DeepUMQA-PA

Almost every accuracy estimator for a predicted protein complex decides which residues are "in contact" by measuring a distance — conventionally 8 Å between Cβ atoms. That test treats two residues pressed face-to-face across a wide buried surface exactly like two that graze at a single point, and records nothing about the water that may or may not sit between them. Burying hydrophobic surface away from solvent is what drives folding and binding in the first place, so a scorer working from distances alone is blind to the physics that makes an interface real.

DeepUMQA-PA replaces the threshold with contact area. Voronoi tessellation over the complex's heavy atoms, each given its van der Waals radius, produces an explicit dividing surface between every touching atom pair; summing those areas over residues gives how much surface a residue buries against its neighbours and how much it still presents to solvent. A second quantity falls out of the same tessellation — the orientation of each contact surface, a vector where the surface crosses the interatomic axis. Together they say not merely that two residues touch, but how much and in which direction. The authors call these physical-aware features, and their ablations show the solvent-contact term is the network's most valuable input.

It comes from Haodong Wang, Meng Sun, Lei Xie, Dong Liu and Guijun Zhang at Zhejiang University of Technology, published in Computational and Structural Biotechnology Journal after a 2024 bioRxiv preprint. It extends DeepUMQA3, the group's interface estimator, and joins a line running from the monomer scorers DeepUMQA and DeepUMQA2 through DeepUMQA-Global to the unified DeepUMQA-X.

#Key Features

• Contact area rather than a distance cutoff: residue-level areas accumulate from Voronoi atom-atom contacts with intra-residue pairs masked, and a separate residue-solvent area captures exposure to the surrounding medium.
• Contact-surface orientation: per-residue vectors summed from the atom-level surfaces encode the geometry of how a residue meets its neighbours, which distance features discard entirely.
• Single-model operation: one PDB file is enough, so it works on a lone AlphaFold output as readily as on a CASP-scale decoy set.
• Equivariant coordinate refinement: an E(n)-equivariant graph neural network coupled to invariant point attention nudges the input coordinates toward the native structure, and that structure embedding rejoins the feature stack before the accuracy readout.
• Pretrained language model embeddings: sequence embeddings from ESM-2 (L×1280) and backbone structure embeddings from ESM-IF1 (L×512) supply the evolutionary signal.

#Technical Details

Four feature classes enter the network: physical-aware contact area and orientation, ultrafast shape recognition and voxelization features inherited from the DeepUMQA line, the two language model embeddings, and Rosetta energies. Fully connected layers project them to an L×128 embedding, which joins the model's coordinates in the EGNN–IPA block; the resulting structure embedding is recombined with it and fed to an attention-equipped residual network of dilated 2D convolutions. Branch blocks emit a 15 Å contact mask and an estogram of Cβ distance deviations, from which per-residue lDDT is computed. Training used DeepUMQA3's non-redundant complex set restricted to structures before January 2022 — 7,590 targets at roughly 240 models each, split 9:1 — for 125 hours on a single A100. On 7,875 CASP15 models across 30 targets, all generated after May 2022 and so genuinely held out, interface residue lDDT prediction reached Pearson 0.608, Spearman 0.574 and AUC 0.765 against DeepUMQA3's 0.587, 0.554 and 0.761, with the next-best method at 0.476 Pearson. On the five nanobody-antigen targets the gap widens to 0.576 against 0.493. Removing the residue-solvent contact area drops Pearson from 0.61 to 0.52.

#Applications

The estimate lands where a structural biologist has a predicted assembly and must decide how far to trust it. Per-residue output makes the answer spatial rather than a single number: it marks which stretches of an interface are firm enough to site a mutation, design a binder against, or feed to a refinement protocol. The authors position it as complementary to the self-assessment scores structure predictors emit — it achieved lower mean absolute error than AlphaFold-Multimer on 43% of targets and AlphaFold 3 on 50%, concentrated on the targets those models scored worst, which is exactly where an independent second opinion matters.

#Impact

DeepUMQA-PA demonstrates that interface accuracy estimation was leaving physical information on the table: swapping distance thresholds for tessellation-derived areas and orientations improved a method that already led its CASP15 category, most of all for nanobody-antigen complexes, the weak-signal case complex prediction handles worst. The paper is candid that interface residues remain hardest to score, with mean absolute error roughly 29% worse than for buried core residues. Distribution is the real limitation: no code, weights or package was released, the availability statement offers only the web server, and that endpoint now shows a maintenance notice directing users to DeepUMQA-X — the in-place supersession this group applied to earlier servers, leaving no way to run this specific model.