Research map / Probabilistic and scientific modeling / Molecules and proteins
Evaluating Representation Learning on the Protein Structure Universe
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.
Introduces ProteinWorkshop, a modular benchmark for evaluating geometric GNN protein structure encoders with denoising-based pretraining on AlphaFoldDB and node- and graph-level downstream tasks.
Problem
Protein structure representation learning lacks a standardized benchmark covering both node-level and graph-level tasks, large-scale pretraining on predicted structures, and comparison across invariant and equivariant geometric GNNs.
Approach
A unified, modular benchmark that combines multiple geometric GNN architectures, several structural featurisation schemes, denoising-based auxiliary and pretraining tasks, and large predicted-structure corpora (AlphaFoldDB 2.27M representatives, ESM Atlas) with storage-efficient dataloaders, and evaluates them on node-level and graph-level protein tasks.
Key results
- Auxiliary denoising tasks improve baseline performance on 50 of 60 model-task combinations.
- Pretraining on AlphaFoldDB improves GCPNet with backbone dihedral features, e.g., fold accuracy rising from 38.34 to 42.81 with structure denoising.
- ESM-2-650M with structural featurisation matches state-of-the-art GNNs on (super)family fold and gene ontology prediction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add sequence or structure denoising as an auxiliary task (ablation) | Geometric GNNs without pretraining · Downstream tasks across models and primary tasks | task metric: 50 out of 60 combinations | Helped |
| Pretrain on AlphaFold structures (ablation) | Rotation-invariant and equivariant GNNs · Downstream tasks | task metric | Helped |
| Add structural features to ESM-2-650M (ablation) | ESM-2-650M · GO-BP | F_max: 0.462* -> 0.472$ pm0.00$ | Helped |
| Add structural features to ESM-2-650M (ablation) | ESM-2-650M · GO-MF | F_max: 0.546* -> 0.583$ pm0.00$ | Helped |
| Add backbone torsion angles to Cα + virtual-angle features (ablation) | Geometric GNNs · Pretraining tasks on AlphaFoldDB | pretraining metric | Helped |
| Add backbone torsion angles to Cα + virtual-angle features (ablation) | GCPNet · Inverse Folding (AlphaFoldDB validation) | perplexity: 6.243 -> 3.839 | Helped |
| Add backbone torsion angles to Cα + virtual-angle features (ablation) | GCPNet · Structure denoising (AlphaFoldDB validation) | RMSE: 0.0679 -> 0.0561 | Helped |
| Add backbone torsion angles to Cα + virtual-angle features (ablation) | GCPNet · Sequence denoising (AlphaFoldDB validation) | accuracy: 44.81 -> 59.54 | Helped |
| Add sidechain torsion angles (ablation) | Geometric GNNs · Downstream tasks | task metric | Mixed |
| Use structure denoising as an auxiliary task (ablation) | MACE · GO and Reaction tasks | training convergence | Helped |
| Use sequence denoising instead of no auxiliary task (ablation) | SchNet with Cα features · Fold classification (Fold) | accuracy: 20.71$ pm0.01$ -> 24.58$ pm0.01$ | Helped |
| Use structure denoising instead of no auxiliary task (ablation) | SchNet with Cα features · GO-BP | F_max: 0.314$ pm0.00$ -> 0.343$ pm0.00$ | Helped |
Surprising findings
- Backbone-only featurisation (C-alpha with virtual angles and backbone torsions) performs as well or better than adding full sidechain torsion information.
- Pretraining benefits equivariant GNNs more than invariant models.
Limitations
- Spherical equivariant models (TFN, MACE) could not be pretrained due to high computational requirements.
- Only denoising-based pretraining objectives were benchmarked; contrastive and generative objectives were not evaluated.
- Mostly C-alpha featurisation was explored because full-atom graphs are computationally expensive.
Experimental setup
Datasets: AlphaFoldDB PDB CATH ASTRAL ESM Atlas Fold classification (SCOP 1.75) Gene Ontology Reaction class Antibody developability PPI site Metal binding site PTM site
Benchmarks: GO-BP GO-MF GO-CC Fold Family Superfamily Reaction Antibody developability PPI site Inverse folding PTM site
Metrics: accuracy F_max AUPRC perplexity ROC-AUC
Baselines: SchNet EGNN TFN MACE GCPNet GearNet-Edge ESM-2-650M
Builds on: ESM-2 GearNet GCPNet NoisyNodes FoldComp
Closest papers by meaning
- Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs (NeurIPS 2025 · similarity 0.90)
- Protein 3D Graph Structure Learning for Robust Structure-Based Protein Property Prediction (AAAI 2024 · similarity 0.89)
- Learning Complete Protein Representation by Dynamically Coupling of Sequence and Structure (NeurIPS 2024 · similarity 0.89)
- Enhancing Graph Contrastive Learning for Protein Graphs from Perspective of Invariance (ICML 2025 · similarity 0.89)
- Clustering for Protein Representation Learning (CVPR 2024 · similarity 0.89)
- AtomSurf: Surface Representation for Learning on Protein Structures (ICLR 2025 · similarity 0.89)
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.