Computational design of functional cyclic peptides using deep learning

Abstract

Cyclic peptides have gained significant traction as a therapeutic modality. Given their ease of synthesis, expansive chemical space, and the promising pharmacokinetic properties of existing cyclic peptide drugs, cyclic peptides have been proposed as a mid-point between biologics and small molecules. Computational design of structured cyclic peptides has been successful using Rosetta, even design of membrane traversing cyclic peptides, but efforts to design binders to protein targets have led to only a handful of successful cases. Deep learning (DL) networks have recently shown considerable opportunities for accurate structure prediction and design of biomolecules that are potent inhibitors of therapeutically relevant protein interfaces. This work describes the application of a cyclic offset to the AlphaFold2 network as well RFdiffusion, resulting in accurate prediction and design of structured de novo cyclic peptides and high affinity cyclic peptide binders against protein targets of interest of diverse shape and function.