Addressing the Problem of Lysine Glycation Prediction in Proteins via Recurrent Neural Networks

A distinguishing feature of the metabolic disorder diabetes involves elevated damage to cellular components. Glycation, in contrast to glycosylation, is regarded as a strictly nonenzymatic process that involves the reaction of sugars (e.g., glucose and fructose) and sugar-derived molecules (e.g., methylglyoxal) with amino groups of biologically highly relevant molecules, such as nucleic acids, lipids, and proteins. The primary form of alteration arises from the chemical interaction between glycating agents and proteinaceous arginine/cysteine/lysine residues. Glycation may result in the formation of advanced glycation end-products (AGEs) which are mostly detrimental and compromise the function of the target molecule irreversibly. There are no clear sequence motifs in proteins that allow a straightforward identification of potential glycation sites. However, the physicochemical properties of amino acids that flank the glycated residue seem to play a key role in determining if glycation occurs or not. Here, we used a curated version of the CPLM database to implement a recurrent neural network strategy for the classification of lysine glycation to better understand which of eight physicochemical properties might influence glycation more than others. By using the most promising property for the characterization of amino acids next to lysine sites, isoelectric point, it was possible to obtain a 59.6% accuracy for correctly predicting lysine glycation. When the properties mass and torsion angle were used together, the accuracy increased to approximately 60%. Overall, our approach contributes to the understanding of glycation principles and can aid the task of narrowing down possible sites of lysine glycation in protein targets for further analysis.