Cyclic peptides are promising for tough targets, but traditional design is slow and computationally heavy. AlphaFold2 works well for proteins, yet applying it to cyclic peptides has two hurdles: little training data and how to encode cyclization. Rettie et al. solved this by adding a cyclic positional encoding to AlphaFold2, creating AfCycDesign. It predicts structures, redesigns sequences, and hallucinates new cyclic peptides – even producing nanomolar binders to MDM2 and Keap1. This article walks through how it works and how specialized design services can turn AI outputs into real leads.
Conformational constraint is both an advantage and a computational challenge for cyclic peptides. Traditional methods rely on extensive sampling and energy evaluation, while deep learning promises speed through pattern recognition – but cyclization encoding and limited training data must be addressed.
For the past decade, cyclic peptide design has relied heavily on Rosetta's KIC algorithm for backbone sampling and sequence design. However, this approach requires generating tens to hundreds of thousands of conformations, and a single energy landscape calculation can take up to 120 CPU hours. Even worse, for cyclic peptides with multiple disulfide bonds (e.g., cyclotides), the connectivity must be specified in advance – the method cannot handle them automatically. These limitations make large-scale screening of cyclic peptide scaffolds impractical.
AlphaFold2 can predict linear peptide structures with high accuracy, but it does not recognize cyclization. The core insight of the Rettie team was that the N- and C-termini of a cyclic peptide should be treated as adjacent residues, not as the most distant pair. They therefore modified the relative positional encoding matrix so that the sequence separation between the N- and C-termini is set to 1 (instead of N-1), and integrated this change into the ColabDesign framework, naming the new tool AfCycDesign. The team also leveraged the fact that, although the AlphaFold2 training set does not include small cyclic peptides, it contains many loop regions from larger proteins – making transfer learning feasible.
The study systematically demonstrates AfCycDesign's performance on three tasks: cyclic peptide structure prediction, sequence redesign, and de novo hallucination. X-ray crystallography confirms atomic-level accuracy.
The researchers evaluated AfCycDesign on a set of experimentally determined cyclic peptides from the PDB that were not included in the AlphaFold2 training data. Using single-sequence inputs without MSAs, the method produced highly accurate structural predictions that closely matched the NMR-derived conformations. Compared to a control without cyclic constraints, AfCycDesign demonstrated significantly improved accuracy. In addition, each prediction could be completed within minutes on a single GPU, offering much higher efficiency than traditional physics-based approaches such as Rosetta.
Structure prediction of native cyclic peptides using AfCycDesign1,4
Starting from a helical cyclic peptide backbone designed with Rosetta, the team used AfCycDesign to redesign the sequence, resulting in substantial sequence variation while retaining the overall structural framework. Energy landscape analysis suggested that the redesigned peptide had a stronger preference for the intended folded conformation compared to the original Rosetta design. Experimental validation by racemic X-ray crystallography confirmed close agreement between the designed model and the observed structure. In broader benchmarks across many backbones, AfCycDesign consistently produced sequences with improved structural confidence and distinct amino acid usage patterns compared to Rosetta-based designs, indicating enhanced folding reliability and design quality.
Sequence design of cyclic peptide backbones using AfCycDesign2,4
The team developed a hallucination-based method that jointly optimizes structural confidence metrics to generate both sequence and structure for cyclic peptides of varying lengths. After clustering the generated models, a wide range of unique structural families was identified, including a subset with high confidence predictions. Representative peptides were synthesized and structurally validated, and the experimental results showed close agreement with the design models. For longer cyclic peptides, the approach successfully produced stable secondary structure elements such as helices and β-sheets, with crystal structures closely matching the predicted conformations. Overall, these findings demonstrate that stable cyclic peptides longer than 10 residues can be designed without relying on disulfide bonds.
The researchers grafted a 5-residue motif from p53 onto high-confidence hallucinated scaffolds, designed sequences using ProteinMPNN, filtered with AfCycDesign predictions, and synthesized 14 cyclic peptides. AlphaLISA screening identified five designs with >50% inhibition. The best binder, RMG_14, had an IC50 of 338 nM. X-ray crystallography confirmed that the binding mode of RMG_14 to MDM2 matched the design model closely (Cα RMSD 1.0 Å). Further incorporation of non-canonical amino acids (5-fluoro-tryptophan, β-cyclobutyl-alanine, D-glutamine) gave RMG_14c, with a ~10-fold improvement in affinity by SPR (KD from micromolar to nanomolar).
Functionalized hallucinated scaffolds inhibit MDM2 in vitro3,4
For Keap1, the team used the EETG motif from a hot loop database, matched it to 775 hallucinated scaffolds, and after design and filtering, synthesized three cyclic peptides. Competitive fluorescence polarization assays showed that KC4 had an IC50 better than the native Nrf2 peptide (a 16-residue linear peptide), and all three cyclic peptides achieved nanomolar inhibition. This demonstrates the potential of hallucinated scaffolds as a general grafting platform.
The Rettie team's work shows how deep learning can dramatically improve the efficiency and accuracy of cyclic peptide design. But applying such AI methods to your own projects still comes with practical hurdles.
Although AfCycDesign achieves high prediction accuracy, not all high-pLDDT sequences fold correctly in the lab. In the MDM2 binder design, only 5 out of 14 synthesized peptides showed >50% inhibition. For Keap1, only 3 out of 6 could be synthesized with sufficient purity, though all three worked. For teams without high-throughput screening capabilities, selecting the few most promising sequences from thousands of computational outputs remains a real bottleneck.
Current AfCycDesign only supports the 20 canonical amino acids. When the team optimized RMG_14c, they had to manually introduce fluoro-tryptophan, β-cyclobutyl-alanine, and D-glutamine – modifications that could not be predicted computationally. For projects that require improved metabolic stability or permeability via non-canonical amino acids, systematic integration of such building blocks into the design workflow is still an open problem.
The cyclic peptides in this study ranged from 7 to 16 residues, and the main cyclization method was head-to-tail amide bond formation. In real projects, low cyclization yields, side products (dimers, epimers), and purification difficulties are common – especially when sequences contain multiple prolines or non-natural residues. Moreover, batch-to-batch consistency is critical for pharmacological evaluation but is rarely considered at the design stage.
We provide modular cyclic peptide design support tailored to different project stages:
Design is guided not only by sequence performance but also by practical synthesis and downstream development needs:
Our team enables seamless transition from AI-designed sequences to SPPS synthesis, purification, and analytical characterization, supporting progression from early discovery to gram-scale supply.
From concept to candidate, cyclic peptide projects ultimately depend on obtaining high-purity, well-characterized, and reproducible molecules. Regardless of how sequences are generated, success relies on effective synthesis, purification, and analytical validation to ensure each candidate is suitable for further development.
Creative Peptides provides comprehensive support across the workflow, including design assistance, synthesis, purification, and characterization, helping to transform early-stage ideas into reliable, scalable cyclic peptide candidates.
Have a cyclic peptide project in mind?
Contact our scientific team today to discuss your requirements and explore a tailored solution for your project.
compounds.
References
