Graph neural networks
Does building 3D symmetry into a GNN improve protein site prediction?
Open access · cc by · source: Europe PMC
A graph neural network whose outputs transform correctly under rotations, reflections and translations predicted protein interaction sites better than earlier methods, even from AlphaFold2-predicted structures.
Study at a glance
- Design
- Computational / modelling — EquiPPIS (E(3)-equivariant GNN) trained on 335 proteins, tested on 60, compared with nine prior predictors plus invariant/no-attention/GCN/GAT ablations; also tested on unbound and AlphaFold2 structures
- N
- N=60 · 60 test proteins (Test_60); 335 training proteins; 31-protein unbound test subset; 42-protein validation set
- Population
- Protein chains from three public PPI benchmark datasets
- Outcome
- Per-residue interaction-site classification: ROC-AUC, PR-AUC, F1, MCC
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
EquiPPIS was the only method to exceed a ROC-AUC of 0.8 on the test set, beating all sequence- and structure-based competitors. The invariant version and plain GCN/GAT networks did worse, pointing to equivariance as the main source of the gain, with attention adding a smaller boost. Using AlphaFold2-predicted structures, EquiPPIS reached a PR-AUC of 0.451, higher than the best competitor achieved even with experimental structures (0.429), and it lost less accuracy on unbound structures.
Methodology
The authors turned each protein into a graph of residues linked when their alpha-carbons lie within 14 angstroms, and fed coordinates plus sequence and structure features into an E(3)-equivariant graph neural network (EquiPPIS) that labels each residue as an interaction site or not. They trained on 335 proteins and tested on 60, against nine existing predictors. To test why it worked they trained versions with the coordinate updates switched off (making it merely invariant), with attention removed, and ordinary GCN and GAT networks, all with the same features.
Limitations
The test set is small (60 proteins, 31 in the unbound subset), so the gaps between methods are estimated from few examples, although the authors ran resampling significance tests. The method predicts sites for a single protein without knowing its partner, so it cannot say which partner binds where. The authors note the network is not interpretable, and they did not test adding multiple sequence alignment information.
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.
Gains over strong baselines were often small (about one to four points for iGRLDTI; small on KIBA for iNGNN-DTI) and came from benchmarks with randomly sampled negatives, single datasets or unequal evaluation protocols, while the protein-structure GNNs showed larger margins.
Same question, contrary or null result.
Open questions
Tensions this paper is part of
From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.
Gains over strong baselines were often small (about one to four points for iGRLDTI; small on KIBA for iNGNN-DTI) and came from benchmarks with randomly sampled negatives, single datasets or unequal evaluation protocols, while the protein-structure GNNs showed larger margins.
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?