GraphCPLMQA

A structure predictor grades its own homework. AlphaFold 2's pLDDT is computed by the same network that produced the coordinates, so wherever the model is confidently wrong, the confidence tends to be wrong in the same direction. An estimator that sees only the finished structure — the sequence it claims to fold, the geometry it actually has — can disagree, and that disagreement is what tells an experimentalist which of five candidate models to take into the lab and which loop in the winner not to believe.

GraphCPLMQA, from Guijun Zhang's group at Zhejiang University of Technology with collaborators at BioMap and MBZUAI, makes that judgment one residue at a time. Its starting observation is that most quality-assessment networks hand-build the features they reason over — one-hot sequence encodings, inter-residue distance matrices — and in doing so discard the sequence-to-structure relationship that pretrained protein language models already carry. GraphCPLMQA reads the submitted structure through ESM instead: residue embeddings from ESM-2 in the single-sequence version, or from the MSA transformer ESM-MSA-1b in the full version, paired with structure embeddings that the inverse-folding model ESM-IF1 derives directly from the backbone coordinates.

Those representations feed a coupled network rather than one monolithic predictor: a graph encoding module that relates sequence to structure, and a convolutional decoding module that relates structure to quality. Splitting the problem lets each half specialize, and the paper's ablations credit that split for much of the gain over the group's earlier convolutional estimators.

#Key Features

• Embeddings in place of hand-built encodings: sequence and geometry both enter as pretrained representations — 1280-dimensional ESM-2 layer-33 embeddings, or 768-dimensional ESM-MSA-1b embeddings plus 144-dimensional row attention, alongside a 512-dimensional ESM-IF1 structure embedding.
• A single-sequence mode that unlocks complexes: the Single version needs no MSA at all, which matters for multimer interfaces where paired alignments are unreliable or absent.
• Topology-aware geometric features: for each residue, two farthest-point references form a triangle and a local coordinate frame encoding how a fragment is oriented within the whole fold, while contact order — normally a global measure of topological complexity — is extended per residue.
• Error and threshold branches: the decoder predicts a distance-error map and a distance-threshold map separately before combining them into the local lDDT score, so the estimate carries an explicit notion of how far off each contact is.

#Technical Details

The encoder updates node and edge features with a graph transformer, AlphaFold-style triangular multiplication and invariant point attention, then an E(n)-equivariant graph network over each residue's K nearest neighbours; the decoder is a residual 2D convolutional stack using dilated convolutions, GELU activations and inverted bottlenecks. Training data came from 15,054 PDB proteins selected as of 19 November 2021 at resolution 2.5 Å or better, 50–400 residues, and under 35% mutual sequence identity. Decoys were generated three ways — dihedral perturbation with fast relaxation, template modelling with RosettaCM and I-TASSER-MTD, and deep-learning-guided conformational change with the group's RocketX — yielding 1,378,676 structure models. Training ran 100 epochs of AdamW at a 0.001 learning rate, roughly 120 hours on one A100.

On the CASP monomer sets (9,390 models across 70 CASP13 targets and 9,645 across 69 CASP14 targets, 19,035 in total), global quality reached Pearson 0.927. On 9,108 models from 34 CASP15 multimer targets, the single-sequence version scored interface residues at Spearman 0.617, Kendall 0.45, AUC 0.844 and MAE 0.144, a 23.6% Pearson improvement over ModFOLDdockR. In the CAMEO blind test from 20 May to 13 August 2022, covering 128,018 residues, local accuracy reached Pearson 0.891, Kendall 0.680, AUC 0.942 and MAE 0.081.

#Applications

The practical use is triage. Given the five models AlphaFold 2 returns for a target, GraphCPLMQA picked a better structure than the predictor's own top-ranked output on average — lower selection error by 0.672 Å RMSD, 0.015 TM-score and 0.008 lDDT — and beat pLDDT on mean absolute error for 253 of 345 CASP14 models. The per-residue output also tells a modeller where to spend effort, marking the regions of an otherwise plausible fold that need rebuilding. The method applies to arbitrary PDB inputs with no fitting on the user's data, and the group runs a public web server for both versions — in practice the only working route, since the checkpoint the repository points to is no longer downloadable.

#Impact

GraphCPLMQA became infrastructure inside its own lab's pipeline: DeepAssembly uses it to pick the final conformation from its refined multi-domain candidates, and its successor networks appear as the GraphCPLMQA2 scorers inside DeepUMQA-X. It also ran as a standing CAMEO-QE server. The honest limits: accuracy on multimer interfaces degrades for both very short and very long targets, the MSA version pays for its extra accuracy with an HHblits search against UniRef30 and BFD, and the released code carries no license file. The checkpoint's only publication route was a single unversioned .pkl on the lab's own web host, and that URL is now dead, leaving the hosted server as the only way to run the method.