Benchmarking protein structure predictors to assist machine learning-guided peptide discovery

Machine learning models provide an informed and efficient strategy to create novel peptide and protein sequences with the desired profiles. Nevertheless, they are primarily trained on sequences where the tridimensional structures of peptides and proteins are often overlooked. We need a fast and reliable approach to estimate the structural diversity of medium-large training sets before building models. This study benchmarked four protein structure prediction methods (Jpred4, PEP2D, PSIPRED, AlphaFold2) using 261 curated and experimentally known structures from the PDBe database. We applied our best predictor to map the structural landscape of GRAMPA, the giant and vastly uncharted repository of 5980 antimicrobial peptides. The dataset was predominantly made of loose helices (65.1%), followed by random coils (17.8%), and beta-stranded and mixed structures accounted for the rest. Machine learning models provide an informed and efficient strategy to create novel peptide and protein sequences with the desired profiles.