Research map / Probabilistic and scientific modeling / Molecules and proteins
De novo Protein Design Using Geometric Vector Field Networks
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 Vector Field Network (VFN), a frame-based structure encoder using learnable vector computations between frame-anchored virtual atoms, applied to protein diffusion (VFN-Diff) and inverse folding (VFN-IF, VFN-IFE with ESM).
Problem
Protein structure encoders for de novo design must model residue backbone frames without atom-level features, and prior frame-based encoders such as IPA are limited by a single pooled distance scalar ("atom representation bottleneck").
Approach
A vector field operator applies learnable linear-like weights to 3D virtual atom coordinates expressed in a common local frame, outputting multiple geometric vectors (decomposed into unit directions and RBF-encoded lengths) that feed attention-based residue and virtual-atom updates; real backbone atoms can also be used as virtual atoms.
Key results
- In protein diffusion, VFN-Diff reaches 67.04% scTM>0.5 designability vs. 53.58% for FrameDiff (8 sequences, 500 steps, noise 1.0).
- VFN-IF achieves 54.74% sequence recovery on CATH 4.2 vs. 51.66% for PiFold; VFN-IF+ trained on the full PDB reaches 57.14%.
- VFN-IFE (VFN with ESM via LoRA) reaches 62.67% recovery on CATH 4.2 vs. 55.65% for LM-Design.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FrameDiff's IPA point attention with VFN | FrameDiff · Protein diffusion, Noise Scale 1.0 | scTM0.5 (%): 53.58% -> 67.04% | Helped |
| Replace FrameDiff's IPA point attention with VFN | FrameDiff · Protein diffusion, Noise Scale 1.0 | diversity (%): 51.98% -> 66.54% | Helped |
| Equip VFN-IF with ESM and LoRA | VFN-IF · CATH 4.2 | sequence recovery (%): 55.65% -> 62.67% | Helped |
| Train VFN-IF on the entire PDB dataset | VFN-IF · CATH 4.2 | sequence recovery (%): 54.74% -> 57.14% | Helped |
| Use VFN-IF rather than PiFold | Inverse folding · CATH 4.2 | sequence recovery (%): 51.66% -> 54.74% | Helped |
| Use VFN-Diff + VFN-IFE instead of FrameDiff + MPNN | Whole protein-design pipeline · One-shot, Noise Scale 1.0 | scTM>0.5 (%): 37.25% -> 45.43% | Helped |
| Remove edge features (ablation) | VFN-IF · CATH 4.2, 15 layers | sequence recovery (%): 54.08 -> 54.7 | Helped |
| Use 32 vectors instead of 16 (ablation) | VFN-IF · CATH 4.2 | sequence recovery (%): 53.97 -> 54.28 | Helped |
| Use 64 vectors instead of 32 (ablation) | VFN-IF · CATH 4.2 | sequence recovery (%): 54.28 -> 53.63 | Hurt |
| Use V-MLP rather than regular MLP (ablation) | VFN-IF · CATH 4.2 | sequence recovery (%): 53.42 -> 54.28 | Helped |
| Add the vector field operator (ablation) | VFN-IF · CATH 4.2 | sequence recovery (%): 35.43 -> 54.28 | Helped |
| Add g_i,j to edge aggregation (ablation) | VFN-IF · CATH 4.2 | sequence recovery (%): 53.05 -> 54.28 | Helped |
Surprising findings
- Removing the entire vector field operator drops recovery from 54.28% to 35.43%, showing the operator is critical.
- VFN-IF with only 5 layers (52.68%) already beats a 10-layer PiFold, and VFN-IF does not saturate with depth as PiFold does.
Limitations
- Structure recovery and whole-pipeline results are one-shot evaluations with limited sample sizes.
- Comparison with RFDiffusion was deliberately excluded as out of scope.
- Diversity and pdbTM metrics were modified from FrameDiff's protocol, making comparisons less standard.
Experimental setup
Datasets: CATH 4.2 TS50 TS500 PDB
Benchmarks: CATH 4.2 TS50 TS500
Metrics: sequence recovery perplexity scTM scRMSD designability diversity pdbTM
Baselines: IPA (FrameDiff) PiFold ProteinMPNN LM-Design ESM-IF GVP AlphaDesign StructGNN GraphTrans GCA
Builds on: FrameDiff PiFold invariant point attention (IPA) ESM LM-Design
Closest papers by meaning
- Sidechain conditioning and modeling for full-atom protein sequence design with FAMPNN (ICML 2025 · similarity 0.86)
- Pre-Training Protein Bi-level Representation Through Span Mask Strategy On 3D Protein Chains (ICML 2024 · similarity 0.86)
- AtomSurf: Surface Representation for Learning on Protein Structures (ICLR 2025 · similarity 0.86)
- Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer (ICLR 2025 · similarity 0.86)
- Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.86)
- Proteus: Exploring Protein Structure Generation for Enhanced Designability and Efficiency (ICML 2024 · 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.