Deep learning for binding site segmentation in protein ensembles

Multiple protein conformations are used to accurately represent binding site structure and dynamics in ensemble docking or de novo drug design. Following conformational selection, these conformations are often generated through molecular dynamics (MD) simulations of the protein in the ligand-free (apo) state. However, it can be difficult to identify within these ensembles the conformations most likely to bind a ligand in the absence of prior information on the ligand-bound (holo) state. This research aims to develop a deep learning (DL) tool that can scan a putative binding site in a protein and recognise the shape and physicochemical features compatible with the binding of chemical fragments using semantic segmentation.