Research map / Probabilistic and scientific modeling / Molecules and proteins
Clustering for Protein Representation Learning
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 a neural clustering framework that iteratively groups amino acids into spherical clusters, aggregates them into medoids, and nominates top-scoring clusters to learn hierarchical protein representations for classification.
Problem
Protein representation learning treats all amino acids equally, ignoring that only a few critical residues determine a protein's structure and function.
Approach
Iteratively apply spherical cluster initialization around each amino acid, attention-based cluster representation extraction, and GCN-based cluster nomination over 1D sequence and 3D structure, so the network progressively keeps only the most representative amino acids for prediction.
Key results
- Achieves Fmax 0.866 on EC number prediction vs 0.820 for CDConv (+5.6%).
- Reaches 63.1% accuracy on fold classification (Fold split) vs 56.7% for CDConv.
- Achieves 89.6% accuracy on enzyme reaction classification vs 88.5% for CDConv.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace neural clustering with GRACLUS (ablation) | Neural clustering framework · EC number prediction | Fmax: 0.866 -> 0.792 | Hurt |
| replace cluster representation extraction with average pooling (ablation) | Neural clustering framework · EC number prediction | Fmax: 0.866 -> 0.817 | Hurt |
| replace learned attention scores with random scores (ablation) | Neural clustering framework · EC number prediction | Fmax: 0.866 -> 0.765 | Hurt |
| reduce clustering iterations from 4 to 1 (ablation) | Neural clustering framework · Enzyme reaction classification | accuracy: 89.6 -> 84.7 | Hurt |
| reduce clustering iterations from 4 to 2 (ablation) | Neural clustering framework · Enzyme reaction classification | accuracy: 89.6 -> 88.1 | Hurt |
| reduce clustering iterations from 4 to 3 (ablation) | Neural clustering framework · Enzyme reaction classification | accuracy: 89.6 -> 89.0 | Hurt |
| increase clustering iterations from 4 to 5 (ablation) | Neural clustering framework · Enzyme reaction classification | accuracy: 89.6 -> 86.3 | Hurt |
| randomly remove 10% of structure coordinates (ablation) | Neural clustering framework · Protein fold classification (Superfamily) | accuracy: 81.2 -> 78.7 | Hurt |
Surprising findings
- Outperforms ESM-1b and other pretrained protein language models on fold classification without any pretraining or self-supervision.
Limitations
- Performance drops notably when many coordinates are missing (e.g., 56.7% at 20% missing on Fold split).
- Accuracy drops at 5 iterations, indicating sensitivity to the number of clustering iterations and hyperparameters (radius, nomination fraction).
Experimental setup
Datasets: EC number prediction dataset (Gligorijevic et al.) GO term prediction dataset protein fold classification dataset (Hou et al.) enzyme reaction dataset (Gligorijevic/IEConv processed)
Benchmarks: EC number prediction GO term prediction protein fold classification enzyme reaction classification
Metrics: Fmax accuracy
Baselines: CDConv GearNet ProNet New IEConv GVP GCN GAT LM-GVP ESM-1b
Builds on: CDConv GearNet ProNet GraphQA
Closest papers by meaning
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.89)
- Learning Complete Protein Representation by Dynamically Coupling of Sequence and Structure (NeurIPS 2024 · similarity 0.86)
- Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs (NeurIPS 2025 · similarity 0.86)
- Boosting Protein Graph Representations through Static-Dynamic Fusion (ICML 2025 · similarity 0.85)
- Protein 3D Graph Structure Learning for Robust Structure-Based Protein Property Prediction (AAAI 2024 · similarity 0.85)
- Enhancing Molecular Representation Learning Through the Combination of 3D and 2D Graph Machine Learning (Student Abstract) (AAAI 2025 · similarity 0.85)
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.