(707c) Computational Algorithms for Molecular Profiling of Activating Mutations in Kinases

Computational algorithms to assess the mutational landscapes of protein kinases mutated in cancer are desirable since the prevalence of mutations (some activating, some passenger) make clinical decisions for patient enrollment in drug trials challenging. While data-driven techniques provide a scalable approach to assess the mutational effect on protein structure and function, they often overlook the mechanism underlying the effect thereby resulting in high false positive rates.

Firstly, we implement a data driven technique that is a significant improvement over SIFT and PolyPhen2 in predicting driver and passenger mutations in kinases based on machine learning algorithm that implements features from the biochemical space. Moreover, through implementing combined computational studies of molecular dynamics (MD) as well as enhanced sampling simulations, we obtain mechanistic understanding of conformational transition in kinase and effectively predict the experimentally determined activation status of mutation in kinase by identifying hydrogen bond fingerprint in the activation loop and the αC-helix regions and this outperforms existing standard methods with drastically reduced false positives when validated against experiments while also leading to the unraveling of convergent mechanisms in the oncogenic activation of mutations in ALK kinase which is known to have implications in pediatric cancer. Finally, through this technique we propose to append explainability and interpretability to the supervised-ML algorithm through engineering features from the structural transition of the kinase from the inactive to the active state, with the structural transition being obtained through enhanced sampling technique of metadynamics.

We acknowledge financial support from the NIH and computational support from XSEDE.