We Can Predict the Effects of Kinase Domain Mutations using Molecular Dynamics and Machine Learning

Abstract

Protein kinases, which are known to be involved in cell proliferation, differentiation, and survival, have been shown to also play a key role in cancer formation. Numerous cancer therapies have been developed that target over-active kinases, but the ease of sequencing cancer genomes has shown that kinase proteins can be mutated at virtually any residue of the kinase catalytic domain. This raises the question of which patients should get which treatments, as these therapies generally target only over-active kinases, but it is generally not known whether each new kinase domain mutation is active without conducting laborious experiments. This has created a strong desire, both in the clinic and in the lab, to be able to computationally predict the effects of kinase domain mutations. Here, we report on a method which utilizes a convergence between standard molecular dynamics, free energy calculations, and machine learning approaches in order to predict the effects of kinase domain mutations. This method has been shown to give few false negatives and very few false positives across dozens of mutations in several different kinases, from both the tyrosine and serine/threonine families, and in both transmembrane receptor as well as soluble kinases. This method is much more accurate, sensitive, and specific than widely used mutation assessment algorithms which are either not mechanistic, not kinase specific, or are not either. As such, kinase specific mechanistic models could have a role in the clinic for directing patient care and the insights gained from these models could also be leveraged to design new generations of kinase specific inhibitors.