Research map / Probabilistic and scientific modeling / Molecules and proteins
Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension
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 PepHAR, a hot-spot-driven autoregressive peptide generator that samples hot spots with an energy-based model, extends fragments via von Mises dihedral distributions, and refines assembly with a correction stage.
Problem
De novo peptide binder design struggles with unequal residue contributions to binding, the geometric constraints of peptide bonds, and a lack of realistic benchmarks such as scaffolding known hot spot residues.
Approach
Generate a few hot-spot residues first from an energy-based density model (NCE-trained, Langevin sampling), then autoregressively extend fragments from them by sampling dihedral angles from von Mises distributions so peptide-bond geometry holds, and finally refine the full peptide with gradient-based backbone and dihedral objectives.
Key results
- In peptide design, PepHAR (K=3) reaches RMSD 2.68 Å and SSR 84.91%, versus 3.83 Å and 80.24% for PepGLAD, but its Success rate (25.54%) is below PepFlow (27.96%).
- In scaffold generation, PepHAR (K=3) achieves the best RMSD (2.15 Å), SSR (83.02%) and BSR (88.02%) among compared methods.
- Ablation: removing hot-spot sampling drops BSR from 86.74% to 74.17% and Stability from 16.62% to 11.23%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use von Mises angle distributions instead of direct angle predictions (ablation) | PepHAR · Peptide Binder Design | RMSD: 3.10 -> 2.68 | Helped |
| Replace randomly positioned and typed residues with sampled hot spots (ablation) | PepHAR · Peptide Binder Design | BSR: 74.17 -> 86.74 | Helped |
| Add the correction stage (ablation) | PepHAR · Peptide Binder Design | Affinity: 14.85 -> 20.53 | Helped |
| Increase the number of hot spots from K=1 to K=3 (ablation) | PepHAR · Peptide Binder Design | RMSD: 3.73 -> 2.68 | Helped |
| Increase the number of hot spots from K=1 to K=3 (ablation) | PepHAR · Peptide Binder Design | Novelty: 81.21 -> 79.11 | Hurt |
| Increase the number of hot spots from K=1 to K=3 (ablation) | PepHAR · Peptide Scaffold Generation | RMSD: 3.72 -> 2.15 | Helped |
| Increase the number of hot spots from K=1 to K=3 (ablation) | PepHAR · Peptide Scaffold Generation | Novelty: 80.61 -> 72.56 | Hurt |
Surprising findings
- PepHAR's Success rate in peptide design (25.54%) is lower than PepFlow (27.96%) and RFDiffusion (25.38% roughly equal), and its stability is below RFDiffusion.
- Increasing the number of hot spots improves geometry and energy but reduces novelty and diversity.
Limitations
- Valid rate (59.31%) and Stability/Success metrics lag some baselines such as RFDiffusion and PepFlow.
- Hot spots for scaffold generation are chosen manually from Rosetta energies, which limits the scalability of that setting.
Experimental setup
Datasets: PepBDB Q-BioLip
Benchmarks: peptide binder design benchmark (158 test complexes from PepBDB and Q-BioLip) peptide scaffold generation benchmark
Metrics: Valid RMSD SSR BSR Stability Affinity Novelty Diversity Success (ipTM > 0.6)
Baselines: RFDiffusion ProteinGenerator PepFlow PepGLAD
Builds on: Invariant Point Attention noise contrastive estimation Rosetta energy function
Closest papers by meaning
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.88)
- Full-Atom Peptide Design based on Multi-modal Flow Matching (ICML 2024 · similarity 0.87)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.85)
- PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion (ICML 2025 · similarity 0.85)
- Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints (ICML 2025 · similarity 0.84)
- Designing Cyclic Peptides via Harmonic SDE with Atom-Bond Modeling (ICML 2025 · similarity 0.84)
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.