Harnessing deep statistical potential for biophysical scoring of protein-peptide interactions
Abstract
Protein-peptide interactions (PpIs) play a critical role in major cellular processes. Recently, a number of machine learning (ML)-based methods have been developed to predict PpIs, but most of them rely heavily on sequence data, limiting their ability to capture the generalized molecular interactions in three-dimensional (3D) space, which is crucial for understanding protein-peptide binding mechanisms and advancing peptide therapeutics. Protein-peptide docking approaches provide a feasible way to generate the 3D models of PpIs, but they often suffer from low-precision scoring functions (SFs). To address this, we developed DeepPpIScore, a novel SF for PpIs that employs unsupervised geometric deep learning coupled with a physics-inspired statistical potential. Trained solely on curated experimental structures without binding affinity data or classification labels, DeepPpIScore exhibits broad generalization across multiple tasks. Our comprehensive evaluations in bound and unbound peptide bioactive conformation prediction, binding affinity prediction, and binding pair identification reveal that DeepPpIScore outperforms or matches state-of-the-art baselines, including popular protein-protein SFs, ML-based methods, and AlphaFold-Multimer 2.3 (AF-M 2.3). Notably, DeepPpIScore achieves superior results in peptide binding mode prediction compared to AF-M 2.3. More importantly, DeepPpIScore offers interpretability in terms of hotspot preferences at protein interfaces, physics-informed noncovalent interactions, and protein-peptide binding energies.
Keywords:
protein-peptide interactions; geometric deep learning; statistical potential; scoring function; AlphaFold