Research map / Probabilistic and scientific modeling / Molecules and proteins
AtomSurf: Surface Representation for Learning on Protein Structures
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.
Adapts DiffusionNet surface encoders for protein learning and fuses them node-wise with graph encoders across all layers via a bipartite message-passing graph, reaching state-of-the-art on Atom3D, binding-site and PINDER tasks.
Problem
Surface-based learning on protein structures has never been fairly benchmarked against graph and other representations, and existing surface methods use surface information in isolation or fuse it only globally at the output.
Approach
Adapt DiffusionNet to protein surfaces (batch support, batch norm, scale-aware learned diffusion times initialized near 10 Å) and jointly train surface and graph encoders connected by a bipartite proximity graph so node/vertex features are shared at every layer.
Key results
- AtomSurf reaches 90.9 AuROC on Atom3D PIP, 71.6 on MSP and 61.7 R_l / 85.7 R_g on PSR, without pretraining and with about 600k parameters.
- AtomSurf-bench, a benchmark-protocol model with 200k parameters, reaches 87.6 PIP AuROC and 70.7 MSP AuROC, while pure surface Surface Diff gets 0.837 PIP and 0.5 MSP.
- On MaSIF-ligand, the method reports 88.2 balanced accuracy versus 81 for the runner-up, and on PINDER-Pair (holo) AuROC 92.8 versus 80.1 for ProNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add batch support, Batch Normalization, and scale-informed diffusion-time initialization (ablation) | DiffusionNet · RNA segmentation | test accuracy: 80.9% -> 84.1% | Helped |
| Use surface encoding alone instead of the integrated surface-graph model | AtomSurf-bench · PIP | AuROC: 0.876 -> 0.837 | Hurt |
| Add the integrated surface-graph representation | AtomSurf-bench · PIP | AuROC: 0.669 -> 0.876 | Helped |
| Add ESM embeddings to graph initial node features (ablation) | AtomSurf · Pinder-Pair Holo | AuROC: 93 -> 92.8 | No clear effect |
| Add ESM embeddings to graph initial node features (ablation) | AtomSurf · Pinder-Site Holo | AuROC: 87.6 -> 88.3 | Helped |
| Replace the sequential architecture with bipartite message passing (ablation) | AtomSurf-bench · MSP | AuROC: 60.9 -> 70.7 | Helped |
| Use attentive, potentially long-distance message passing instead of localized message passing (ablation) | AtomSurf-bench · benchmark tasks | performance | Hurt |
| Use several message-passing blocks and more interconnected networks (ablation) | AtomSurf-bench · benchmark tasks | performance | Helped |
| Remove ESM embeddings from graph initial node features (ablation) | AtomSurf · PINDER | AuROC: 92.8 vs 93; 88.4 vs 87.6; 87.1 vs 87; 88.3 vs 87.6; 84.2 vs 82.4; 82 vs 80.9 | Mixed |
Surprising findings
- Pure surface encoding (Surface Diff) underperforms even on protein-protein interaction and at chance level on MSP (0.5 AuROC) when input features are held minimal.
- Removing ESM embeddings reduces performance only modestly on PINDER, with the drop shrinking under stricter splits, suggesting sequence features may aid memorization.
Limitations
- High memory requirements of surface operations limit batch size.
- Surface-only learning underperforms graph methods under fixed minimal input features.
- Validation on real-life scenarios is left as future work.
Experimental setup
Datasets: Atom3D PIP Atom3D MSP Atom3D PSR RNA segmentation MaSIF-ligand AbAg PINDER
Benchmarks: Atom3D PIP Atom3D MSP Atom3D PSR RNA segmentation MaSIF-ligand AbAg PINDER
Metrics: AuROC R_l R_g balanced accuracy MCC accuracy
Baselines: 3DCNN ENN Graph (GCN) ProNet GVP GearNet-ESM HMR MaSIF GEP DGCNN DeltaConv
Builds on: DiffusionNet MaSIF ProNet GVP Atom3D
Closest papers by meaning
- Pre-Training Protein Bi-level Representation Through Span Mask Strategy On 3D Protein Chains (ICML 2024 · similarity 0.89)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.89)
- Surface-VQMAE: Vector-quantized Masked Auto-encoders on Molecular Surfaces (ICML 2024 · similarity 0.87)
- Generalized Implicit Neural Representations for Dynamic Molecular Surface Modeling (AAAI 2025 · similarity 0.87)
- Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.86)
- Atom-level Protein Representation Learning Improves Protein Structure Prediction (NeurIPS 2026 · 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.