Reinforcement Learning-Based Target-Specific De Novo Design of Cyclic Peptide Binders

Abstract Cyclic peptides are promising therapeutic agents for challenging targets, especially protein–protein interactions. However, computationally designing cyclic peptide binders remains challenging. Here, we present CYC_BUILDER, a reinforcement learning-based framework that assembles peptide fragments and performs efficient cyclization via head-to-tail amide or disulfide bonds, which uses a Monte Carlo Tree Search to guide fragment selection, peptide growth, and structure refinement. We show that CYC_BUILDER was able to successfully regenerate native binding sequences and poses for known cyclic peptide–protein complexes. We have applied CYC_BUILDER to generate cyclic peptide binders for TNFα and found that the design results outperformed those from AfCycDesign and Anchor Extension in binding energy, structural diversity, and efficiency. We experimentally tested the activity of nine designed peptides, and four of them demonstrated potent binding and cellular activity. CYC_BUILDER offers a powerful tool for cyclic peptide discovery with broad applications in therapeutics and synthetic biology.