Research map / Probabilistic and scientific modeling / Molecules and proteins
Designing Cyclic Peptides via Harmonic SDE with Atom-Bond Modeling
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 CpSDE, a harmonic-SDE all-atom docking model (AtomSDE) plus a residue type predictor (ResRouter) combined via routed sampling to design diverse cyclic peptides conditioned on protein targets and cyclization graphs.
Problem
Designing cyclic peptides for protein targets is hard due to scarce 3D cyclic peptide-target complex data, geometric constraints from cyclization, and non-canonical amino acids in cyclization bonds; prior generative methods handle only one cyclization type or linear peptides.
Approach
Model all atoms and bonds (rather than residue frames) with a harmonic SDE whose prior and perturbation kernel are shaped by the ligand chemical graph Laplacian, and alternate structure denoising with residue-type prediction via routed sampling over an atom73 superposition representation so any cyclization type can be generated.
Key results
- Head-to-tail CpSDE cyclic peptides reach average stability -568.04 and affinity -50.86 versus PepGLAD linear -359.44 and -45.06, with diversity 0.79 matching the best baseline.
- RFDiffusion achieves the best energy (affinity avg -70.30) but low diversity (0.55), attributed to generating alpha-helices.
- In an SMYD2 case study, designed head-to-tail peptide H2T-6 had 100 ns MD peptide RMSD 3.05 Å and MM-PBSA binding free energy -24.02 kcal/mol, versus -19.00 for the ground truth linear peptide.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CpSDE instead of the linear co-design baselines | Cyclic peptide design on 100 large-volume protein pockets · 100 large-volume protein pockets — mixed cyclization | Stability (Avg.): -576.39 -> -580.67 | Helped |
| Use CpSDE instead of the linear co-design baselines | Cyclic peptide design on 100 large-volume protein pockets · 100 large-volume protein pockets — mixed cyclization | Affinity (Avg.): -47.88 -> -55.71 | Helped |
| Use CpSDE instead of RFDiffusion | Peptide design on 100 large-volume protein pockets · 100 large-volume protein pockets — comparison of linear and cyclic designs | Stability (Avg.): -633.51 vs -580.67 | Hurt |
| Use CpSDE instead of RFDiffusion | Peptide design on 100 large-volume protein pockets · 100 large-volume protein pockets — comparison of linear and cyclic designs | Affinity (Avg.): -70.30 vs -55.71 | Hurt |
| Use CpSDE instead of RFDiffusion | Peptide design on 100 large-volume protein pockets · 100 large-volume protein pockets — comparison of linear and cyclic designs | Diversity: 0.55 -> 0.79 | Helped |
| Choose head-to-tail or head-to-side cyclization instead of side-to-tail or side-to-side cyclization | CpSDE · 100 large-volume protein pockets — cyclic peptide types | Stability (Avg.): -568.04; -564.81; -547.14; -537.09 | Mixed |
| Choose head-to-tail or head-to-side cyclization instead of side-to-tail or side-to-side cyclization | CpSDE · 100 large-volume protein pockets — cyclic peptide types | Affinity (Avg.): -50.86; -49.81; -46.92; -49.73 | Mixed |
| Design head-to-tail cyclic peptide H2T-6 instead of the ground-truth linear peptide | CpSDE for SMYD2 · SMYD2 inhibitor case study | Binding free energy (MM-PBSA): -19.00 -> -24.02 kcal/mol | Helped |
| Design head-to-tail cyclic peptide H2T-6 instead of the PepFlow-generated linear peptide | CpSDE for SMYD2 · SMYD2 inhibitor case study | Peptide RMSD: 4.59 Å -> 3.05 Å | Helped |
| Design side-to-side cyclic peptide S2S-4 instead of the ground-truth linear peptide | CpSDE for SET8 · SET8 inhibitor case study | Binding free energy (MM-PBSA): -6.39 kcal/mol -> -12.48 kcal/mol | Helped |
| Design side-to-side cyclic peptide S2S-4 instead of the ground-truth linear peptide | CpSDE for SET8 · SET8 inhibitor case study | Peptide RMSD: 4.06 Å -> 2.54 Å | Helped |
Surprising findings
- Head-to-tail and head-to-side cyclic peptides outperform side-to-tail and side-to-side cyclic peptides, which the authors attribute to C-N bonds being more frequent in training data than S-S and C-S bonds.
- RFDiffusion gets the best raw energy scores among all co-design methods but has low diversity, because it tends to generate alpha-helices in a repetitive pattern.
Limitations
- Evaluation relies on Rosetta energy scores and MD on few case studies rather than experimental validation.
- Cyclic peptide training data is scarce, so the model relies on linear peptide and small-molecule data for learning cyclization bonds.
Experimental setup
Datasets: PDBBind RCSB PDB Propedia PepBDB
Benchmarks: Rosetta stability and affinity on 100 large-pocket protein targets
Metrics: Stability (Rosetta total energy, lower better) Affinity (Rosetta interface binding energy, lower better) Diversity (1 - pairwise TM-score)
Baselines: RFDiffusion ProteinGenerator PepFlow PepGLAD
Builds on: score-based SDE diffusion Protpardelle atom73 superposition PepGLAD PepFlow
Closest papers by meaning
- Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints (ICML 2025 · similarity 0.91)
- Full-Atom Peptide Design based on Multi-modal Flow Matching (ICML 2024 · similarity 0.87)
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.86)
- Unified all-atom molecule generation with neural fields (NeurIPS 2025 · similarity 0.85)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.84)
- Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension (ICLR 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.