DeepUMQA2

A single-model accuracy estimator is handed one predicted structure and asked how good it is, with no experimental answer and no pool of rivals to average over. The usual way to answer is to interrogate the coordinates: pack the local atomic neighbourhood of each residue into a voxel grid, add energy terms, add a description of where the residue sits in the fold. All of that describes the submitted model, and none of it describes the protein. Two structures can be equally self-consistent and equally well packed while only one of them agrees with what the residue's own family has been doing for hundreds of millions of years.

DeepUMQA2, from Guijun Zhang's group at Zhejiang University of Technology, closes that gap by re-deriving evidence from outside the structure being scored. The sequence is pulled off the submitted model, run through HHblits against UniRef30 and BFD to build a multiple sequence alignment, and run through HHsearch against PDB100 to collect homologous templates. Co-evolution statistics from the alignment and geometric constraints from the top templates are concatenated with the model-dependent features inherited from DeepUMQA into an initial residue-pair representation — so the network can ask whether the contacts present in the model are the contacts the family implies, rather than only whether they are physically plausible.

The second change is the backbone. Where the first version stacked residual blocks, DeepUMQA2 updates the pair representation with triangular multiplication and axial attention, letting information flow between residue pairs that share an index rather than only between spatial neighbours. Ablations separate the two contributions: the new backbone alone beats the predecessor on every global metric, and the sequence and template features improve it further.

#Key Features

• Evidence from beyond the submitted model: one-hot sequence encoding, residue position entropy, a position-specific frequency matrix and an inverse covariance matrix with average product correction give the network the family's coupling pattern to check the structure against.
• Templates at three levels: the top 20 homologous templates contribute pairwise distances and orientations, per-residue alignment probability and secondary-structure match, and whole-alignment identity scores; when no adequate template is found, the model scores without them.
• Triangular and axial pair updates: two triangular multiplication layers, two axial attention layers and a feedforward layer form a pair-update module iterated three times.
• Structured output, not a bare score: two residual branch networks predict an inter-residue distance-deviation distribution and a 15 Å contact map, from which per-residue and global lDDT are computed in closed form.
• Fixed checkpoint, later targets: the network is trained once against databases frozen in May 2018 and applied unchanged to targets and sequence databases years newer.

#Technical Details

Training reuses DeepUMQA's corpus: 7,615 PISCES proteins at ≤40% sequence redundancy, roughly 150 decoys each from stochastic modelling, comparative modelling, deep learning and native-structure perturbation. Distance deviations are binned into 15 intervals at ±0.5, ±1, ±2, ±4, ±10, ±15 and ±20 Å; the loss combines distance-deviation cross-entropy, contact cross-entropy and per-residue lDDT mean-squared error with weights of 1, 1 and 10, trained for 100 epochs with AdamW on TITAN RTX GPUs. Each branch network holds six residual blocks of dilated 2D convolutions. On CASP13 (76 targets, 10,739 models) and CASP14 (70 targets, 10,380 models) the global Pearson correlation reaches 0.919 and 0.899 and top-1 loss falls to 0.049 and 0.035, improvements of 20.5% and 20.4% in top-1 loss over the first version. Over a three-month CAMEO blind test (11 March to 4 June 2022, 192 targets and 1,882 models) it reached a local AUC of 0.797 at a 0.2 false-positive-rate cutoff and ranked first among competing server methods.

#Applications

The immediate use is model selection: given the hundreds of candidate structures a modelling pipeline produces, rank them by predicted global lDDT and commit to one. The per-residue output then does a second job, marking which segments of the chosen structure are reliable enough to build an experiment or a docking run on and which are modelling noise worth refining. Because scoring requires a fresh MSA and template search, it fits at the end of a pipeline that has already paid that cost, not as a lightweight filter inside a sampling loop.

#Impact

DeepUMQA2 marks the point where this lab's accuracy estimators stopped treating a structure as self-contained evidence, a choice carried forward through DeepUMQA3 to complex interfaces and into GraphCPLMQA and DeepUMQA-X, which swap the alignment features for protein language model embeddings. Its practical reach is limited by distribution: it shipped as a hosted web server with no source package and no downloadable checkpoint, and the address published with the paper now serves the lab's later DeepUMQA-X system, so the trained network described here is no longer independently runnable. The database searches it depends on also make each assessment considerably heavier than a purely structure-based estimator.