid and small-molecule interfaces.
- ipSAE calculates pDockQ, pDockQ2, ipSAE, ipTM and LIS for predicted complexes.
- AlphaFold2 predicts multimers with pLDDT, PAE and ipTM.
- Boltz-2 predicts complexes with proteins, nucleic acids and ligands, with an optional PAE matrix.
- RMSD calculator compares whole structures when interface accuracy is not the question.
For choosing a predictor in the first place, see protein complex structure prediction.
Frequently asked questions
What is a good DockQ score?
A DockQ of 0.23 or more is an acceptable model by CAPRI standards, 0.49 or more is medium quality and 0.80 or more is high quality.[1] Most structure prediction benchmarks count a model as a success at 0.23.[2]
What is the difference between DockQ and pDockQ?
DockQ is measured by comparing a model with a reference structure. pDockQ is an estimate of DockQ made from the model's own confidence scores, the interface pLDDT and the number of contacts, so it needs no reference but is only as good as its calibration.[1][2]
What does a pDockQ2 of 0.23 mean?
It is the predicted equivalent of the CAPRI acceptable line. AlphaFold DB treats pDockQ2 above about 0.23 as a practical guide to a likely acceptable interface and uses it, with ipSAE of at least 0.6, to choose which predicted dimers to display.[6]
What is DockQ v2?
The 2024 reimplementation of DockQ. It keeps the original scores and adds multi-interface scoring with automatic chain mapping, nucleic acid interfaces, small-molecule LRMSD and clash counts.[4]
Can DockQ score protein-ligand docking?
DockQ v2 reports a symmetry-corrected, pocket-aligned LRMSD for small molecules but no DockQ quality class. The ligand must be a separate chain in both the model and the reference.[4][5]
Why is my DockQ low when the model looks right?
Check the chain mapping first, especially when the reference has several copies of a chain. Then look at iRMSD. A few misplaced interface residues can push it above 2 Å, which removes most of the iRMSD third of the score even when nearly all contacts are correct, as in the CDK2 example above.
