Mechanism-Driven Features Enable Asn Deamidation Reactivity Prediction via Machine Learning Methods

The spontaneous deamidation of Asparagine (Asn) residues is a common post-translational modification of proteins that can occur on disparate time scales, ranging from hours to thousands of years. This variability in the reaction rate reflects the influence of structural and environmental factors on the multistep mechanism of the deamidation reaction. Understanding the fine connection between reactivity and these modulating factors is essential to advance our knowledge of the deamidation kinetics in proteins and improve the prediction of deamidation-prone residues. In this work, we assessed the step-specific structural-dynamics parameters underlying the chemical basis of the first two reaction stages (the deprotonation and ring-closure steps) and developed novel descriptors derived from molecular dynamics (MD) simulations, which encompass solvation, hydrogen bonds, conformational free energy, and an environment electrostatic effect. These descriptors were evaluated across 63 Asn residues from six distinct proteins and used as input features for three machine learning models, Random Forest, Naive Bayes, and Logistic Regression, to classify Asn residue reactivity. Among these, the Random Forest classifier achieved the best predictive metrics, underscoring the significance of mechanism-tailored features in discriminating Asn reactivity and unveiling the key physicochemical factors that govern deamidation rates in proteins.