Title: MuCO: Generative Peptide Cyclization Empowered by Multi-stage Conformation Optimization

Abstract:

Accurately modeling peptide cyclization is essential for virtually screening peptide candidates that possess favorable pharmaceutical and physical characteristics. This process presents significant difficulties, as cyclic peptides frequently adopt a variety of ring-shaped conformations that deterministic models, originally designed for linear peptide folding, fail to adequately represent. To address this, we introduce MuCO (Multi-stage Conformation Optimization), a novel generative approach that models the distribution of cyclic peptide structures conditioned on their linear counterparts.

MuCO simplifies the complex task of peptide cyclization by dividing it into three distinct phases: topology-aware backbone design, generative side-chain packing, and physics-aware all-atom optimization. This hierarchical, coarse-to-fine strategy facilitates the generation and refinement of cyclic peptide conformations. The framework supports an efficient parallel sampling method, enabling the rapid discovery of diverse, low-energy structures.

Evaluations conducted on the large-scale CPSea dataset reveal that MuCO consistently surpasses current state-of-the-art techniques across key metrics, including physical stability, structural diversity, secondary structure recovery, and computational efficiency. These results position MuCO as a highly promising computational resource for the exploration and design of cyclic peptides. A demonstration of the method is available at https://github.com/mianqiu00/MuCO.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC