Research map / Probabilistic and scientific modeling / Molecules and proteins
Learning Complete Protein Representation by Dynamically Coupling of Sequence and Structure
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 CoupleNet, a protein representation model that couples sequence and 3D structure via two dynamically thresholded graphs (sequential and radius) with complete geometric features, convolving on nodes and edges simultaneously.
Problem
Learning protein representations that effectively model the intricate inter-dependencies between amino acid sequences and 3D tertiary structures, which existing sequence-only, structure-only, and GNN-based hybrid methods handle poorly, leading to weak function prediction especially for low-similarit
Approach
Construct two graph types (sequential graph by residue index distance and radius graph by C-alpha distance) whose thresholds grow with network depth, use complete geometric features at amino acid and backbone levels, and perform node and edge convolutions simultaneously in a complete message passing scheme.
Key results
- CoupleNet reaches 60.6% accuracy on the Fold test set and 82.1% on SuperFamily, versus 56.7% and 77.7% for the prior best CDConv.
- On GO prediction CoupleNet achieves Fmax 0.467 (BP), 0.669 (MF), 0.494 (CC), outperforming CDConv's 0.453, 0.654, 0.479.
- On EC number prediction CoupleNet achieves Fmax 0.866 versus 0.820 for CDConv; ablation shows removing structure drops Fold accuracy to 26.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CoupleNet instead of CDConv | protein representation learning · Fold classification — Fold | accuracy: 56.7 -> 60.6 | Helped |
| use CoupleNet instead of CDConv | protein representation learning · Fold classification — SuperFamily | accuracy: 77.7 -> 82.1 | Helped |
| use CoupleNet instead of CDConv | protein representation learning · Fold classification — Family | accuracy: 99.6 -> 99.7 | Helped |
| use CoupleNet instead of CDConv | protein representation learning · Enzyme reaction classification | accuracy: 88.5 -> 89.0 | Helped |
| use CoupleNet instead of CDConv | protein representation learning · GO-BP | Fmax: 0.453 -> 0.467 | Helped |
| use CoupleNet instead of CDConv | protein representation learning · GO-MF | Fmax: 0.654 -> 0.669 | Helped |
| use CoupleNet instead of CDConv | protein representation learning · GO-CC | Fmax: 0.479 -> 0.494 | Helped |
| use CoupleNet instead of CDConv | protein representation learning · EC number prediction | Fmax: 0.820 -> 0.866 | Helped |
| add backbone-level representations beyond amino-acid-level representations (ablation) | CoupleNetaa · Fold classification — Fold | accuracy: 57.8 -> 60.6 | Helped |
| add sequence information (ablation) | CoupleNet without sequence information · GO-BP | Fmax: 0.441 -> 0.467 | Helped |
| add structure information (ablation) | CoupleNet without structure information · Fold classification — Fold | accuracy: 26.1 -> 60.6 | Helped |
| add backbone torsion angles Φ, Ψ, Ω (ablation) | CoupleNet without backbone torsion angles · Fold classification — Fold | accuracy: 60.3 -> 60.6 | Helped |
Surprising findings
- Removing structure information causes a much larger performance drop than removing sequence information, indicating structure carries more discriminative signal in this setup.
Limitations
- Requires both protein sequences and 3D structures as input, so it cannot be applied to sequence-only data.
- No large-scale pretraining; the authors note a future multimodal large-scale model is needed.
- Evaluated only on supervised downstream tasks trained from scratch on the benchmarks.
Experimental setup
Datasets: Fold classification dataset (16,712 proteins) Enzyme reaction dataset (37,248 proteins) GO term prediction dataset EC number prediction dataset
Benchmarks: Fold classification (Fold, SuperFamily, Family) Enzyme reaction classification GO-BP GO-MF GO-CC EC number prediction
Metrics: accuracy Fmax
Baselines: ResNet Transformer 3DCNN_MQA IEConv GraphQA GVP ProNet GearNet CDConv
Builds on: ProNet CDConv GearNet ComENet trRosetta geometries
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 0.87)
- Boosting Protein Graph Representations through Static-Dynamic Fusion (ICML 2025 · similarity 0.87)
- Clustering for Protein Representation Learning (CVPR 2024 · similarity 0.86)
- Atom-level Protein Representation Learning Improves Protein Structure Prediction (NeurIPS 2026 · similarity 0.86)
- Multi-Scale Representation Learning for Protein Fitness Prediction (NeurIPS 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.