Research map / Probabilistic and scientific modeling / Molecules and proteins
Surface-VQMAE: Vector-quantized Masked Auto-encoders on Molecular Surfaces
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 Surface-VQMAE, an unsupervised masked auto-encoder pretrained on protein molecular surface point clouds, with Morton-order patching, a SurfFormer Transformer, and a vector-quantized (Gumbel-softmax) codebook for masked tokens.
Problem
Self-supervised representation learning for proteins has focused on sequences and 3D structures, leaving molecular surface information largely unexploited; surface point clouds are sparse, unordered, and the MAE latent space is unconstrained between encoder and decoder.
Approach
Partition surface point clouds into FPS/KNN patches ordered by Morton code, encode them with a geometric point network and a geometry-aware SurfFormer, and replace MAE mask tokens with Gumbel-softmax-relaxed codebook vectors (VQ) so the latent posterior is discrete, with reconstruction of patch centers and curvature prediction as pretext targets.
Key results
- Binding site scoring: AUROC 0.9478 and balanced accuracy 87.16%, beating PIsToN (AUROC 0.9355).
- Antibody-antigen affinity on crystal structures: Pearson 0.411 vs NERE 0.340 on the SAbDab test set.
- Mutant effect (SKEMPI.v2): per-structure Pearson 0.4694 and Spearman 0.4324, the best reported among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Surface-VQMAE instead of PIsToN | Protein interface scoring · Protein interface scoring test set | AUROC: 93.55 -> 94.78 | Helped |
| Use Surface-VQMAE instead of NERE | Antibody-antigen affinity prediction · SAbDab ZDock structures | Pearson correlation: 0.234 -> 0.358 | Helped |
| Use Surface-VQMAE instead of RDE-Net | Mutant effect prediction · SKEMPI.v2 | Per-structure Pearson: 0.4448 -> 0.4694 | Helped |
| Use Surface-VQMAE instead of DDGPred | Mutant effect prediction · SKEMPI.v2 | Overall Pearson: 0.6580 -> 0.6482 | Hurt |
| Add SurfFormer to the backbone (ablation) | Surface-VQMAE · Downstream tasks | Performance | Helped |
| Add the MAE framework (ablation) | Surface-VQMAE · Downstream tasks | Performance | Helped |
| Add vector quantization (ablation) | Surface-VQMAE · Downstream tasks | Performance | Helped |
Limitations
- Relies on a single downsampling technique (FPS) without exploring alternatives.
- Does not thoroughly examine the influence of noise or outliers in surface data.
- Generalizability may be limited to the specific datasets used (SKEMPI, SAbDab).
Experimental setup
Datasets: PDB-REDO PRISM ZDock benchmark PDBBind SAbDab SKEMPI.v2 Myung et al. (2022) antibody-antigen affinity set
Benchmarks: PRISM ZDock benchmark PDBBind SAbDab SKEMPI.v2
Metrics: AUROC AP balanced accuracy F1 precision recall Pearson Spearman RMSE MAE
Baselines: PIsToN dMaSIF MaSIF NERE RDE-Net PPIFormer DDGPred ESM-IF
Builds on: MAE PointGPT dVAE dMaSIF Point-MAE
Closest papers by meaning
- AtomSurf: Surface Representation for Learning on Protein Structures (ICLR 2025 · similarity 0.87)
- SurfPro: Functional Protein Design Based on Continuous Surface (ICML 2024 · similarity 0.84)
- Generalized Implicit Neural Representations for Dynamic Molecular Surface Modeling (AAAI 2025 · similarity 0.84)
- Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.84)
- Mol-AE: Auto-Encoder Based Molecular Representation Learning With 3D Cloze Test Objective (ICML 2024 · similarity 0.83)
- nach0-pc: Multi-task Language Model with Molecular Point Cloud Encoder (AAAI 2025 · similarity 0.83)
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.