Abstract
Cyclic peptides exhibit remarkable stability, membrane permeability, and binding affinity, positioning them as promising therapeutics. However, their synthesis, particularly on-resin head-to-tail cyclization, remains challenging, with cyclization site selection critically influencing yield. Here, we introduce a machine learning (ML) approach to predict cyclization outcomes, leveraging CycloBot, our fully automated cyclic peptide synthesis platform. Using this system, we generate a standardized dataset of 306 cyclic peptides (2-14 residues) and develop an ML model achieving an average prediction accuracy of 84%. Experimental validation with 74 random and therapeutic peptides showed an 86% prediction consistency. To facilitate practical use, we built CycloPepper, a user-friendly platform available through both web and software interfaces, enabling rapid cyclization site assessment. This tool effectively identified potential cyclization sites for disease-targeting peptides, including cancer biomarkers. Our work illustrates the potential of ML-assisted synthesis to streamline cyclic peptide synthesis and accelerate therapeutic discovery.