Design of peptide inhibitors targeting β-catenin using generative deep learning and molecular dynamics simulations

Abstract Computational design of therapeutic peptides that enables fast exploration of a combinatorial number of amino acid possibilities has emerged as a popular strategy in antibiotics discovery but identifying peptide inhibitors targeting specific protein-protein interactions (PPIs) still faces significant challenges. Here, we present a new approach for de novo design of target-specific peptides, which combines a Gated Recurrent Unit (GRU) based variational autoencoder (VAE) with Rosetta FlexPepDock for peptide sequence generation and potency assessment, followed by the refinement with all-atom molecular dynamics (MD) simulations. We have utilized the computational protocol to design eight peptide inhibitors specifically disrupting the interaction of β-catenin with TCF/LEF family transcription factors. Four of these peptide inhibitor candidates were synthesized and experimentally tested, of which two displayed enhanced binding affinities for β-catenin compared to the best-known peptide made of all natural amino acids. We also constructed a random peptide library of 710,000 unique peptides, and none of these peptides showed a higher binding affinity than our computationally designed peptides. Our study exemplifies the success of combining deep learning and structure-based modeling and simulation for target-specific peptide design. Besides the development of therapeutic peptides targeting β-catenin/TCF, our approach should be transferable to other PPIs, such as p53/MDM2, Bcl-2/Bax, and CD40/CD40L.