Research map / Graph learning / Graph self-supervised learning
Enhancing Graph Contrastive Learning for Protein Graphs from Perspective of Invariance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes biology-aware graph augmentations for protein GCL: FCI (spectral, chemically weighted 2D edge sampling) and 3-PSI (dihedral-angle and secondary-structure rotations in 3D), combined in a GCN-based framework evaluated on four protein tasks.
Problem
Existing graph contrastive learning for protein representation learning uses 2D topology-only augmentations that ignore biological properties, while 3D structure-based protein graph augmentation is unexplored and traditional 3D perturbations or homology-model augmentations distort or reduce diversit
Approach
Augmentations guided by protein invariances: FCI preserves functional communities via spectral bounds with side-chain chemical similarity as edge weights, and 3-PSI perturbs backbone dihedral angles and rotates secondary structures to keep peptide planes and helices/sheets intact.
Key results
- Combined 3-PSI-Alpha + FCI reaches 89.0% average on FOLD and 0.885 Fmax on EC, versus 86.3% and 0.853 for the backbone without augmentation.
- FCI alone yields a 4.4% relative gain in GOCC over the best 2D topology-based GCL method.
- 3-PSI-Diag achieves an 11.4% relative improvement in GOCC over the best baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FCI to the protein GCL framework (ablation) | proposed protein GCL · EC | Fmax: 0.853 -> 0.878 | Helped |
| add 3-PSIDiag to the protein GCL framework (ablation) | proposed protein GCL · EC | Fmax: 0.853 -> 0.873 | Helped |
| add 3-PSIAlpha to the protein GCL framework (ablation) | proposed protein GCL · EC | Fmax: 0.853 -> 0.874 | Helped |
| add 3-PSIDiag and FCI (ablation) | proposed protein GCL · EC | Fmax: 0.853 -> 0.883 | Helped |
| add 3-PSIAlpha and FCI (ablation) | proposed protein GCL · EC | Fmax: 0.853 -> 0.885 | Helped |
| use FCI instead of the strongest listed 2D topology-based GCL baseline | protein GCL · GO-CC | Fmax: 0.440 -> 0.446 | Helped |
| use 3-PSIDiag instead of traditional 3D augmentation | protein GCL · GO-CC | Fmax: 0.432 -> 0.449 | Helped |
| combine 3-PSIAlpha and FCI instead of CDConv | protein representation learning · Reaction | accuracy: 88.6 -> 89.0 | Helped |
| increase α-helix and β-sheet rotations from 1 to 2 | 3-PSIAlpha · FOLD and GO-BP | accuracy / Fmax | Mixed |
| increase α-helix and β-sheet rotations from 1 to 3 | 3-PSIAlpha · Reaction and GO-CC | accuracy / Fmax | Mixed |
| set loss weight λ to 1.0 | proposed method · EC | Fmax: 0.882 -> 0.885 | Helped |
Surprising findings
- Combining 3-PSI-Diag and 3-PSI-Alpha was avoided because it disrupts protein structure, so they are used separately.
Limitations
- Combining both 3-PSI strategies was found to disrupt protein structure, so they must be used separately.
- Augmentation strength (epsilon, number of rotations) requires task-specific tuning.
Experimental setup
Datasets: Protein Fold classification (FOLD) Enzyme Reaction classification Gene Ontology (GO) BP/MF/CC Enzyme Commission (EC) PDBbind
Benchmarks: FOLD Reaction GO EC PDBbind
Metrics: accuracy Fmax RMSE Pearson Spearman
Baselines: GVP 3DCNN GraphQA IEConv GearNet-Edge-IEConv ProNet CDConv GraphCL Auto-GCL MolCLR T-MGCL GCL-Span
Builds on: GraphCL CDConv ProNet CI-GCL GraphTrans
Closest papers by meaning
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.89)
- Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs (NeurIPS 2025 · similarity 0.88)
- Community-Invariant Graph Contrastive Learning (ICML 2024 · similarity 0.87)
- SOLA-GCL: Subgraph-Oriented Learnable Augmentation Method for Graph Contrastive Learning (AAAI 2025 · similarity 0.87)
- Protein 3D Graph Structure Learning for Robust Structure-Based Protein Property Prediction (AAAI 2024 · similarity 0.87)
- Hierarchical Topology Isomorphism Expertise Embedded Graph Contrastive Learning (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.