BPS2026 – Solvation-based, rotation-invariant Zernike moment representation for protein-protein interface prediction

Protein activity typically involves interactions with other macromolecular partners. Accurate prediction of protein-protein interaction (PPI) interfaces is important for complex structure determination and provides clues to possible regulatory mechanisms, often of therapeutic relevance. One class of approach to interface prediction relies on analysis of the biophysical characteristics and topography of the protein surface. Although grounded in first principles and offering physical insight, these methods struggle to integrate local and non-local features across complex surfaces. We present a novel method for PPI prediction that leverages a unique solvation-based surface descriptor combined with a convolutional neural network (CNN) operating on surface patches represented by Zernike polynomials. The solvation term is derived from our semi-explicit hydration model, which captures both the spatial arrangement and energetics of the first protein hydration shell. Unlike static hydrophobicity scales, it accounts for both the local environment of solvent-exposed groups and global properties of the macromolecule, including overall charge and surface topography. To capture non-local aspects of PPIs, protein feature maps are divided into circular patches of 10 Å radius, which are embedded in Zernike polynomials to generate rotation-invariant descriptors. Interface location within each patch is then predicted by our CNN using circular convolution in Zernike space. The model was trained on an extensive set of interfaces extracted from 15,000+ protein structures annotated in the 3D-complex database. Validation on the external DB5 protein docking data set demonstrates superior performance over current state-of-the-art approaches, for both bound and unbound protein geometries. This approach opens new possibilities for accurate mapping of interaction interfaces and for guiding the design of modulators of protein-protein interactions.