Research map / Image generation / Autoregressive and latent generation
Proteus: Exploring Protein Structure Generation for Enhanced Designability and Efficiency
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.
Proteus trains a protein backbone diffusion model from scratch (no structure-predictor pretraining) using a graph-based triangle block and multi-track folding blocks, reaching higher designability and speed than RFdiffusion, FrameDiff, Genie and Chroma on monomer and oligomer generation.
Problem
Protein backbone diffusion models that avoid pretraining on large structure predictors (e.g., FrameDiff, Genie) lag far behind RFdiffusion in designability, while efficient pretraining-free generation remains hard to scale to long proteins.
Approach
Replaces full triangle attention (O(N^3)) with a KNN-graph triangle attention of O(NK^2) plus structure-biased, sequence-gated edge updates inside an IPA-based SE(3) diffusion network, and augments training data with single chains split from oligomers.
Key results
- Proteus reaches 0.921 designability vs 0.705 for RFDiffusion on monomer generation with 100 sampling steps and 19.8M parameters.
- Success efficiency at length 400-600 is 23.09 s and 50.62 s for Proteus versus 262.64 s and 1494.28 s for RFDiffusion.
- Of 16 experimentally ordered designs, 9 of 300-residue and 3 of 500-residue designs showed monodisperse SEC peaks and CD-confirmed secondary structure.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Proteus instead of RFDiffusion | unconditional monomer generation · Lengths 100–600 | Designability: 0.705 -> 0.921 | Helped |
| Use Proteus instead of RFDiffusion | unconditional monomer generation · Lengths 100–600 | Sample time(s): 120.24 -> 18.20 | Helped |
| Add single-chain training data derived from oligomers (ablation) | Proteus · Unconditional monomer generation | Success rate: 36.9% -> 59.2% | Helped |
| Add rotation-based pair-wise self-condition features (ablation) | Proteus · Unconditional monomer generation | Success rate: 59.2% -> 63.2% | Helped |
| Add structure featurizer (ablation) | Proteus · Unconditional monomer generation | Success rate: 63.2% -> 87.5% | Helped |
| Add graph triangle attention (ablation) | Proteus · Unconditional monomer generation | Success rate: 87.5% -> 92.1% | Helped |
| Add structure bias to triangle attention (ablation) | Proteus · Unconditional monomer generation | Designability | Helped |
| Increase sampling noise level ζ from 0.1 to 0.5 (ablation) | Proteus · Sampling parameter benchmark | < 2 ˚A scRMSD (↓): 92.0% -> 87.8% | Hurt |
| Increase sampling noise level ζ from 0.1 to 1.0 (ablation) | Proteus · Sampling parameter benchmark | < 2 ˚A scRMSD (↓): 92.0% -> 37.3% | Hurt |
| Reduce sampling steps from 100 to 50 (ablation) | Proteus · Sampling parameter benchmark | < 2 ˚A scRMSD (↓): 92.0% -> 89.8% | Hurt |
| Increase sampling steps from 100 to 200 (ablation) | Proteus · Sampling parameter benchmark | < 2 ˚A scRMSD (↓): 92.0% -> 91.0% | Hurt |
| Increase sampling steps from 100 to 500 (ablation) | Proteus · Sampling parameter benchmark | < 2 ˚A scRMSD (↓): 92.0% -> 92.2% | Helped |
Surprising findings
- Adding oligomer-derived single chains to the monomer dataset substantially raised designability (ablation success rate 36.9% to 59.2%).
- Proteus reached 92.0% success at 100 steps, similar to 200 and 500 steps, so extra sampling steps gave little gain.
Limitations
- Designability still measured in silico with ProteinMPNN/ESMFold; experimental validation covers only 16 designs.
- Chroma becomes more time-efficient than Proteus for lengths above 400.
- Source code to be released only after acceptance.
Experimental setup
Datasets: PDB (curated 50,773 single chains) UniProt IDs for deduplication DSSP for loop filtering
Benchmarks: unconditional monomer generation (lengths 100-600) oligomer complex generation (dimer, trimer, tetramer) experimental validation of 16 designs (SEC, CD)
Metrics: designability (scRMSD < 2 Å / scTM) sample time success efficiency diversity (TM-score clustering) success rate
Baselines: RFDiffusion Genie FrameDiff Chroma
Builds on: FrameDiff AlphaFold2 triangle attention Invariant Point Attention Chroma RFdiffusion
Closest papers by meaning
- ProtPainter: Draw or Drag Protein via Topology-guided Diffusion (ICLR 2025 · similarity 0.88)
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.88)
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.86)
- ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids (ICLR 2025 · similarity 0.86)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.86)
- ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search (NeurIPS 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.