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

TL;DR

MuCO addresses the challenge of generating diverse, physically plausible cyclic peptide conformations from linear sequences by decoupling cyclization into three stages: topology-aware backbone generation, generative side-chain packing, and physics-aware all-atom refinement. It employs SE(3) flow matching for backbones, torsional flow matching with cyclic-aware encoding for side chains, and Charmm36-based energy minimization to ensure ring closure and low energy. This hierarchical, parallelizable framework enables efficient exploration of a rugged conformational landscape, achieving lower energies and richer diversity than state-of-the-art baselines on the CPSea-derived datasets. The approach offers a scalable computational tool for cyclic peptide design, with potential impact on drug discovery and materials science, while future work aims to extend to multi-chain systems and experimental validation.

Abstract

Modeling peptide cyclization is critical for the virtual screening of candidate peptides with desirable physical and pharmaceutical properties. This task is challenging because a cyclic peptide often exhibits diverse, ring-shaped conformations, which cannot be well captured by deterministic prediction models derived from linear peptide folding. In this study, we propose MuCO (Multi-stage Conformation Optimization), a generative peptide cyclization method that models the distribution of cyclic peptide conformations conditioned on the corresponding linear peptide. In principle, MuCO decouples the peptide cyclization task into three stages: topology-aware backbone design, generative side-chain packing, and physics-aware all-atom optimization, thereby generating and optimizing conformations of cyclic peptides in a coarse-to-fine manner. This multi-stage framework enables an efficient parallel sampling strategy for conformation generation and allows for rapid exploration of diverse, low-energy conformations. Experiments on the large-scale CPSea dataset demonstrate that MuCO significantly outperforms state-of-the-art methods in consistently in physical stability, structural diversity, secondary structure recovery, and computational efficiency, making it a promising computational tool for exploring and designing cyclic peptides.

Figures (14)

• Figure 1: An illustration of the 3-stage scheme of MuCO. The generation process is decoupled into three stages, including $i)$Topology-Aware Backbone Generation for backbone scaffold $B$ by $\mathrm{SE(3)}$ flow matching on condition of sequence embedding, $ii)$ Generative Side-chain Packing for side chain sampling via conditional flow matching with cyclic RPE, $iii)$ Physics-Aware Optimization by Charmm36 forcefield to yield final energy minimized conformation.
• Figure 2: MuCO framework overview. The process decouples into three stages: (1) cyclic backbone generation on $\mathcal{M}_\mathcal{B}$, (2) side-chain packing on $\mathcal{M}_\mathcal{C}$ with Cyclic RPE, and (3) physics-aware refinement to reach the local energy minimum $\bm{X}$.
• Figure 3: (a) Success Rate (%): Percentage of successful cyclization. (b) Mean Potential Energy ($E$): Physical stability calculated via Charmm36. (c) Structural Diversity ($H_\mathcal{C}$): Shannon entropy of conformational clusters. (d) Mode distribution coverage on the CPTrans dataset.
• Figure 4: (a) Secondary Structural Diversity ($H_{SS}$) across different test sets; MuCO maintains superior ensemble diversity compared to all baselines. (b) Secondary structure composition (Helix, Sheet, and Coil) on the CPSea-PDB dataset. MuCO significantly improves the recovery of regular secondary structures.
• Figure 5: Case study on structural fidelity. Visual comparison of predicted conformations for a representative sample from the CPSea PDB dataset. Under single sampling, MuCO reconstructs the native-like $\alpha$-helical motif and achieves significantly lower potential energy ($E$) compared to representative baselines.