Relation in ligand dynamics and protein mutation using unsupervised machine learning.

Abstract

Drug resistance mutation, which decreases ligand binding affinity to the target biomolecule, is a crucial issue all over the world. To overcome this issue, many approaches are proposed and tried. Free energy calculations are an effective and widely used method to estimate the ligand binding affinity. Relative free energy calculations can estimate the influence of the mutation on the ligand binding. However, the accuracy of the estimation should be limited partly because of insufficient sampling and relaxation of the mutants in the short molecular dynamics (MD) simulations. On the other hand, longer MD simulations make the calculation cost incredibly expensive. To improve the accuracy and computational cost, we have performed a machine-learning-assisted analysis of MD simulations [Yasuda, Commun. Biol. (2022)] of mutant protein-ligand systems. We used nilotinib-Abl-kinase systems, which were previously reported by other groups [Hauser, Commun. biol. (2018), Aldeghi, ACS Cent. Sci. (2019) ]. Our approach can compare the difference in the protein dynamics on ligand binding using unsupervised learning and extracts features of the dynamics of each mutant protein. We will show the relation of the feature to the ligand-binding affinity. Furthermore, interpretation of the feature enables us to understand the part of the ligand that shows different dynamics induced by the mutants. This would provide hints for the mechanics of the binding and improvements of the ligand.