Classification and functional prediction of fungal glycosyltransferases using machine learning and deep learning...

Glycosyltransferases (GTs) are ubiquitous multifunctional enzymes, responsible for catalyzing the transfer of sugar residues to generate structural diversity in biological systems. The extensive variability in structure, function, and mechanisms of action of GTs has resulted in multiple classification systems, hindering identification of significant patterns. This study aims to address this complexity by using machine learning and deep learning algorithms to uncover structural patterns associated to their function. A dataset of 3340 GTs from 88 representative fungal species across various divisions was utilized. Employing their predicted three-dimensional folding as well as their amino acid structural and biochemical characteristics, a Convolutional Variational Autoencoder (CVAE) model was used to summarize this information into five vectors, which were enriched with new derived variables and descriptive statistics. Cluster analysis and classification were conducted using the k-means algorithm to identify similarity patterns among the proteins. The clustering generated from enriched data exhibited greater coherence with the groups formed by k-means than the one generated solely from structural data. These findings demonstrated that the methodology employed, which integrates all variables in a unified manner, facilitates a comprehensive understanding of the relationships between GTs compared to approaches based on disparate criteria. The use of a CVAE, along with the innovative integration of biochemical and structural variables, resulted in a complementary and comprehensive, data-driven classification system of GTs. This approach represents a significant methodological advance beyond traditional sequence-based approaches as well as in the use of deep learning methodologies for the study of GTs and their biological functions.