Research map / Image generation / Autoregressive and latent generation
Proteina: Scaling Flow-based Protein Structure Generative Models
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.
Proteina scales non-equivariant flow-matching transformers (up to ~400M params, 21M synthetic structures) for protein backbone generation with hierarchical CATH fold conditioning, CFG/autoguidance, LoRA fine-tuning, and new distributional metrics.
Problem
Protein backbone generative models have been small, equivariant, trained on at most ~0.6M structures, lack semantic control, and are evaluated with per-sample metrics that do not measure distributional fit.
Approach
Scale a non-equivariant transformer flow-matching model to large synthetic AFDB data, condition hierarchically on CATH fold labels with classifier-free guidance, and introduce FID-like distribution metrics (FPSD, fJSD, fS) computed with a fold-class predictor.
Key results
- M_21M reaches 99.0% designability in unconditional generation (noise scale 0.3), the highest among compared models, while the FS model yields higher diversity.
- On the RFDiffusion motif-scaffolding benchmark Proteina gets 2094 unique successes vs 1445 for Genie2 (second best).
- Classifier-free guidance on the fold-class-conditional model raises designability to 91.4% (omega=1.0) and improves beta-sheet content, e.g. mainly-beta class design ability 90.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase noise scale from γ=0.35 to γ=0.5 (ablation) | Proteina M_FS · Unconditional backbone generation | diversity (Cluster): 0.49 (239) -> 0.71 (323) | Helped |
| Increase noise scale from γ=0.35 to γ=0.5 (ablation) | Proteina M_FS · Unconditional backbone generation | designability (%): 98.2 -> 91.4 | Hurt |
| Train on the larger, strongly filtered D_21M dataset instead of D_FS | Proteina · Unconditional backbone generation | designability (%): 98.2 -> 99.0 | Helped |
| Remove triangle layers and pair representation updates (ablation) | Proteina M_FS · Unconditional backbone generation | designability (%): 96.4 -> 93.8 | Hurt |
| Remove triangle layers and pair representation updates (ablation) | Proteina M_FS · Unconditional backbone generation | diversity (Cluster): 0.63 (305) -> 0.62 (292) | No clear effect |
| Fine-tune with LoRA on designable PDB proteins (ablation) | Proteina M_FS · Unconditional backbone generation | designability (%): 96.4 -> 96.6 | Helped |
| Fine-tune with LoRA on designable PDB proteins (ablation) | Proteina M_FS · Unconditional backbone generation | FPSD vs. PDB: 388.0 -> 274.1 | Helped |
| Fine-tune with LoRA on designable PDB proteins (ablation) | Proteina M_FS · Unconditional backbone generation | diversity (Cluster): 0.63 (305) -> 0.43 (208) | Hurt |
| Increase CFG guidance weight from ω=1.0 to ω=2.0 (ablation) | Proteina M_cond_FS · Fold class-conditional backbone generation | designability (%): 91.4 -> 83.8 | Hurt |
| Increase CFG guidance weight from ω=1.0 to ω=2.0 (ablation) | Proteina M_cond_FS · Fold class-conditional backbone generation | FPSD vs. PDB: 121.1 -> 103.0 | Helped |
| Add hierarchical fold-class conditioning with CFG | Proteina · Fold class-conditional backbone generation | designability (%): 57.0 -> 91.4 | Helped |
| Increase CFG guidance weight (ablation) | Proteina M_cond_21M · Fold class re-classification, C-level classes | correct re-classification probability: ω=0.0: 0.585 / 0.017 / 0.450; ω=2.5: 0.993 / 0.979 / 0.997 | Helped |
Surprising findings
- Non-equivariant transformer trained without equivariant architecture matches or exceeds fully equivariant baselines on protein backbone generation, with models learning approximate SO(3) equivariance from centering and rotation augmentation.
- Chroma performs competitively on standard designability/diversity/novelty metrics but fails at fold-class-specific generation when checked by re-classification.
Limitations
- Designability relies on auxiliary ProteinMPNN and ESMFold, which have their own biases.
- Main models use noise-scale reduction at inference to boost designability, trading off diversity.
- Rare fold classes are reliably generated worse than common ones (re-classification probabilities near zero for rare T classes).
Experimental setup
Datasets: AFDB Foldseek clusters (D_FS) AFDB high-quality filtered subset (D_21M) TED domain labels (CATH) PDB designable proteins (LoRA fine-tuning) RFDiffusion motif-scaffolding benchmark
Benchmarks: unconditional backbone generation designability/diversity/novelty FPSD fJSD fS fold-class conditional generation RFDiffusion motif-scaffolding benchmark long chain generation up to 800 residues
Metrics: designability (%) diversity (cluster, TM-score) novelty vs PDB/AFDB FPSD fJSD fold score secondary structure % unique motif-scaffolding successes sampling time max batch size
Baselines: FrameDiff FoldFlow FrameFlow ESM3 Chroma RFDiffusion Proteus Genie2
Builds on: rectified flow classifier-free guidance FrameDiff RFDiffusion Genie2
Closest papers by meaning
- Learning conformational ensembles of proteins based on backbone geometry (NeurIPS 2025 · similarity 0.91)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.90)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.89)
- Generating Highly Designable Proteins with Geometric Algebra Flow Matching (NeurIPS 2024 · similarity 0.89)
- NMA-tune: Generating Highly Designable and Dynamics Aware Protein Backbones (ICML 2025 · similarity 0.89)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.89)
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.