Cyclic peptide design conditioned on target protein structure, generating all four cyclization types via all-atom, all-bond harmonic SDE modeling.
No providers recorded yet. Browse all providers
Cyclic peptides are attractive therapeutics: closing a peptide into a macrocycle makes it far more resistant to enzymatic hydrolysis than its linear counterpart and often improves both conformational stability and target affinity. Computational design of cyclic peptides has nevertheless lagged behind linear peptide design for three structural reasons — there are very few solved 3D structures of protein targets bound to cyclic peptide ligands, cyclization imposes hard geometric constraints that residue-frame representations cannot express, and the linking chemistry frequently involves non-canonical amino acids.
CpSDE, from ByteDance Seed with collaborators at the University of Chinese Academy of Sciences, CASIA, Tsinghua University, and Shanghai Jiao Tong University, addresses all three by abandoning the residue frame entirely and modeling peptides as atoms and bonds. It pairs two networks: AtomSDE, a generative structure predictor built on a harmonic stochastic differential equation, and ResRouter, a residue-type predictor that reads denoised structures. A routed sampling algorithm alternates between them, updating structure and sequence together until a complete cyclic peptide emerges in the binding pocket of a given target.
Because atoms and bonds are the shared vocabulary of small molecules, linear peptides, and macrocycles, this representation lets CpSDE learn from abundant protein–ligand complex data rather than the scarce cyclic peptide structures. Presented at ICML 2025, it is the first generative method able to design all four cyclization topologies directly against a 3D protein target.
Both components are SE(3)-equivariant message-passing networks operating over two graphs — a k-nearest-neighbour graph capturing protein–ligand interactions and the ligand's chemical bond graph. AtomSDE trains on a curated small-molecule set of 14,348 PDBBind complexes plus peptide complexes with fewer than 200 heavy atoms; ResRouter trains on the full curated peptide set of 20,033 complexes drawn from the RCSB PDB, Propedia, and PepBDB, with ligands under 30 residues. Splits are clustered at 30% receptor sequence identity. Training used AdamW at a constant learning rate of 1e-4, converging in 48 hours (AtomSDE) and 18 hours (ResRouter) on 8 NVIDIA H100 GPUs; routed sampling runs 1,000 steps, invoking ResRouter only below t = 0.5.
On 100 large-volume protein pockets, Rosetta-scored designs from the mixed-topology setting reach an average total energy of -580.67 and interface binding energy of -55.71, against -576.39 / -46.98 for ProteinGenerator, -576.16 / -47.88 for PepFlow, and -359.44 / -45.06 for PepGLAD. Structural diversity is 0.79 across all four topologies, versus 0.55 for RFdiffusion paired with ProteinMPNN, which scores the best raw energy but collapses toward a repeated α-helical motif.
CpSDE targets the macrocyclic peptide segment of drug discovery, where the goal is a protease-resistant binder against a protein surface that small molecules cannot address. Given a receptor structure and a desired cyclization chemistry, it produces candidate macrocycles across a range of lengths for downstream Rosetta relaxation, molecular dynamics triage, and synthesis. The paper illustrates this with two case studies: head-to-tail inhibitors of the oncogenic methyltransferase SMYD2 and side-to-side inhibitors of SET8, each followed by MD simulation of the designed conformational ensembles.
CpSDE establishes atom-and-bond modeling as a practical route around the data bottleneck that has held back cyclic peptide design, and it makes topology a controllable input rather than a fixed architectural assumption. The authors report that generated macrocycles can still exhibit invalid conformations — distorted bond lengths or receptor clashes — that require expensive Rosetta relaxation to repair, and note that self-consistency evaluation of the kind standard in protein design is not yet possible because no accurate cyclic peptide structure predictor or docking model exists. Validation is entirely computational; no wet-lab synthesis is reported. Neither code nor trained weights have been released.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.