Cyclic peptides are promising drug candidates due to their structural stability, proteolytic resistance, and potential oral bioavailability. While recent generalized biomolecular models have expanded to support noncanonical amino acids (NCAAs), high-precision structure prediction for small cyclic peptides containing NCAAs and diverse cyclization chemistries remains challenging. We adapted AGDIFF, an all-atom diffusion generative model originally designed for small molecule conformer generation, and retrained it on the 36,198-member Conformer Rotamer Ensembles of Macrocyclic Peptides (CREMP) data set. Using a 2D molecular graph representation, the model inherently supports NCAAs and complex linkages. After applying a stereochemical correction step, AGDIFF achieved high accuracy on benchmark peptides (average RMSD 0.79 Å; ring torsion fingerprint deviation 6.55°), resolving stereochemical insensitivity and reliably producing the correct antipodes of enantiomeric residues. Ramachandran analyses confirmed the conformational plausibility of the generated ensembles. These results demonstrate the effectiveness of diffusion-based deep learning for cyclic peptide modeling and highlight its potential for the rational design of NCAA-containing macrocyclic peptides in drug discovery.
Wu et al. (Fri,) studied this question.