Graph neural networks
Can an equivariant graph network judge protein models better?
Open access · cc by · source: Europe PMC
A graph network built to respect 3D rotations and translations estimated the quality of predicted protein structures more accurately than earlier methods, and even better than AlphaFold2's own confidence scores.
Study at a glance
- Design
- Computational / modelling — New equivariant graph neural network (EnQA) trained on CASP and AlphaFold2 structural models and compared with existing quality-assessment methods on held-out test sets
- N
- No single N: separate test sets of 9501 CASP14 models (64 targets), 945 CAMEO models (38 targets) and 56 AlphaFold2 test targets
- Population
- Predicted 3D protein structural models from CASP, CAMEO and AlphaFold2
- Outcome
- Agreement (MSE, MAE, Pearson correlation) between predicted and true lDDT quality scores at residue and model level
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
On AlphaFold2 models the best variant reached a per-model correlation of 0.9001 with true quality, versus 0.8376 for AlphaFold2's own score, 0.4966 for DeepAccNet and 0.34 for VoroMQA. AlphaFold2's self-score systematically overestimated quality, while EnQA's predictions did not differ significantly from the truth. Removing AlphaFold2's confidence score as an input caused the largest drop in performance, but the model still beat older methods; one hand-crafted feature (buriedness) barely mattered.
Methodology
The authors represented each predicted protein structure as a graph of residues and built EnQA, a network whose layers are equivariant to rotating or moving the molecule, so its quality estimate cannot depend on arbitrary orientation. It combined geometric features, protein language-model attention maps and information from AlphaFold2 predictions. They trained on older CASP rounds and AlphaFold2 models, then tested on CASP14, CAMEO and a family-disjoint AlphaFold2 test set, comparing with DeepAccNet, VoroMQA, ProQ4 and AlphaFold2's self-reported score.
Limitations
The AlphaFold2 test set is small (56 targets) and the CAMEO set covers only 38 targets over a few months, so rankings could shift on other data. On CASP14 the method was slightly worse than a simple AlphaFold2-consensus baseline on per-residue MAE, so it is not uniformly better. The authors note that AlphaFold2's apparent overestimation may partly reflect errors in the experimental reference structures. The paper does not isolate how much of the gain comes from equivariance itself versus the extra AlphaFold2 and language-model features.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
Building 3D symmetry into the network helped on protein structure tasks.
E(3)-equivariant GNNs, which respect 3D rotations and translations, outperformed alternatives: EquiPPIS was the only method above 0.8 ROC-AUC for interaction sites, and its non-equivariant version and plain GCN/GAT networks did worse.
Evidence for the claim as stated.
Building 3D symmetry into the network helped on protein structure tasks.
E(3)-equivariant GNNs, which respect 3D rotations and translations, outperformed alternatives: EquiPPIS was the only method above 0.8 ROC-AUC for interaction sites, and its non-equivariant version and plain GCN/GAT networks did worse.
Scope note — Did not isolate equivariance from its extra AlphaFold2 and language-model features; removing AlphaFold2's confidence input caused the largest drop.
Limits the claim's scope: a different population, assay, or outcome.
Estimating model quality is its own learning problem.
Deep models for model-quality assessment improved ranking modestly: DeepQA tied the leading single-model method on CASP11 stage one and led on stage two, and a later equivariant GNN beat AlphaFold2's own confidence score (0.90 vs 0.84 correlation) on AlphaFold2 models.
Evidence for the claim as stated.
Related papers in this topic
Same topic cluster — not a recommendation engine.
- Can spreading signals through a protein network find disease genes?
- Can a known gene network make microarray classifiers interpretable?
- Can a graph neural network sort unknown phage DNA into families?
- Can self-supervised learning predict how mutations change binding?
- Does letting each node choose its own depth fix GNN over-smoothing?