Fast end-to-end surface interpretation of SARS-CoV-2 variants by differentiable molecular surface interaction fingerprinting method

Abstract

Confronting the challenge of persistent mutations of SARS-CoV-2, researchers have turned to deep learning methods to predict the mutated structures of spike proteins and to hypothesize potential changes in their structures and drug efficacies. However, limited works are focused on the surface learning of spike proteins even though their biological functions are usually defined by the geometric and chemical features of 3D molecular surfaces. In addition, the current used geometric deep learning methods are based on mesh representations of proteins to identify potential binding targets for drugs. However, the use of meshes has limitations and is not applicable for many important tasks in molecular biology. To address these limitations, we adopt the differentiable molecular surface interaction fingerprinting (dMaSIF) method which is based on the 3D point clouds and a novel efficient geometric convolutional layer to fast predict the interaction sites on the protein surface. The different binding site patterns for Delta, Omicron and its subvariants are clearly visualized. We find that Delta and Omicron show the similar surface binding patterns while BA.2, BA.2.13, BA.3 and BA.4 present similar ones. BA.4 possesses higher positive interaction site ratio than the others which may account for its higher transmission and infection among humans. In addition, the positive interaction site ratios of BA.2, BA.2.13, BA.3 are higher than Delta and Omicron, which are accordant with their transmission and infection rates. Hopefully our work offers a new effective route to analyze the protein-protein interaction for the SARS-CoV-2 variants.