CycloPepper: Predicting Peptide Cyclization Sites
CD ComputaBio provides software-based computational services to support research and development. We do not offer free software packages.
Overview
Designing a promising cyclic peptide does not establish how best to synthesize it. Even when the final ring is unchanged, selecting a different bond for head-to-tail closure changes the linear precursor, its terminal residues, and its behavior on resin. CycloPepper addresses this practical retrosynthetic decision: which precursor arrangements are likely to cyclize under a defined synthetic protocol?
In a 2026 Nature Communications study, Yourong Pan, Chengrui Hu and colleagues, with corresponding authors Xin Hong and Chengxi Li, combined automated synthesis with supervised machine learning. The team, working across Zhejiang University and Henan Normal University, generated 306 experimentally labeled precursors, compared 12 classifiers, and selected a support vector classifier. Its mean five-fold cross-validation accuracy was 84%; subsequent experiments agreed with 64 of 74 predictions, approximately 86%. These results support a focused feasibility filter, rather than a predictor of isolated yield or therapeutic performance.1
Why the Same Ring Can Require Different Synthetic Routes
Head-to-tail cyclization closes a peptide through an amide bond between its termini. A target ring can be opened computationally at different peptide bonds to generate alternative linear precursors. These are cyclic rotations, rather than arbitrary rearrangements of the amino-acid order. FPGM and PGMF, for example, can describe different precursor routes to the same cyclic sequence, but their terminal environments differ during synthesis.
Successful closure depends on bringing the termini into a productive geometry while competing with intermolecular reactions. Backbone preferences, local steric effects, ring size, and resin attachment can change that balance. Consequently, a sequence-level decision must retain the proposed disconnection. Treating all rotations as identical training examples would discard the very synthetic distinction the model is intended to learn.
The paper's broader lesson is that synthetic accessibility belongs in candidate selection alongside binding and developability. However, its scope is specifically on-resin head-to-tail closure using the authors' chemistry. It does not establish a general feasibility rule for disulfide, side-chain, stapled, or other macrocyclization reactions.
Automation Creates the Dataset Before Learning Begins
The authors used a diaminonicotinic acid (DAN) linker system with automated solid-phase peptide synthesis. Isoamyl nitrite activates the linker to a reactive triazolopyridine intermediate; base addition then supports intramolecular closure. CycloBot hardware and AutoXpert control software standardize reagent delivery, temperature, and processing, reducing the methodological variation encountered when collecting outcomes from unrelated publications.
The reported dataset protocol used 40 mg of DAN resin at a loading of 0.36 mmol/g. Linear synthesis occurred at 90 °C, activation at 25 °C, and cyclization at 50 °C. Products were assessed by LC-MS: detection of the target molecular weight defined a positive label, and its absence defined a negative label. This is a binary analytical endpoint. It does not quantify purified yield, product purity, epimerization, or the amount of competing material.
Of the 306 entries, 174 were positive and 132 negative. The length groups contained 45 dipeptides, 53 tripeptides, 120 sequences of four to six residues, 69 of seven to nine residues, and 19 of ten or more residues. The distribution therefore favors shorter peptides; the upper length range has much less experimental coverage.
Dataset design combined randomly generated seeds with manual additions, deletions, and substitutions. Each canonical amino acid appeared in more than 20 distinct sequences. This improves compositional coverage but does not make the dataset a random sample of medicinal-chemistry programs. Closely related variants can be informative experimentally while still requiring careful separation during machine-learning evaluation.
Sequence Features and a Modest Model Comparison
CycloPepper represents each precursor through four complementary feature groups. Position-specific one-hot encoding preserves residue order; amino-acid composition captures overall content; dipeptide composition describes adjacent residue pairs; and handcrafted indicators encode empirical synthetic observations. Together these generate 716 features: 280, 20, 400, and 16 dimensions, respectively. The Methods report a reduction to 414 dimensions after removing all-zero dimensions.
The handcrafted features include terminal-residue patterns associated with different cyclization outcomes. These are useful chemical priors, but their proposed explanations, such as steric or conformational effects, should be read as interpretations rather than independent mechanistic proof. The model predicts labels from descriptors; it does not explicitly simulate the ring-closing transition state.
| Classifier | Five-fold accuracy, mean ± SD | Interpretation |
|---|---|---|
| Support vector classifier | 0.84 ± 0.06 | Selected for subsequent prediction |
| k-nearest neighbors | 0.83 ± 0.06 | Similar mean performance |
| Gradient boosting | 0.82 ± 0.04 | Competitive conventional baseline |
| Gaussian naive Bayes | 0.74 ± 0.05 | Lowest reported mean |
Twelve individual and ensemble classifiers were evaluated. The authors report no statistically significant performance differences among the algorithms and overlapping variability. Accordingly, SVC is the selected model within this comparison, not evidence that one algorithm decisively outperforms the alternatives. Consistent experimental labels are at least as central to this study as model architecture.
