Research map / Probabilistic and scientific modeling / Molecules and proteins
Fold2Seq: A Joint Sequence(1D)-Fold(3D) Embedding-based Generative Model for Protein Design
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 |
|---|---|---|---|
| Use Fold2Seq instead of Graph_trans | Protein sequence design · ID Test | fold-level per-residue perplexity: 9.6 (2.9) -> 9.0 (5.3) | Helped |
| Use Fold2Seq instead of Graph_trans | Protein sequence design · OD Test | fold-level per-residue perplexity: 11.5 (3.3) -> 12.0 (2.4) | Hurt |
| Use Fold2Seq instead of Graph_trans | Protein sequence design · ID and OD Test | fold-level sequence recovery: 28.8 (11.3) -> 27.2 (6.3); 27.1 (4.0) -> 25.2 (3.2) | Mixed |
| Use Fold2Seq instead of Graph_trans | Protein sequence design · ID and OD Test coverage subsets | fold-level coverage: Fold2Seq provides better coverage in 68%-88% folds | Helped |
| Replace 1D string fold representation with 3D voxel representation (ablation) | Fold2Seq · ID Test | fold-level sequence recovery: 20.0 (8.31) -> 22.5 (7.34) | Helped |
| Replace 1D string fold representation with 3D voxel representation (ablation) | Fold2Seq · OD Test | fold-level sequence recovery: 19.2 (3.45) -> 21.3 (3.33) | Helped |
| Add sequence encoder, sequence reconstruction loss, and cosine similarity loss (ablation) | Fold2Seq · ID and OD Test | fold-level sequence recovery: 22.5 (7.34) -> 22.8 (8.01); 21.3 (3.33) -> 21.9 (2.34) | Helped |
| Add two fold-classification losses (ablation) | Fold2Seq · ID and OD Test | fold-level sequence recovery: 22.8 (8.01) -> 25.6 (6.34); 21.9 (2.34) -> 23.7 (2.34) | Helped |
| Add cyclic loss (ablation) | Fold2Seq · ID and OD Test | fold-level sequence recovery: 25.6 (6.34) -> 27.2 (6.3); 23.7 (2.34) -> 25.2 (3.2) | Helped |
| Train all loss terms together rather than in two stages (ablation) | Fold2Seq · ID and OD Test | fold-level sequence recovery: 22.2 (4.3) -> 27.2 (6.3); 20.3 (3.2) -> 25.2 (3.2) | Hurt |
| Use principal-axis orientations or denser 45-degree augmentation (ablation) | Fold2Seq · OD Test | fold-level per-residue perplexity: 12.2 (2.7); 12.0 (2.5) | No clear effect |
| Average voxel features across NMR ensemble structures rather than use a single structure (ablation) | Fold2Seq · NMR ID and OD Test | fold-level sequence recovery: 24.1 (3.9) -> 25.2 (3.5); 22.2 (3.8) -> 24.1 (4.2) | Helped |
Closest papers by meaning
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.90)
- FoldToken: Learning Protein Language via Vector Quantization and Beyond (AAAI 2025 · similarity 0.89)
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.89)
- ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search (NeurIPS 2025 · similarity 0.89)
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.88)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · 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.