Model Agnostic Conditioning of Boltzmann Generators for Peptide Cyclization

Cyclic peptides offer strong therapeutic potential due to their enhanced binding affinity and protease resistance, but their design remains challenging due to limited structural data and existing models that consider only ground states rather than conformational ensembles. We introduce CycLOPS, a model-agnostic framework that conditions Boltzmann generators to sample valid cyclic conformations without retraining. To address cyclic peptide data scarcity, we reformulate design as conditional sampling over linear peptides via chemically informed loss functions; CycLOPS encompasses 18 inter-amino acid crosslinks across 6 diverse chemical reactions and leverages tetrahedral geometry constraints through 6 interatomic distances defining a kernel density-estimated joint distribution from MD simulations. It is readily extensible to many more chemistries, too. We demonstrate the versatility of CycLOPS across two distinct generative model architectures---the Sequential Boltzmann Generator and the Equivariant Continuous Normalizing Flow++. CycLOPS successfully biases sampling toward chemically plausible macrocycles in both cases.