Fast Generation of Simulation-Quality Structural Ensembles of Mixed-Chirality Cyclic Peptides via Diffusion Models.

Cyclic peptides are an emerging therapeutic modality, with recent computational efforts focusing on the design of cyclic peptides that predominantly adopt a single conformation. However, many cyclic peptides adopt multiple conformations in solution, existing as structural ensembles. This conformational flexibility is often integral to their function: chameleonic switching between alternative states can enhance membrane permeability, and specific conformations may be required for molecular recognition and binding. Consequently, the ability to predict their structural ensembles is crucial for advancing the de novo design of cyclic peptide therapeutics. Here, we introduce diffusion models to efficiently and accurately predict structural ensembles of mixed-chirality cyclic peptides. The models are trained directly on molecular dynamics (MD) simulation data; in particular, each frame of the simulation becomes a single training instance in which a structure is represented as sine and cosine values of backbone dihedral angles. The trained diffusion model can not only generate MD-quality structures of cyclic peptides, but also the generated structures follow the Boltzmann distribution sampled in the MD simulation, enabling a deeper understanding of the physicochemical basis of cyclic peptide properties and allowing efficient computational design of cyclic peptides targeting biologically relevant systems.