What the 86% Validation Result Actually Measures
The experimental test enumerated potential closure positions rather than checking only a favored precursor. Figure 4 presents eight randomly generated cyclic sequences and four bioactive cyclic peptides from Pseudostellariae radix. Their alternative precursors total 74 experiments: 49 for the random group and 25 for the natural-product group. The authors report 39 experimentally cyclizable precursors and 35 non-cyclizable precursors.
Across these tests, 64 predictions matched the experimental labels, giving 64/74 = 86.5%, reported as 86%. Correct negative predictions contribute to this number. It therefore cannot be described as an 86% synthesis success rate, an 86% yield, or an 86% chance of obtaining a useful drug candidate.
Performance also varied by target ring. All tested positions were classified correctly for several sequences, whereas the two seven-residue examples cyclo-QEYQHEK and cyclo-WMLWEPF each had five correct predictions out of seven. Natural-product examples ranged from 75% to approximately 86% agreement. Aggregate accuracy is useful for screening, but does not remove the need to test the chosen precursor.
Reporting note. One sentence in the paper describes the 74 experiments as spanning "seven" cyclic peptide sequences. Figure 4, the preceding descriptions, and the site counts instead identify 12 rings. The 12-ring interpretation used here follows those itemized results; the headline 64/74 outcome is unchanged.
From Site Prediction to Candidate Prioritization
The study packages its SVC model as CycloPepper, with standalone and web interfaces. A sequence is encoded and assessed across possible disconnections, producing positional feasibility calls within seconds. The standalone software supports CSV batch input. A displayed success label indicates at least one predicted feasible site; it remains a computational output pending synthesis.
The authors also examined 561 designed cyclic peptides derived from antibody complementarity-determining regions associated with seven targets: CD52, PD-L1, TNF-α, CD41, CEACAMs, HER2, and APP. They report 341, or 60.8%, as feasible in this computational application. Because the paper uses both sequence and site terminology around this result, it should be treated as a reported feasibility-screening output rather than a count of experimentally confirmed products.
No target-binding, cellular efficacy, permeability, or in vivo validation is established by that screen. Likewise, excluding predicted failures could reduce experiments in a future workflow, but the study does not demonstrate a measured percentage reduction in project costs or development time. Its immediate decision value is narrower: nominate precursor routes for experimental evaluation.
Boundaries and Practical Validation Priorities
Transfer to another synthesis setup requires checking protocol compatibility. A model trained on DAN-resin chemistry may behave differently under alternative coupling reagents, resin loadings, solvents, or solution-phase conditions. The input encoding covers 20 canonical amino acids and sequences up to 14 residues; noncanonical residues, extensive backbone modification, and longer macrocycles need separate evidence.
Dataset size also limits generalization claims. The authors measured train-to-validation sequence identity and conducted a rotation-invariant audit for initially high-similarity pairs at a 60% identity threshold. This recognizes that circularly equivalent representations can obscure relatedness. A future assessment should additionally group related rings and precursor families, challenge genuinely new chemical space, and examine false positives and false negatives separately.
For development teams, the next useful endpoint is prospective validation on their own chemistry, recording quantitative yield and purity alongside binary feasibility. Comparing predicted sites with chemist-selected routes would show whether the filter changes practical decisions. Binding and developability measurements should remain separate: an easier synthetic route does not by itself improve pharmacology.
Overview of What CD ComputaBio Can Provide
The study motivates connecting synthetic feasibility with computational peptide research. The following established service areas relate to adjacent decisions; they do not imply ownership of CycloPepper, access to CycloBot, or reproduction of the paper's accuracy.
| Research Need | Related CD ComputaBio Support | Connection to This Article |
|---|---|---|
| Assess ring conformations | Cyclic Peptide Modeling Services | Explore structural hypotheses for cyclic peptide candidates. |
| Compare precursor sequence features | Protein Sequence Analysis Service | Annotate sequence differences relevant to candidate review. |
| Organize candidate collections | Peptide Library Design Services | Plan sequence diversity and candidate prioritization. |
| Explore target recognition | Protein-Peptide Docking Service | Evaluate putative binding poses separately from synthetic feasibility. |
| Examine conformational flexibility | Molecular Dynamics Simulation Service | Investigate modeled dynamics without claiming reaction-yield prediction. |
References
- Pan Y, Hu C, Li J, Wan F, Hong X, Li C. CycloPepper: a machine learning platform for predicting cyclization outcomes and optimizing synthesis of therapeutic cyclopeptides. Nature Communications. 2026;17:2803. https://doi.org/10.1038/s41467-026-69441-w.
* For Research Use Only.
