Research map / Probabilistic and scientific modeling / Molecules and proteins
PPFLOW: Target-Aware Peptide Design with Torsional 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.
Proposes PPFlow, a target-aware peptide generator using conditional flow matching on torus (torsion angles), SE(3) and multinomial amino-acid spaces, plus a new 15.6k-pair protein-peptide dataset PPBench2024.
Problem
Structure-based design of therapeutic peptides binding specific protein targets, which existing small-molecule and protein SBDD methods handle poorly, and a lack of large peptide-protein complex datasets.
Approach
Model peptide backbones via torsion angles (phi, psi, omega) on a hypertorus with a conditional torus flow matching, combined with SE(3) flow matching for global pose and multinomial flow matching for sequence, all conditioned on the receptor through an attention-based network (LoCS/ITA) with equivariance guarantees.
Key results
- PPFlow reports the highest IMP%-S (most stable designed complexes) and the best ΔG/IMP%-B on PPDBench peptide generation versus DiffPP, per the authors.
- On flexible re-docking, DiffPP performs best on L-RMSD and Success%, while PPFlow models ligand centroid most accurately; classical docking is dozens of times slower.
- RDE-PP with pretraining outperforms Rosetta and unpretrained RDE-PP on three of four torsional angles in side-chain packing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PPFlow for peptide generation instead of the baselines | PPFlow · PPDBench peptide generation | ΔG | Helped |
| Use PPFlow for peptide generation instead of the baselines | PPFlow · PPDBench peptide generation | IMP%-B | Helped |
| Use PPFlow for peptide generation instead of the baselines | PPFlow · PPDBench peptide generation | IMP%-S | Helped |
| Use PPFlow rather than DiffBP-PP for peptide generation | PPFlow · PPDBench peptide generation | Validity | Helped |
| Use PPFlow rather than DiffPP for peptide generation | PPFlow · PPDBench peptide generation | Diversity | Helped |
| Use RDE-predicted side chains instead of Rosetta-packed side chains (ablation) | PPFlow-FA vs PPFlow-BB · PPDBench peptide generation | IMP%-S | Helped |
| Use DiffPP rather than PPFlow for flexible re-docking | DiffPP · PPDBench flexible re-docking | L-RMSD | Helped |
| Use DiffPP rather than PPFlow for flexible re-docking | DiffPP · PPDBench flexible re-docking | Success% | Helped |
| Use PPFlow rather than DiffPP for flexible re-docking | PPFlow · PPDBench flexible re-docking | C-RMSD | Helped |
| Pretrain RDE-PP before fine-tuning on protein-peptide complexes (ablation) | RDE-PP · Side-chain packing | MAE | Helped |
| Pretrain RDE-PP before fine-tuning on protein-peptide complexes (ablation) | RDE-PP · Side-chain packing | NLL | Helped |
| Increase the number of optimization steps (ablation) | PPFlow · Peptide optimization | binding affinity | No clear effect |
Surprising findings
- Increasing optimization steps contributes little to binding affinity improvement, though PPFlow still yields more high-affinity peptides.
- DiffBP-PP (atom-level diffusion) shows the lowest validity due to its high degrees of freedom, so it was dropped from later comparisons.
Limitations
- Docking performance is not competitive with baselines such as DiffPP and classical docking.
- Side-chain packing gains over classical Rosetta are small.
- Evaluation relies on generated-peptide metrics computed by ADCP and FoldX rather than experimental validation.
Experimental setup
Datasets: PPBench2024 PropediaV2.3 PepBDB PPDBench PDB-REDO RCSB PDB
Benchmarks: PPDBench
Metrics: binding energy ΔG (ADCP) IMP%-B stability ΔG (FoldX) IMP%-S validity novelty (TM-score, SeqOL) diversity L-RMSD C-RMSD Success%
Baselines: DiffPP DiffBP-PP Rosetta RDE-PP (w/o pretraining) HDock VinaDock
Builds on: Riemannian flow matching AlphaFold2 NeRF DiffAB TargetDiff
Closest papers by meaning
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number (NeurIPS 2025 · similarity 0.89)
- Full-Atom Peptide Design based on Multi-modal Flow Matching (ICML 2024 · similarity 0.89)
- Multi-domain Distribution Learning for De Novo Drug Design (ICLR 2025 · similarity 0.87)
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.86)
- Harmonic Self-Conditioned Flow Matching for joint Multi-Ligand Docking and Binding Site Design (ICML 2024 · similarity 0.86)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 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.