Graph neural networks
Can a graph neural network judge predicted protein shapes?
Open access · cc by · source: Europe PMC
Treating a protein as a graph of residues and contacts let a fairly simple neural network rank predicted protein structures about as well as the best existing single-model tools.
Study at a glance
- Design
- Computational / modelling — Message-passing GCN trained on scored CASP 9-12 decoys, tested on CASP13 stage-2 decoys, with ablations over co-optimisation, depth/connectivity and features.
- N
- About 85k training/validation decoys from CASP 9-12; test set of roughly 14k decoys across 72 CASP13 targets; ablations averaged over 10 runs on CASP 11.
- Population
- Computationally generated protein structure models (decoys) from CASP editions
- Outcome
- Agreement between predicted and true quality scores (per-target Pearson R for global GDT_TS, Spearman correlation for per-residue LDDT/CAD, RMSE)
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
On CASP13, GraphQA ranked decoys within a target better than other single-model methods on the per-target correlation measure, and matched the best methods on per-residue scores. It clearly outperformed ProQ4, which uses the same input features but a 1D-CNN, and a variant using only amino-acid identity performed almost as well as the full model. Learning local and global scores together lowered global error, but did not help local error.
Methodology
The authors represented each predicted protein structure as a graph whose nodes are amino acids and whose edges link neighbouring residues in the chain or residues close in 3D space. A message-passing graph network was trained to predict both whole-model and per-residue quality scores on decoys from earlier CASP competitions, then tested on CASP13 targets. They ran ablations on network depth, contact cutoff, input features and whether local and global scores are learned together.
Limitations
GraphQA was tested offline after the fact, whereas rival methods' numbers come from the blind CASP13 competition, so the comparison is not fully fair, as the authors themselves note. The graph uses only alpha-carbon positions, not full atomic detail. The refinement experiment is only a weak signal (a small average cosine similarity), so it does not show the model can actually improve structures.
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.
The graph structure itself carried much of the signal.
GraphQA clearly outperformed ProQ4, which uses the same input features with a 1D-CNN, for ranking protein models on CASP13, and a variant using only amino-acid identity was almost as good.
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?