DeepUMQA

A structure predictor hands back a coordinate file and no error bars. Estimation of model accuracy supplies the missing half — which parts of a predicted fold can be trusted — and the hardest version of it is the single-model case: judge one structure on its own evidence, with no pool of alternatives to average over. Doing that well means describing every residue twice: what its immediate atomic neighbourhood looks like, and where it sits in the fold as a whole. Voxelization, the prevailing feature when this work appeared, covers only the first. It grids the atoms around a residue in a local frame set by its backbone, which is informative about packing but cannot tell a buried core residue from a surface one on a different lobe.

DeepUMQA supplies the missing global term by adapting Ultrafast Shape Recognition, a descriptor built for matching small-molecule shapes, to the residue level. For a given residue, three reference positions are chosen in sequence — the residue itself, the residue farthest from it, and the residue farthest from that one — and for each the mean distance from every residue in the chain is taken. Those three first moments say where the residue lies relative to the entire structure, at almost no computational cost. Ablations show the two feature types are largely complementary: removing either degrades nearly every correlation metric, and using both is best overall.

It was built by Sai-Sai Guo, Jun Liu, Xiao-Gen Zhou and Gui-Jun Zhang at the College of Information Engineering, Zhejiang University of Technology, and published in Bioinformatics in 2022, founding an accuracy-estimation line the same group later extended to complexes and interface scoring, ending with DeepUMQA-X.

#Key Features

• Residue-level shape recognition: the USR first moments give each residue a rotation- and translation-invariant summary of its place in the global topology, without a second 3D convolution stack.
• lDDT derived, not regressed directly: two output arms predict a Cβ distance-deviation distribution and a contact map thresholded at 15 Å, and per-residue lDDT is computed from those probabilities.
• Single-model operation: scoring needs one PDB file, so it works where consensus methods, which rank by similarity to neighbours in a pool, have nothing to score against.
• Physics terms alongside geometry: 13 Rosetta centroid energy terms, covering one-body, two-body and hydrogen-bonding contributions, enter as normalised inputs beside the geometric features.
• Optional four-model ensemble: inference runs from a single downloaded checkpoint, or averages four.

#Technical Details

Voxelized residue environments pass through 3D convolutions, are flattened and concatenated with 1D features (BLOSUM62 vectors, Meiler amino-acid properties, DSSP secondary structure), then striped into a 2D tensor and joined with the 2D maps — inter-residue distances, orientations, sequence separation and pairwise energies. A convolution with instance normalisation and ELU activation lifts these to 128 channels for a residual trunk of 15 dilated blocks, branching into two arms of four blocks apiece. Training used a PISCES-derived non-redundant set — 40% maximum sequence redundancy, 2.5 Å resolution cutoff, chains of 50–300 residues — of 7,615 proteins split 7,226 train / 389 validation, with roughly 150 decoys per protein from Rosetta stochastic modelling, RosettaCM, native perturbation and trRosetta-guided folding. Across 51 CASP13, 44 CASP14 and 195 CAMEO targets — 17,057 models in total — global Pearson correlation with true lDDT reached 0.8332, 0.7891 and 0.7522 respectively, against 0.8157, 0.7778 and 0.7303 for DeepAccNet, the strongest baseline; ROC AUC for separating good from poor models was 0.938 on CASP13 and 0.902 on CASP14.

#Applications

The practical use is triage: given a set of predicted models for one target, the per-residue lDDT profile identifies the best structure to carry forward and marks the loops and termini modelled badly — what a refinement protocol needs to know where to work. Because scoring is single-model, it applies equally to a lone prediction and to a large decoy set, and the per-residue output doubles as a confidence track when deciding which regions justify a mutagenesis or binding experiment. The group entered it in the CAMEO-QE blind assessment, where it placed first over a one-month window spanning 26 November to 18 December 2021.

#Impact

DeepUMQA arrived as accuracy estimation was catching up with AlphaFold 2: predictions got dramatically better, so a quality estimator's value shifted from ranking mediocre decoys toward flagging the residues where an otherwise good model is wrong. Its cheap global-topology descriptor carried into the group's later systems and into the wider practice of pairing local and global features. Distribution is the weak point: the repository named in the paper is no longer reachable, the surviving community mirror carries no licence file and still contains the authors' hardcoded absolute paths, and the weights are served from a lab download page rather than a stable archive. The published server URL now hosts the successor system.