Loading…
Protein model accuracy estimation returning a complex fold score and an interface score from a single structure, with no candidate pool required.
No providers recorded yet. Browse all providers
A structure predictor's confidence score comes from the same network that produced the structure, so it inherits that network's blind spots. An independent judge has to work from the coordinates alone. The usual way around that is consensus — score a candidate by how closely it resembles the others in the pool — which reports agreement rather than accuracy, and misreads a protein with two genuine conformational states as an inconsistent pool. Single-model methods judge one structure on its own evidence, and have historically lagged behind both consensus scoring and built-in confidence.
DeepUMQA-Global answers it with a structure–sequence cross-consistency test run in both directions. In one, ProteinMPNN runs over the input coordinates to recover the sequence preferences that fold implies, and those are compared against the sequence actually supplied — a structure-aware global sequence, or SAGS. In the other, the sequence is folded and used to retrieve structural templates whose residue–residue distances are accumulated into a per-pair histogram with an entropy term, giving a sequence-conditioned expectation for the geometry — a sequence-aware structural profile, or SASP. A correct model satisfies both; a fold no sequence would choose, or a geometry no template supports, fails one. The network emits two scalars: pScore for the whole complex, and ipScore restricted to residues with an inter-chain atom within 5 Å.
The method comes from Guijun Zhang's IOBIO Lab at Zhejiang University of Technology and builds on the group's DeepUMQA3 protocol. It is a separate artifact from the same group's DeepUMQA-X server, which it treats as prior work and benchmarks against.
Residue-wise features pass through a self-attention graph pooling layer at a sampling ratio of 0.5 and four graph attention layers (eight heads, dropout 0.25, GELU) to form the single representation; inter-residue features go through a 2D CNN with a stride-2 downsampling block and a single-layer Pair2Pair row–column attention block (four heads, 128 channels) to form the pair representation, which the single representation gates elementwise before averaging over rows and columns. Training used 7,840 non-redundant monomeric PDB targets at 2.5 Å resolution and 50–400 residues, filtered at 40% sequence identity, with roughly 140 decoys apiece for 1,065,994 structures, supervised on TM-score under a log-cosh loss for about two days on a single A100. Against AlphaFold 3 self-assessment on 26 CASP16 complex targets, Pearson correlation with true TM-score rose from 0.287 to 0.453 and Spearman from 0.249 to 0.371; against HDOCK confidence on 107 PDB dimers, Pearson rose from 0.575 to 0.616 and Spearman from 0.251 to 0.418, with DiffDock-PP also compared. On the CASP16 retrospective benchmark of 30 targets and 10,521 models, it took the highest Z-score among single-model methods in both the fold and interface evaluations, 109.568 and 108.718.
The practical use is triage: given hundreds of candidates from sampling, docking or several predictors, rank them and commit to one before spending experimental effort. Being pool-independent, it also works as a common yardstick across pipelines reporting incomparable internal confidences. A free hosted server accepts a PDB upload and returns both scores. Running it locally is heavier: the feature pipeline calls out to PyRosetta, Foldseek-Multimer, Voronota and ProteinMPNN and needs the PDB100 and PDB_AFDB template databases on disk, hence the authors' Singularity container.
A single-model method reached the top of the CASP16 field on retrospective global accuracy — second overall in the fold evaluation against every participating method, consensus approaches included — without ever looking at a second candidate. That narrows a gap the CASP16 assessors called substantial, and makes external accuracy estimation practical where no model pool exists. The caveats are about distribution rather than the science: the checkpoint committed to the repository is a two-byte placeholder, the container carrying the real weights is linked as an unpublished deposit, no license accompanies the code, and the documented local workflow expects a reference structure from a hosted prediction service. The paper also notes that its absolute Pearson correlations stay modest because it slightly overestimates accuracy — it ranks better than it calibrates.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.