Research map / Graph learning / Graph self-supervised learning
Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs
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 SSHG, a two-level hierarchical protein graph (residue-level SCHull subgraphs per DSSP secondary-structure motif plus a motif-level inter-structural graph with local-frame orientation edges), with two GNN stages, claiming maximal expressiveness and efficiency gains on EC reaction classification and LBA.
Problem
Residue-level protein GNNs struggle to capture multiscale secondary-structure features and long-range dependencies efficiently; large-radius radial graphs are computationally expensive.
Approach
Segment proteins into DSSP secondary-structure motifs, build a sparse SCHull graph within each motif and a coarse SCHull graph over motif centers with local-frame relative-orientation edge features, then apply two GNNs in sequence; the framework is modular in the choice of GNN and is shown to preserve maximal expressiveness.
Key results
- ProNet+SSHG reaches 87.2% EC accuracy with ~140 s/epoch vs 210 for ProNet-Backbone, matching IEConv (87.2%) with 1.3M vs 9.8M params.
- Mamba+SSHG achieves 88.4% EC accuracy, the best reported, and LBA RMSE 1.399 / Pearson 0.614 on PDBbind.
- On GVP-GNN, SSHG raises EC accuracy from 68.5% to 73.6% and cuts memory from 8,930 to 1,416 MiB at comparable settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SSHG to Mamba | Mamba · EC reaction classification | test accuracy: 85.9$\pm 0.2$ -> 88.4$\pm 0.3$ | Helped |
| Add SSHG to ProNet-Backbone | ProNet-Backbone · EC reaction classification | test accuracy: 86.4$\pm 0.2$ -> 87.2$\pm 0.2$ | Helped |
| Add SSHG to GVP-GNN | GVP-GNN · EC reaction classification | test accuracy: 68.5$\pm 0.1$ -> 73.6$\pm 0.1$ | Helped |
| Add SSHG to GCN | GCN · EC reaction classification | test accuracy: 66.5 -> 71.2 | Helped |
| Add SSHG to Mamba | Mamba · LBA | RMSE: 1.457 ± 0.004 -> \textbf{1.399}$\pm 0.003$ | Helped |
| Add SSHG to Mamba | Mamba · LBA | Pearson correlation: 0.565 ± 0.003 -> \textbf{0.614}$\pm 0.003$ | Helped |
| Add SSHG to Mamba | Mamba · LBA | Spearman correlation: 0.554 ± 0.004 -> \textbf{0.610}$\pm 0.004$ | Helped |
| Add SSHG to ProNet-Backbone | ProNet-Backbone · LBA | RMSE: 1.458 -> 1.435$\pm 0.004$ | Helped |
| Add SSHG to ProNet-Backbone | ProNet-Backbone · LBA | Pearson correlation: 0.546 -> 0.579$\pm 0.004$ | Helped |
| Add SSHG to ProNet-Backbone | ProNet-Backbone · LBA | Spearman correlation: 0.550 -> 0.591$\pm 0.003$ | Helped |
| Use SSHG graph construction instead of increasing radial cutoffs (ablation) | ProNet · EC reaction classification | test accuracy: 87.0 -> 87.2 | Helped |
| Use SSHG graph construction instead of increasing radial cutoffs (ablation) | GVP-GNN · EC reaction classification | test accuracy: 69.2 -> 73.6 | Helped |
Surprising findings
- Assigning more parameters to the first-stage (residue-level) GNN slightly improves accuracy than a larger inter-structural stage.
- Non-injective mean/sum pooling is used in practice despite the injectivity assumption required by the expressiveness theorem.
Limitations
- Expressiveness theorem assumes injective aggregation/readout, which is not enforced in practice.
- Evaluated only on EC reaction classification and LBA; other tasks (fold classification, PPI) left to future work.
- Secondary-structure segmentation depends on DSSP and fixed motif definitions.
Experimental setup
Datasets: EC reaction classification dataset (enzyme reaction) PDBbind (ligand binding affinity, 30% identity split)
Benchmarks: EC reaction classification (React) PDBbind LBA
Metrics: test accuracy RMSE Pearson correlation Spearman correlation training time per epoch memory usage number of edges parameters
Baselines: GCN IEConv DWNN GearNet HoloProt GVP-GNN ProNet-Backbone TAPE
Builds on: SCHull graph DSSP ProNet GVP-GNN Mamba
Closest papers by meaning
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.90)
- Enhancing Graph Contrastive Learning for Protein Graphs from Perspective of Invariance (ICML 2025 · similarity 0.88)
- Boosting Protein Graph Representations through Static-Dynamic Fusion (ICML 2025 · similarity 0.88)
- MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation (ICLR 2025 · similarity 0.88)
- Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer (ICLR 2025 · similarity 0.88)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.87)
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.