Accurate structure prediction of cyclic peptides containing unnatural amino acids using HighFold3

Abstract In recent years, cyclic peptides have emerged as a research hotspot in drug development due to their excellent stability, high specificity, and potential to penetrate intracellular targets. However, existing computational models still face challenges in making accurate predictions of the structures of cyclic peptides containing unnatural amino acids (unAAs), thereby limiting their application in drug design. The release of AlphaFold 3 (AF3) has significantly improved the modeling capability of biomolecular complexes and supports the definition of unAAs through residue modifications using the CCD database (Chemical Component Dictionary). Yet, its reliance on structures already present in the training library limits its ability to accurately predict cyclic peptide structures. Base on the AlphaFold 3 framework, we developed HighFold3, which comprises two sub-models: HighFold3-Linear and HighFold3-Cyclic, designed for predicting the structures of linear and cyclic peptides, respectively. Our research results show that, by incorporating a cyclic peptide positional preference matrix, HighFold3 achieves superior performance over other models (HighFold, HighFold2, and CyclicBoltz1) in predicting cyclic peptide structures, particularly excelling in complex cyclic peptides containing unAAs. HighFold3 provides an efficient tool for cyclic peptide drug design, promising to advance the application of cyclic peptides in targeted therapy, antibacterial treatments, and cellular penetration.