Research map / Probabilistic and scientific modeling / Molecules and proteins
Full-Atom Peptide Design based on Multi-modal Flow Matching
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.
Introduces PepFlow, a multi-modal conditional flow-matching model that generates full-atom target-specific peptides by jointly flowing backbone frames (SE(3)), side-chain torsions (torus), and residue types (simplex via logits).
Problem
Designing target-specific peptide binders that jointly model backbone orientation, side-chain conformations, and sequence, which prior generative models (backbone-only diffusion) do not capture.
Approach
Applies Riemannian/conditional flow matching separately to each residue modality (Euclidean for translation, SO(3) geodesics for rotation, flat-torus geodesics for side-chain angles, logit-space linear flows for residue types), trained jointly with an IPA-based decoder that predicts clean structures (reparameterized CFM) and sampled via Euler integration; partial sampling enables fix-backbone desi
Key results
- In sequence-structure co-design, PepFlow w/Bb+Seq+Ang achieves RMSD 2.07 Å and affinity 21.37%, versus RFdiffusion's 4.17 Å and 16.53%.
- In fix-backbone sequence design, PepFlow w/Bb+Seq reaches 56.40% AAR, above ProteinMPNN's 53.28%.
- In side-chain packing, PepFlow w/Bb+Seq+Ang gets 62.79% correct rate, above DiffPack's 62.58%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add sequence modeling to backbone-only PepFlow (ablation) | PepFlow · Sequence-structure co-design | AAR (%): 50.46 -> 53.25 | Helped |
| Add sequence modeling to backbone-only PepFlow (ablation) | PepFlow · Sequence-structure co-design | RMSD (Å): 2.30 -> 2.21 | Helped |
| Add side-chain angle modeling to backbone-and-sequence PepFlow (ablation) | PepFlow · Sequence-structure co-design | Affinity (%): 19.39 -> 21.37 | Helped |
| Add side-chain angle modeling to backbone-and-sequence PepFlow (ablation) | PepFlow · Sequence-structure co-design | Diversity: 0.50 -> 0.42 | Hurt |
| Add sequence modeling to backbone-only PepFlow (ablation) | PepFlow · Sequence-structure co-design | BSR (%): 82.17 -> 85.19 | Helped |
| Add side-chain angle modeling to backbone-and-sequence PepFlow (ablation) | PepFlow · Sequence-structure co-design | BSR (%): 85.19 -> 86.89 | Helped |
| Add side-chain angle modeling to backbone-and-sequence PepFlow (ablation) | PepFlow · Fix-backbone sequence design | AAR (%): 56.40 -> 54.32 | Hurt |
| Add side-chain angle modeling to backbone-and-sequence PepFlow (ablation) | PepFlow · Fix-backbone sequence design | Designability (%): 59.72 -> 65.48 | Helped |
| Add side-chain angle modeling to backbone-and-sequence PepFlow (ablation) | PepFlow · Fix-backbone sequence design | Diversity: 23.38 -> 20.65 | Hurt |
Surprising findings
- Baselines outperform PepFlow on Stability and Designability in co-design, attributed to training on entire PDB structures.
- Adding side-chain angle modeling slightly lowers sequence recovery in fix-backbone design.
Limitations
- Limited diversity due to deterministic ODE-based sampling.
- No property-guided generation capability.
- Baselines trained on entire PDB score higher on stability and designability.
Experimental setup
Datasets: PepBDB QBioLiP
Benchmarks: PepBDB/QBioLiP peptide test split (158 complexes)
Metrics: AAR RMSD SSR BSR stability affinity designability diversity MSE correct rate
Baselines: RFdiffusion ProteinGenerator ProteinMPNN ESM-IF DiffPack AttnPacker DLPacker SCWRL4
Builds on: conditional flow matching riemannian flow matching invariant point attention RFDiffusion ProteinMPNN
Closest papers by meaning
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.91)
- PPFLOW: Target-Aware Peptide Design with Torsional Flow Matching (ICML 2024 · similarity 0.89)
- Generating Highly Designable Proteins with Geometric Algebra Flow Matching (NeurIPS 2024 · similarity 0.88)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.88)
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.88)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 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.