Structural biology · interface quality

How do we judge a docked structure?

DockQ turns a deceptively visual judgment — “did these two proteins meet in the right way?” — into three complementary measurements and one interpretable score.

Step 1 · Start with the question

Same proteins is not the same interface

A docking prediction contains the right two molecular partners, but that is only the beginning. Imagine protein A as a rigid receptor and protein B as a ligand. B can be near A, yet rotated, shifted, or using a completely different patch. An evaluation must compare the relationship between them, not merely recognize the ingredients.

Formal version: what is compared?

For an experimentally solved reference complex and a predicted complex, the chains are matched by identity. Evaluation then asks whether corresponding interface residues contact one another and whether corresponding atoms occupy the same positions after a specified alignment.

Step 2 · Did it keep the handshakes?

Native contacts make the interface discrete

Call a residue pair a contact when the two residues lie close across the interface (a common convention uses a 5 Å heavy-atom cutoff). The reference structure supplies the list of native contacts. We then ask how many of those exact pairs the prediction recovered. This is \(F_\text{nat}\): the fraction of native contacts reproduced.

If the reference has \(N_\text{native}\) contacts and the prediction reproduces \(N_\text{shared}\), then \[F_\text{nat}=\frac{N_\text{shared}}{N_\text{native}}.\] It is a recall-like quantity: extra, wrong contacts do not repair missing native ones.

Formal version: the exact convention behind \(F_\text{nat}\)

DockQ follows the CAPRI-style idea of native contacts. Its implementation identifies interface residue pairs in the reference, then counts the fraction that are also in contact in the model. The precise atom/contact convention should be reported, because changing a cutoff changes the contact list.

Step 3 · Is the whole pose right?

Distance needs an alignment rule

RMSD is a root-mean-square distance between corresponding atoms. But before measuring it, we must decide what to line up. This choice is not a technicality: aligning on the interface asks whether its local geometry matches; aligning on the receptor asks whether the ligand landed in the right global pose.

iRMSD aligns the interface residues and measures their RMSD. It is unforgiving about the local fit where the proteins touch. LRMSD aligns the receptor, then measures the ligand's RMSD; it exposes a ligand that is rotated or translated relative to a correctly fixed receptor.

Formal version: RMSD in one line

For \(n\) corresponding atoms, after the chosen optimal superposition, \[ \mathrm{RMSD}=\sqrt{\frac{1}{n}\sum_{i=1}^{n}\left\|\mathbf{x}_i-\mathbf{y}_i\right\|^2}. \] Here \(\mathbf{x}_i\) and \(\mathbf{y}_i\) are predicted and reference coordinates. The alignment rule determines which rigid motion is removed before this average is taken.

Step 4 · Three clues, one score

DockQ rewards agreement without a cliff edge

A score based only on contacts would miss a near-correct pose; a score based only on distances could forgive the wrong handshake pattern. DockQ averages one contact term with two smoothly scaled distance terms. The scaling matters: a 1 Å error should count as strong evidence, while a 20 Å error should not become twenty times “more wrong.”

\[ \mathrm{DockQ}=\frac{1}{3}\left(F_\text{nat}+\frac{1}{1+(\mathrm{iRMSD}/1.5)^2}+\frac{1}{1+(\mathrm{LRMSD}/8.5)^2}\right). \] The constants \(1.5\) Å and \(8.5\) Å set the scales where the two distance contributions have fallen to one half.

Why these particular shapes?

Each distance term has the form \(1/(1+(d/d_0)^2)\). At \(d=0\), it contributes 1; at its scale \(d=d_0\), it contributes \(1/2\); and it approaches 0 smoothly as the mismatch becomes large. This avoids a brittle pass/fail cutoff while keeping the score bounded between 0 and 1.

Step 5 · Turn a score into a judgment

DockQ makes categories comparable

DockQ is continuous, so it preserves the difference between “barely useful” and “almost perfect.” For familiar CAPRI-style labels, common DockQ boundaries are: incorrect below 0.23, acceptable from 0.23, medium from 0.49, and high from 0.80.

These labels are a communication aid, not an excuse to ignore the components. Two structures can have similar DockQ values for different reasons: one may have the right contacts but a slightly shifted ligand, while another is geometrically close but missing much of the intended interface. Inspect \(F_\text{nat}\), iRMSD, and LRMSD alongside the final score.

Putting it together

One interface, three views

DockQ is not a magic number. It formalizes the plain test we began with: did the predicted proteins shake the same hands, at the same meeting place? \(F_\text{nat}\) checks the handshakes; iRMSD checks the local meeting surface; LRMSD checks the ligand's global landing. Their bounded average gives a fair, smooth way to compare docked complexes.

Where this shows up: benchmarking protein-protein docking methods, evaluating multimeric structure predictions, and comparing model-quality scores across targets whose interfaces have different sizes and geometries.