CovaTide: Covalent Topology-Conditioned All-Atom Generation of Linear and Cyclic Peptides

Peptide binders offer a versatile therapeutic modality for targeting protein surfaces, but conformational flexibility and proteolytic degradation can limit their utility. Cyclization can stabilize binding conformations and improve proteolytic resistance, yet cyclic peptide design must coordinate global receptor-bound geometry with local closure chemistry. Existing methods often cyclize generated poses post hoc or transfer linear-peptide generators to cyclic peptide design without cyclic-complex training data. We introduce CovaTide, a receptor-conditioned all-atom flow model jointly trained on linear and cyclic complexes. Operating directly on atomic coordinates, CovaTide generates structures and sequences in a unified design space spanning linear peptides and head-to-tail, disulfide, and isopeptide cyclization. Multiscale topology conditioning couples residue connectivity with atom-level chemistry throughout generation, avoiding post-generation endpoint mutation. Compared with the evaluated baselines, CovaTide improves sequence and binding-pose recovery for linear peptides and achieves higher cyclization and energy-based success rates for cyclic peptides after relaxation. These results support explicit covalent topology conditioning for unified receptor-conditioned peptide design.