GraphGPSM

A two-domain protein model can put every residue in a locally correct environment and still be wrong in the way that matters — the two domains hinged apart at the wrong angle. An accuracy estimator that predicts lDDT residue by residue and then averages sees nothing amiss, because each domain scores well on its own, while the model's TM-score against the native structure collapses. CASP15 made the gap concrete for this lab: its interface estimator DeepUMQA3 led the interface-residue category while doing poorly at estimating global folding accuracy.

GraphGPSM answers that by training a network to output one number — the TM-score a submitted structure would achieve against its unknown native — rather than a per-residue profile summarised after the fact. Developed by Guangxing He, Jun Liu, Dong Liu and Guijun Zhang at Zhejiang University of Technology and published in Briefings in Bioinformatics in 2023, it represents a structure as a residue graph and reasons over it with an E(n)-equivariant graph neural network, so rotating or translating the input leaves the score unchanged.

Every node and edge descriptor comes from the submitted coordinates or from Rosetta's energy model — no multiple sequence alignment, no template search — which is what lets the method judge orphan proteins, where alignment depth is the thing that is missing.

#Key Features

• Fold accuracy as the training target: the network regresses TM-score directly, reporting on a model's overall topology rather than on the correctness of each residue's neighbourhood.
• Residue-level ultrafast shape recognition: the two successively farthest residues are located and the first moments of all distances to that triple are taken, compactly describing where a residue sits inside the whole fold.
• Gaussian radial basis encoding of the backbone: each residue's 30 nearest Cα neighbours contribute a unit direction vector and a distance, both expanded through Gaussian radial basis functions — the largest single contributor in the ablations.
• Physics alongside geometry: Rosetta one-body terms (rama, p_aa_pp, omega, ref) sit on the nodes and two-body terms (fa_atr, fa_rep, fa_sol, fa_elec, hydrogen bonding) on the edges, next to trRosetta-style inter-residue distances and orientations.
• A search objective, not only a filter: cheap enough to evaluate on every proposed conformation, the score can enter a folding protocol's acceptance criterion rather than being applied once at the end.

#Technical Details

The EGNN runs 50 graph propagation steps and feeds a multilayer perceptron of dropout, activation and linear layers that emits the global score. Training data was 15,054 PDB chains culled through PISCES on 19 March 2022 at under 35% mutual identity, 50–400 residues and 2.5 Å or better, plus 1,118 human AlphaFold DB models with pLDDT above 90 — 16,172 proteins in all. About 150 decoys per protein came from dihedral perturbation with fast relaxation, template modelling with RosettaCM and Modeller, and deep-learning modelling with the group's in-house RocketX, clustered and filtered to roughly 1,473,706 structures. Training used Adam at a 0.001 learning rate, batch size 8 and mean squared error, about a week on one Tesla V100.

Correlation between the predicted score and true TM-score reached Pearson 0.812 on CASP13 (12,150 models across 81 targets), 0.719 on CASP14 (10,281 across 69) and 0.809 on CAMEO (1,810 models across 181 targets, May to July 2022), ahead of ModFOLD7/8, ProQ3D, VoroMQA-A, DeepAccNet, QDeep, QMEAN and REF2015. In the CASP15 blind season, entered as the server GuijunLab-Threader, its average predicted TM-score of 0.730 was the closest among top-ranked servers to the true average of 0.716, and its average bias of 0.126 the lowest.

#Applications

The straightforward use is ranking. Across 66 CASP14 targets, models picked by GraphGPSM from AlphaFold 2's five outputs averaged TM-score 0.875 against 0.859 for the predictor's own first-ranked model. The more interesting use is guidance: added alongside REF2015 inside Rosetta's ClassicAbinitio fragment assembly, with each proposed conformation accepted or rejected under a Metropolis criterion, it raised mean TM-score by 66.5% over REF2015 alone on 484 test proteins. Re-predicting from AlphaFold 2 starting models, 35 orphan proteins improved from 0.478 to 0.541 mean TM-score and 57 multi-domain proteins from 0.611 to 0.656 — the latter from corrected interdomain orientation rather than better individual domains.

#Impact

GraphGPSM staked out global topology as a scoring target distinct from local lDDT, and made the case that a scoring model can improve a structure prediction rather than merely grade it. That thread continued in DeepUMQA-X, where fold-level, interface and per-residue accuracy get separate heads. The limitations are real: inference needs a Rosetta installation for the energy features, no code or weights were ever released, and the availability statement offered only a hosted server — since discontinued in favour of the lab's per-residue estimator GraphCPLMQA, leaving no route to run the method. The authors also acknowledged that scoring complex interfaces remained its weaker side.