Research map / Probabilistic and scientific modeling / Molecules and proteins
Learning inverse folding from millions of predicted structures
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add 12M AlphaFold2-predicted structures to training (mixed 1:80 with CATH) | GVP-GNN-large (21M) · CATH 4.3 topology split test | sequence recovery %: 39.2% to 50.8% | Helped |
| add 12M AlphaFold2-predicted structures to training | GVP-Transformer · CATH 4.3 topology split test | sequence recovery %: 38.3% to 51.6% | Helped |
| add 12M AlphaFold2-predicted structures to training | GVP-GNN (1M) · CATH 4.3 topology split test | sequence recovery %: 42.2% to 38.6% | Hurt |
| scale GVP-GNN from 1M to 21M parameters on CATH only (ablation) | GVP-GNN · CATH 4.3 topology split test | sequence recovery %: 42.2% to 39.2% | Hurt |
| use GVP-Transformer (Transformer decoder over GVP encoder) instead of GVP-GNN-large when training with predicted structures (ablation) | predicted-structure training · CATH 4.3 topology split test | sequence recovery %: 50.8% to 51.6% | Helped |
| add Gaussian noise (0.1 Å) to predicted training structures (ablation) | GVP-Transformer · CATH 4.3 topology split test | sequence recovery | Helped |
| span masking of backbone coordinates instead of random masking during training (ablation) | GVP-Transformer · Masked backbone perplexity (Figure 4) | perplexity on masked regions | Helped |
| mix predicted structures at higher ratio during training (ablation) | GVP-GNN-large and GVP-Transformer · CATH 4.3 topology split test | sequence recovery | Helped |
| add predicted structures to training (ablation) | GVP-GNN-large and GVP-Transformer on complex stability (SKEMPI) · Atom3D SKEMPI complex stability | zero-shot classification performance | No clear effect |
| add ACE2 coordinates as input for zero-shot binding affinity prediction (ablation) | GVP-GNN-large+AF2 · ACE2-RBD mutational scan (Starr et al.) | Spearman correlation: 0.52 to 0.69 | Helped |
| adding 12M predicted structures to training improves stability predictions for mini-proteins | GVP-Transformer and GVP-GNN-large · Rocklin de novo mini-protein stability | stability prediction correlation | Helped |
| multi-state design conditioned on two conformations vs single-state | GVP-Transformer · PDBFlex multi-conformation | sequence perplexity at flexible residues | Helped |
Closest papers by meaning
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.90)
- Fast Uncovering of Protein Sequence Diversity from Structure (ICLR 2025 · similarity 0.90)
- ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search (NeurIPS 2025 · similarity 0.89)
- Protein Inverse Folding From Structure Feedback (NeurIPS 2025 · similarity 0.89)
- AlphaFold Database Debiasing for Robust Inverse Folding (NeurIPS 2025 · similarity 0.89)
- FoldToken: Learning Protein Language via Vector Quantization and Beyond (AAAI 2025 · similarity 0.88)
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.