DeepUMQA3

A predicted protein complex can be right and wrong at the same time. Each chain folds to something close to its experimental structure, every per-residue confidence score looks healthy, and the two subunits are still docked against the wrong face. Accuracy estimators built for monomers cannot see that failure, because nothing in a single chain's local environment records how it meets its partner — and the interface is the part a biologist wants to act on, since it is where mutations and binders go. CASP15 made that gap explicit when it shifted its estimation-of-model-accuracy category from monomers to complexes.

DeepUMQA3 answers it by describing a submitted assembly at three levels at once. At the level of the overall complex, the structure is treated as one object and order-independent features are extracted: an overall ultrafast shape recognition descriptor, residue voxelization, inter-residue distances and orientations, and amino acid property vectors. At the intra-monomer level each chain is described on its own through an ESM-1b sequence embedding, DSSP secondary structure and Rosetta energy terms. The inter-monomer level carries the genuinely new part: an attention map over the paired inter-monomer sequence, and an inter-monomer USR that places each residue of one chain relative to the topology of the other chains. The original USR descriptor, introduced by the same group for monomers, summarised where a residue sits inside its own fold in a handful of numbers; extending it across the chain boundary buys an alignment-free account of the assembly's global arrangement, not just its local contacts.

The work comes from Jun Liu, Dong Liu and Gui-Jun Zhang at the College of Information Engineering, Zhejiang University of Technology, and was published in Bioinformatics in 2023. It is the complex-scoring member of a line that begins with DeepUMQA and DeepUMQA2, both monomer estimators, and continues in DeepUMQA-X.

#Key Features

• Inter-monomer ultrafast shape recognition: a residue-level shape descriptor computed across the chain boundary, so a residue is characterised by its position relative to the other monomers' topology and not only its own.
• Three-level feature ensemble: complex-wide geometry, per-monomer sequence and energetics, and cross-monomer sequence attention are combined in a single input tensor rather than scored separately and merged afterwards.
• Interface and whole-structure output: the network predicts interface residue accuracy and per-residue lDDT for every residue in the assembly, so high- and low-precision regions away from the interface are also separated.
• Single-model scoring: one PDB file is enough; there is no dependence on a pool of alternative models, which is what consensus estimators need and often do not have.
• Batch web service: structures can be pasted, uploaded individually or submitted as a zip archive, with results returned as a 3D view, a per-residue lDDT plot and per-chain interface accuracy.

#Technical Details

The three-level features are consumed by an improved deep residual network coupled with triangular update and axial attention, which predicts per-residue lDDT and interface residue accuracy directly. In the CASP15 blind test of interface residue accuracy estimation, entered as GuijunLab-RocketX, it reached Pearson 0.564, Spearman 0.535 and AUC 0.755 against true interface lDDT, and 0.505, 0.456 and 0.714 under the CAD measure — the strongest correlations in that category, with the runner-up at 0.476 Pearson on lDDT. The CASP15 companion analysis reports 0.570 Pearson under the assessor's own evaluation, the only entry above 0.5. Across the 39 assessed targets it had the highest lDDT Pearson among the top five methods on 25, led on both homomers and heteromers, and was best on four of five nanobody–antigen and all three antibody–antigen targets. Server runtimes were measured CPU-only on the 35 CASP15 complexes under 3000 residues.

#Applications

The natural place for it is immediately after complex prediction: a pipeline running AlphaFold-Multimer or a docking protocol produces many candidate assemblies, and DeepUMQA3 supplies the interface-level accuracy estimate that decides which one to carry forward and which contacts within it are real. Because the output is per-residue rather than a single global number, it also tells an experimentalist which parts of a predicted interface are firm enough to design a mutation or a binder around, and gives a refinement protocol a map of where to spend effort. The authors are explicit that the server is aimed at interface residues, and point elsewhere for global complex fold accuracy.

#Impact

DeepUMQA3 established that features spanning chain boundaries, rather than better local scoring, were what complex accuracy estimation needed, and its CASP15 margin made that case unambiguously. The same three-level design carried into the group's later estimators. Distribution is the limitation: no weights, inference code or package was released, the paper's availability statement offers the web server alone, and that address has since been taken over in place by DeepUMQA-X, leaving no standalone DeepUMQA3 endpoint to call.