Docking Peptides into HIV/FIV Protease with Deep Learning and Focused Peptide Docking Methods

Molecular docking is a structure-based computational drug design technique for predicting the interaction between a small molecule (ligand) and a macromolecule (receptor). Over the past three decades various docking software programs have been developed, mostly for drug-like molecules. With the recent interest in peptides as therapeutic molecules, several peptide docking methods have also been developed. AutoDock CrankPep (ADCP), is a state-of-the-art peptide docking tool that predicts the interaction of peptide with up to 20 amino acids in a user defined region of a macromolecule, i.e.focused docking. Recent advances in deep learning (DL) approaches have shown remarkable success in docking linear peptides composed of natural amino acids only. Unlike ADCP, these methods provide a confidence level in their predictions. Here we explore whether ADCP and various DL methods (AlphaFold2 Monomer, AlphaFold2 Multimer, and OmegaFold) and their prediction confidence metric can be used to discriminate native and non-native substrates for HIV and FIV proteases. We found that ADCP successfully predicts the interactions of native peptides but fails to discriminate non-native ones. This was expected as conventional docking methods report solutions maximizing ligand receptor interactions for any ligand. Surprisingly, DL methods underperform when docking native peptides into these particular docking targets but achieve high success rates with non-native peptides. While AlphaFold managed to successfully dock a few of the native peptides, OmegaFold failed to successfully dock any of them. Overall, none of these methods is currently able to distinguish between native and non-native peptides, warranting further exploration of specialized methodologies.