Mining the endogenous peptidome for peptide binders with deep learning-driven optimization and molecular simulations

The endogenous peptidome is the complete set of naturally occurring peptides, some of which are involved in processes like intercellular communication, and in innate immunity. The peptidome is also a potential source of uncharacterized peptide-based therapeutics, which has remained largely unexplored due to challenges in identifying bioactive peptides in vast peptidomic datasets. Here, we present a generalized computational pipeline that mines datasets for peptides with high affinity binding capacities. The approach employs protein-peptide docking, binding interface scoring and a deep ensemble model to dynamically learn the mapping between peptide embeddings and binding capacity. We demonstrate the utility of the pipeline by identifying endogenous peptide binders for the LPS-binding site of CD14 using experimentally defined wound fluid peptidomes. Promising candidates underwent conformational refinement and validation through all-atom molecular dynamics simulations. This pipeline offers a systematic way to uncover novel peptide binders within large peptidomes, paving the way for accelerated discovery of peptide-based therapeutics.