Computational Intelligence-driven DPP-4 Inhibitory Peptide Prediction and Bispecific Antibody Design: Integrating Deep Learning, Molecular Dynamics, and Nanodisc Validation

Abstract

Investigating peptides and antibodies targeting DPP-4 in vitro experiments requires the presence of the native lipid bilayer, as catalytic activity depends on it. We propose a methodology that involves molecular docking, 100-ns molecular dynamics simulations of DPP-4 in a POPC lipid bilayer mimicking nanodisc conditions, a novel dual-channel BiLSTM deep learning model for predicting DPP-4 inhibitor peptides, CDR affinity maturation based on Rosetta, and optimizing the structure of DPP-4 antibodies. For predicting peptides, we developed a novel dual-channel model based on the ESM-2 protein language model and docking data with expected R2 and Spearman correlations of 0.80-0.85 and 0.70-0.80, respectively, based on the same peptide data used in the experiment within the nanodisc system. In optimizing the structure of DPP-4 antibodies, we use IgFold for predicting the structure of DPP-4 antibodies, ClusPro/HADDOCK for mapping binding sites, and virtual saturation mutagenesis for the optimization of CDRs, resulting in a 2-5 times increased binding affinity. A bispecific DPP-4 inhibitor antibody with GLP-1R agonist activity is designed following the principle of knob-into-hole and using AlphaFold-Multimer.