Research map / Image generation / Autoregressive and latent generation
Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints
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 CP-Composer, a geometric-conditioned latent diffusion model that composes type and distance unit constraints with classifier-free-style guidance to generate target-binding cyclic peptides zero-shot from linear-peptide training data.
Problem
Designing target-specific cyclic peptides is underexplored because training data for cyclic peptides is scarce, and existing methods rely on post-filtering or hard-coded cyclization.
Approach
Decompose cyclization strategies into unit geometric constraints (node-level type constraints and edge-level distance constraints), encode them as vector embeddings (one-hot and RBF) in a latent diffusion denoiser, train jointly with random constraint masking on linear peptides, and at inference compose unseen constraint combinations via classifier-free-style score extrapolation.
Key results
- CP-Composer reaches success rates of 38.57% (stapled), 74.42% (head-to-tail), 82.50% (disulfide) and 84.62% (bicycle) at guidance weight w=5.0, versus 0-30% for PepGLAD and energy-guidance baselines.
- On disulfide peptides CP-Composer (41.25%) outperforms DiffPepBuilder (23.07%) at matched settings.
- MM/PBSA on two targets gives cyclic peptides lower binding energies than native linear binders (-10.66 vs -9.73 and -20.41 vs -15.17 kcal/mol) with lower RMSD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CP-Composer instead of PepGLAD without guidance | target-specific cyclic peptide generation · Stapled peptide | success rate: 22.80% -> 38.57% | Helped |
| Use CP-Composer instead of PepGLAD without guidance | target-specific cyclic peptide generation · Head-to-tail peptide | success rate: 30.23% -> 74.42% | Helped |
| Use composable constraint guidance instead of PepGLAD without guidance | target-specific cyclic peptide generation · Disulfide peptide | success rate: 0 -> 82.50% | Helped |
| Use composable constraint guidance instead of PepGLAD without guidance | target-specific cyclic peptide generation · Bicycle peptide | success rate: 0 -> 84.62% | Helped |
| Use CP-Composer instead of DiffPepBuilder | disulfide peptide generation · Disulfide peptide | success rate: 23.07% -> 41.25% | Helped |
| Use CP-Composer instead of DiffPepBuilder | multi-cycle peptide generation · 2*-S-S- | success rate: 32.78% -> 62.00% | Helped |
| Increase guidance weight from w=1.0 to w=5.0 (ablation) | CP-Composer · Disulfide peptide | success rate: 21.25% -> 82.50% | Helped |
| Increase guidance weight from w=5.0 to w=10.0 (ablation) | CP-Composer · Disulfide peptide | success rate: 82.50% -> 62.50% | Hurt |
| Add RBF distance features (ablation) | CP-Composer · Bicycle peptide | success rate: 26.92% -> 30.76% | Helped |
| Increase RBF channels from 16 to 32 (ablation) | CP-Composer · Bicycle peptide | success rate: 30.76% -> 30.76% | No clear effect |
| Replace linear peptide with a generated cyclic peptide | PDB 3RC4 complex · PDB: 3RC4 | RMSD (Å): 2.57±0.51 -> 1.44±0.23 | Helped |
| Replace linear peptide with a generated cyclic peptide | PDB 4J86 complex · PDB: 4J86 | RMSD (Å): 3.37±0.73 -> 1.56±0.40 | Helped |
Surprising findings
- Energy-based classifier guidance (PepGLAD w/ EG) fails completely on disulfide and bicycle peptides (0% success).
- Increasing guidance weight beyond 5.0 (w=10) lowers success rate and sharply raises KL divergence on several strategies.
Limitations
- Evaluation of MD is limited to two target proteins and one cyclic peptide per strategy due to computational cost.
- Success rates depend strongly on guidance weight w, which trades off constraint satisfaction against distribution fidelity.
- Stapled and high-order 2*Stapled combinations remain hard, with low success rates (7.5% for 2*Stapled).
Experimental setup
Datasets: PepBench ProtFrag LNR
Benchmarks: LNR PepBench
Metrics: success rate AA-KL backbone dihedral KL side-chain dihedral KL MM/PBSA binding energy RMSD
Baselines: PepGLAD PepGLAD w/ energy guidance DiffPepBuilder CADS
Builds on: PepGLAD classifier-free guidance dyMEAN DDPM energy-based classifier guidance
Closest papers by meaning
- Designing Cyclic Peptides via Harmonic SDE with Atom-Bond Modeling (ICML 2025 · similarity 0.91)
- PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion (ICML 2025 · similarity 0.86)
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.85)
- Unified all-atom molecule generation with neural fields (NeurIPS 2025 · similarity 0.85)
- Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension (ICLR 2025 · similarity 0.84)
- Full-Atom Peptide Design based on Multi-modal Flow Matching (ICML 2024 · 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.