Research map / Probabilistic and scientific modeling / Molecules and proteins
SurfPro: Functional Protein Design Based on Continuous Surface
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 SurfPro, a surface-based generative model that designs protein sequences from a biochemically annotated molecular surface using a hierarchical geometric encoder and autoregressive decoder, evaluated on inverse folding, binder design, and enzyme design.
Problem
Designing functional proteins requires specifying both geometric shape and biochemical properties of a surface, whereas standard inverse folding only conditions on backbone geometry.
Approach
Encode a smoothed, octree-compressed molecular surface point cloud with hydrophobicity and charge features using local equivariant graph convolutions followed by frame-averaged global multi-head attention (FAMHA), then decode the sequence autoregressively.
Key results
- On CATH 4.2, SurfPro achieves 57.78% recovery rate and 3.13 perplexity, versus 54.16% and 4.47 for LM-DESIGN.
- In binder design, SurfPro-Pretrain reaches an average success rate of 26.22% across six targets, versus 22.29% for SurfPro and the best prior baseline at about 16.37% (LM-DESIGN).
- In enzyme design, SurfPro achieves 42.23% average success rate versus 37.58% for LM-DESIGN, rising to 43.63% with PDB pretraining; ESP score is comparable to LM-DESIGN (0.8931 vs 0.9037).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SurfPro instead of LM-DESIGN | inverse folding · CATH 4.2 | recovery rate (%): 54.16 -> 57.78 | Helped |
| Use SurfPro instead of LM-DESIGN | inverse folding · CATH 4.2 | perplexity: 4.47 -> 3.13 | Helped |
| Use SurfPro instead of PiFold | binder design · six targets, average | AF2 pAE_interaction: 20.5785 -> 16.9485 | Helped |
| Pretrain on PDB surfaces | SurfPro binder design · six targets, average | success rate (%): 22.29 -> 26.22 | Helped |
| Pretrain on PDB surfaces | SurfPro enzyme design · five substrates, average | success rate (%): 42.23 -> 43.63 | Helped |
| Use SurfPro instead of LM-DESIGN | enzyme design · five substrates, average | ESP score: 0.9037 -> 0.8931 | Hurt |
| Add global landscape modeling (ablation) | SurfPro · CATH 4.2 | recovery rate (%): 15.51 -> 57.78 | Helped |
| Add local perspective modeling (ablation) | SurfPro · CATH 4.2 | recovery rate (%): 34.70 -> 57.78 | Helped |
| Use two biochemical features rather than five (ablation) | SurfPro · CATH 4.2 | recovery rate (%): 56.78 -> 57.78 | Helped |
| Use both hydrophobicity and charge rather than hydrophobicity only (ablation) | SurfPro · CATH 4.2 | recovery rate (%): 51.60 -> 57.78 | Helped |
| Use both hydrophobicity and charge rather than charge only (ablation) | SurfPro · CATH 4.2 | recovery rate (%): 21.34 -> 57.78 | Helped |
| Add biochemical features (ablation) | SurfPro · CATH 4.2 | recovery rate (%): 6.45 -> 57.78 | Helped |
Surprising findings
- Removing the sorting of surface vertices sharply degrades performance (recovery 14.95% vs 57.78%).
- On binder design, SurfPro's average AF2 pAE_interaction is worse than MaSIF (16.95 vs 16.13), though MaSIF is far slower.
Limitations
- Refinement-oriented: starts from existing positive binders rather than de novo design.
- Requires a precomputed molecular surface (MSMS) and can fail on some structures.
- Binder and enzyme success rates rely on AF2 pAE_interaction and ESP scores as proxies for function, not experimental validation.
Experimental setup
Datasets: CATH 4.2 binder design complexes (Bryant/ AF2 pAE study data) enzyme design datasets (EnzymeMap-style collections) PDB surfaces
Benchmarks: CATH 4.2 binder design (6 targets) enzyme design (5 substrates)
Metrics: perplexity recovery rate AF2 pAE_interaction success rate ESP score
Baselines: ProteinMPNN PiFold LM-DESIGN MaSIF Random Baseline
Builds on: EGNN frame averaging dMaSIF MaSIF LM-Design
Closest papers by meaning
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.88)
- Generating Highly Designable Proteins with Geometric Algebra Flow Matching (NeurIPS 2024 · similarity 0.86)
- Learning to engineer protein flexibility (ICLR 2025 · similarity 0.86)
- ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids (ICLR 2025 · similarity 0.86)
- Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates (ICML 2024 · similarity 0.86)
- All-atom inverse protein folding through discrete flow matching (ICML 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.