Loading…
Protein model accuracy estimation for single chains and complexes, predicting per-residue lDDT, interface QS-score, and overall fold TM-score.
No providers recorded yet. Browse all providers
A modern complex-prediction pipeline does not produce one answer; it produces a pile of them. Sampling hundreds or thousands of candidate assemblies is now the main lever on accuracy, which moves the hard problem downstream: given a stack of plausible structures and no experimental answer, which one is right, and which parts of it can be trusted? That question — estimation of model accuracy, or EMA — has its own track at CASP, and it has historically been answered two incompatible ways. Consensus methods score a structure by how much it resembles its neighbours in the pool, which works well when the pool is good and fails quietly when it is not; single-model methods judge one structure on its own evidence, which is harder and more useful. And no single number settles it: a complex can have both chains folded correctly and the interface between them badly wrong.
DeepUMQA-X, from Guijun Zhang's group at Zhejiang University of Technology, is built around both of those observations. It runs a separately trained network for each accuracy metric rather than one estimator with three output heads bolted on — GraphCPLMQA2L for per-residue lDDT, GraphCPLMQA2Q for interface QS-score, and GraphCPLMQA2S together with DeepUMQAS for overall fold TM-score. Those single-model scores can then be used to seed an optional consensus pass, which ranks the submitted pool, keeps the structures placing in the top 25% on both fold and local accuracy, and aligns those against everything else the user uploaded. That inverts the usual dependency: consensus is anchored by per-structure predictions instead of substituting for them.
Published in the 2025 Nucleic Acids Research Web Server issue, it continues a line of EMA networks from the same lab, GraphCPLMQA2 adding protein-language-model and pair-representation features to an earlier convolutional design.
GraphCPLMQA2 refines its evolutionary features through a graph transformer, then feeds them with the model's rotations and translations into IPA modules that generate geometric constraints approximating the native structure. Those, with the non-evolutionary features, pass through a transformer-strategy ResNet predicting a distance-deviation map and a contact map; per-metric fusion blocks turn those maps into lDDT, TM-score, and QS-score. Training data predates January 2022, which makes the CASP15 benchmark — generated after May 2022 — a genuinely held-out test; on it, local interface lDDT prediction reached Pearson 0.809, Spearman 0.763 and AUC 0.847. In the CAMEO-QE one-year blind test, GraphCPLMQA2L ranked first over the full dataset with a partial ROC AUC of 0.75 against 0.72 for the runner-up, and 0.84 on the low-quality model subset. In CASP16, it led all single-model methods on local interface accuracy with a Z-score of 142.48.
The direct users are groups running large-scale sampling for protein complexes — massive AlphaFold 2 sampling, docking ensembles, assembly pipelines — who need to rank a pool and commit to one structure before doing anything expensive with it. The free web server takes PDB files or a zip archive, distinguishes single-chain from complex assessment, and typically returns single-model results within an hour. Per-residue and per-interface scores matter after selection too: they tell a wet-lab collaborator which contacts in a predicted assembly are worth designing an experiment around and which are modelling noise.
DeepUMQA-X placed among the top CASP16 EMA methods, including second in per-residue self-assessment and third in complex local interface accuracy, and its lightweight alignment made previously impractical targets tractable — roughly 300 candidate models of the 5689-residue target H1227 were aligned in about an hour where conventional tools needed days. The tradeoff is cost: consensus assessment averaged 9.21 hours per target against 1.57 hours for a docking-score baseline, though it selected better structures. The sharper limitation is distribution — DeepUMQA-X ships as a hosted server, with no downloadable weights or standalone package for the trained scorers, so it cannot be embedded directly in a local pipeline.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.