Loading…
Macrocyclic peptide binder design against protein targets, cyclizing a diffusion backbone generator's positional encoding so it closes rings.
No providers recorded yet. Browse all providers
Macrocyclic peptides occupy the size range between small molecules and biologics, which is exactly where many undruggable targets live — flat interfaces with no deep hydrophobic pocket, and intracellular proteins a full-size binder cannot reach. The standard route to a macrocyclic binder is to screen a display library of trillions of random peptides. That works, but it says nothing about how the winner binds, and determining the binding mode afterwards is often the bottleneck.
RFpeptides removes the screen. It is a denoising-diffusion pipeline from the Institute for Protein Design at the University of Washington — the Bhardwaj, DiMaio and Baker groups, with collaborators in Ireland, Germany and Massachusetts — that designs head-to-tail cyclic peptide binders directly against a protein target. The insight is small and structural. A backbone generator encodes the relationship between residues i and j by their linear separation along the chain, which has no way to express that the N and C termini are bonded to each other. RFpeptides cyclizes that encoding: once j is more than halfway around the peptide from i, the relative offset flips from positive to negative, so the network sees a closed ring rather than a linear fragment. Only the designed chain is cyclized; the target chain and the binder–target index pairs keep standard encodings.
Because the change lives in the positional encoding rather than the weights, no retraining was needed. RFpeptides runs the released RFdiffusion checkpoints unchanged behind two inference flags, and the protocol now ships inside the RFdiffusion repository.
The full pipeline chains four stages: diffusion generates cyclic backbones against the target; ProteinMPNN with iterative Rosetta Relax designs four sequences per backbone; AfCycDesign — a cyclic-aware modification of AlphaFold2 — re-predicts each complex from sequence, with RoseTTAFold2 as an orthogonal second check; and Rosetta interface metrics (ddG, contact molecular surface, spatial aggregation propensity) filter what survives. Diffusion runs 50 timesteps. Generation is genuinely diverse: 48,000 sampled backbones at 10 and 12 residues yielded 9,045 and 8,913 structurally unique structures respectively, with self-consistency measured by re-predicting LigandMPNN sequences.
Four targets were designed against and tested. Campaigns generated 9,965 to 20,000 backbones per target, expanding to 40,000–80,000 design models, from which no more than 27 candidates per target were selected for synthesis, and 20 or fewer per target were actually tested. Against MCL1 the best macrocycle bound at 2 µM and against MDM2 at 1.9 µM; against GABARAP two designs bound at 6 nM and 36 nM with AlphaScreen IC50 values of 0.7 nM and 2.5 nM; against Acinetobacter baumannii Rhombotarget A, designed from a predicted structure, the best binder reached 9.4 nM. Five X-ray structures were deposited, and the designed macrocycles matched their design models to Cα RMSDs of 0.7 to 1.7 Å.
RFpeptides suits programs that need a binder to a specific, chosen epitope rather than whatever a library happens to yield — peptide therapeutics against flat or shallow interfaces, chimeric degraders built on autophagy adaptors such as GABARAP, diagnostic reagents, and antibacterial leads against pathogen surface proteins with no crystal structure. Because designs arrive with a modeled interface, they slot directly into structure-guided medicinal chemistry.
RFpeptides demonstrates that a targeted change to positional encoding can open an entire molecule class to a frozen generative backbone model, without new weights or new training data — a template other RoseTTAFold-based networks can follow. It sits alongside other computational routes to cyclic binders in this catalog, including CycleDesigner, the tree-search approach of HighPlay, and the all-atom chemistry of Vilya-1. The costs are real: campaigns burn tens of thousands of design models per target, hit rates run at a few binders per handful of peptides tested, and synthesis attrition is significant — 13 of 27 selected MCL1 designs never cyclized in usable yield. The method is also limited to head-to-tail cyclization of canonical amino acids; non-canonical residues, crosslinkers and alternative cyclization chemistries remain future work.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.