Research map / Image generation / Autoregressive and latent generation
ProtPainter: Draw or Drag Protein via Topology-guided Diffusion
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 ProtPainter, a DDPM-based protein backbone generator conditioned on 3D curves via a CurveEncoder-predicted sketch and Helix-Gating fusion schedule, evaluated on a new topology-conditioned restoration benchmark with scTF.
Problem
Existing protein backbone diffusion models lack flexible, precise topology control beyond linear blueprint/SSE parametrizations, limiting backbone space navigation and topology editing.
Approach
Use 3D curves as topology conditions: a CurveEncoder (EGNN + curvature CNN + attention) predicts SSE labels for curves to build a naive sketch, which guides RoseTTAFold-based DDPM sampling through a low-pass-filtered fusion term, with Helix-Gating scheduling the guidance strength by comparing helix percentages.
Key results
- On the GPCR set ProtPainter reaches scTF 0.800 and fit designability FD 0.792 versus 0.262 scTF and 0.050 FD for RFDiffusion on med/GPCR-style comparisons, with the highest scTF on all three datasets among the compared methods.
- On HHH_ems, ProtPainter's FD of 0.654 exceeds RFDiffusion's 0.640 and TopoDiff's 0.160.
- Without refolding, 61.5% of CATH Mainly Alpha (class 1) topologies reach scTF > 0.7 and 39.4% reach scTF > 0.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Condition backbone generation on 3D curves | ProtPainter · HHH_ems | scTF: 0.693 -> 0.767 | Helped |
| Condition backbone generation on 3D curves | ProtPainter · med | FD: 0.050 -> 0.734 | Helped |
| Condition backbone generation on 3D curves | ProtPainter · GPCR | FD: 0.000 -> 0.792 | Helped |
| Add Helix-Guiding (ablation) | ProtPainter · GPCR | fit-designability: 0.656 -> 0.676 | Helped |
| Add Helix-Guiding (ablation) | ProtPainter · HHH_ems | fit-designability: 0.53 -> 0.56 | Helped |
| Add Helix-Gating (ablation) | ProtPainter · GPCR | fit-designability: 0.588 -> 0.676 | Helped |
| Add Helix-Gating (ablation) | ProtPainter · HHH_ems | fit-designability: 0.48 -> 0.56 | Helped |
| Add Two-Phase transition (ablation) | ProtPainter · GPCR | fit-designability: 0.668 -> 0.676 | Helped |
| Add Two-Phase transition (ablation) | ProtPainter · HHH_ems | fit-designability: 0.392 -> 0.56 | Helped |
| Add CurveEncoder (ablation) | ProtPainter · GPCR | fit-designability: 0.112 -> 0.676 | Helped |
| Add CurveEncoder (ablation) | ProtPainter · HHH_ems | fit-designability: 0.196 -> 0.56 | Helped |
| Set γ=0.2 and η=0.7 instead of γ=0.2 and η=0.8 (ablation) | ProtPainter · HHH_ems | average: 0.754 -> 0.782 | Helped |
Surprising findings
- Adding up to 1 Å-scale noise to curves slightly improves CD on the med set (0.924 at noise level 1), and fit designability drops sharply at noise levels of 2 Å and above.
- Point-cloud Chroma conditioning yields near-zero fit designability on GPCR, showing vague spatial conditioning fails for topology control.
Limitations
- Performance drops sharply for complex topologies such as DNA polymerase (CATH 1.10.150, scTF>0.8 only 0.076).
- Beta-sheet topologies are not yet handled; only helix-based control is demonstrated.
- Generation and refolding are time-consuming, which limits real-time design.
Experimental setup
Datasets: PDB CATH HHH_ems 1a0b_cluster GPCR
Benchmarks: Protein Restoration Task (HHH_ems, med/1a0b_cluster, GPCR) CATH topology families
Metrics: scTM scTF scRMSD pLDDT CD FD maxPDBTM
Baselines: RFDiffusion Chroma TopoDiff
Builds on: RFDiffusion DDPM Chroma FrameDiff TopoDiff
Closest papers by meaning
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.88)
- Proteus: Exploring Protein Structure Generation for Enhanced Designability and Efficiency (ICML 2024 · similarity 0.88)
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.88)
- ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids (ICLR 2025 · similarity 0.88)
- ReQFlow: Rectified Quaternion Flow for Efficient and High-Quality Protein Backbone Generation (ICML 2025 · similarity 0.87)
- Learning conformational ensembles of proteins based on backbone geometry (NeurIPS 2025 · similarity 0.87)
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.