Loading…
De novo cyclic peptide binder design against a protein target, chaining a cyclized diffusion sampler, sequence design and structure prediction.
No providers recorded yet. Browse all providers
Cyclic peptides are attractive binders for protein-protein interfaces: the closed backbone resists proteolysis and pays a smaller entropic penalty on binding than a linear peptide of the same length. Designing them de novo has lagged behind linear protein design, because the generative models that made binder design routine assume an open chain. RFdiffusion represents residue relationships with a linear relative-position encoding, so a head-to-tail macrocycle is outside what it can express, and there are far too few solved cyclic peptide complexes to retrain it on.
CycleDesigner is a workflow that gets cyclic backbones out of RFdiffusion anyway. Chenhao Zhang, Hongliang Duan and colleagues at Macao Polytechnic University, Zhejiang University of Technology and Shenzhen Highslab Therapeutics replace the model's positional encoding with one that describes a circle, leaving the pretrained weights untouched. The modified sampler — CycRFdiffusion — generates a cyclic backbone against a specified target and hotspot set; ProteinMPNN designs sequences onto it; HighFold, the same group's cyclic-peptide adaptation of AlphaFold, predicts the resulting complex; and Rosetta energy terms screen what survives. The work appeared on bioRxiv in December 2024 and in the Journal of Chemical Information and Modeling in 2025.
It sits alongside the group's structure-prediction line, HighFold and HighFold2, and their reinforcement-learning binder search HighPlay, as the diffusion-based route to the same goal.
Targets were drawn from HighFold's cyclic peptide complex dataset in the PDB, restricted to single-chain proteins to match RFdiffusion's input constraints, with missing residues repaired by PDBfixer and non-protein entities stripped. The cyclic position matrix is built from residue indices and replaces the default encoding in RFdiffusion's input features; everything downstream is stock. HighFold predicts the peptide-target complex using its cyclic position offset encoding matrix. Runs were containerized with Docker on two RTX 3090 GPUs.
The number of diffusion steps behaves unusually here. Sweeping diffuser.T over 20, 25, 30 and 35 gave mean RMSD to the generated backbone of 2.151, 1.758, 1.769 and 2.212 Å and mean ipTM of 0.527, 0.541, 0.529 and 0.511 — accuracy peaks at 25 steps and degrades beyond it, which the authors attribute to RFdiffusion's bias toward α-helical secondary structure asserting itself with longer sampling, whereas macrocycles are loop-dominated. In the main run, 23 targets each received five backbones (115 total), each backbone five ProteinMPNN sequences (575 sequences), and each sequence a HighFold prediction, yielding 2,875 complex structures with mean pLDDT 91.26, pTM 0.912, ipTM 0.533 and iPAE 0.465. Filtering at iPAE < 0.35, ipTM > 0.5, pLDDT > 80 and RMSD < 3.5 Å left 305 designs, whose means improved to pLDDT 95.15, ipTM 0.778 and iPAE 0.274. Rosetta screening on dGseparated/dSASA × 100 selected 245 structures, of which 74 scored below −1.5. Designs adopted binding poses similar to the native ligands while differing substantially in sequence and conformation.
The workflow targets computational chemistry groups pursuing macrocyclic binders for interfaces that small molecules cannot reach. Given a target structure and a known binding site, it produces cyclic backbones, sequences and predicted complexes with confidence and interface-energy scores attached, narrowing a synthesis shortlist. Using it means installing stock RFdiffusion first and copying CycleDesigner's sources over it; the repository ships source patches rather than a package, distributes no standalone weights, and carries no license file. Molecular dynamics trajectories from the study are archived on Zenodo.
CycleDesigner is a clean demonstration that a capability a generative model was never trained for can be unlocked by editing the representation it reasons over rather than its weights — the same manoeuvre HighFold applies to AlphaFold's positional encoding, transplanted to the diffusion side. The diffusion-step finding is a useful negative result for anyone adapting protein backbone samplers to peptides, since more sampling is normally assumed to be better. The evaluation is entirely computational: every metric reported is a predicted structure score or a Rosetta energy, with no synthesized peptides and no measured binding affinities, so the designs remain candidates rather than validated binders.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.