Integrative modeling of NS3 helicase function with molecular kinetics, molecular simulations and machine learning

Abstract

The viral NS3 helicase is a crucial enzyme involved in the RNA replication of positive strand ssRNA viruses of the Flaviviridae family. The NS3 is an active motor that binds RNA and couples ATP hydrolysis to the mechanical RNA unwinding process. This work explores allosteric processes that regulate NS3 activity. From the protein-ligand interaction perspective, we characterize the novel Zika virus (ZIKV) NS3 helicase. Mutations in conserved motifs and non-binding site residues reveal mechanisms by which NS3 helicase activity can be altered by regulatory residues. We report that ZIKV NS3 function can be enhanced by mutations that increase its RNA affinity and ATP hydrolysis activity, and unwinding deprecation can primarily occur by the loss of RNA binding affinity. We identify RNA annealing as a non-canonical function of ZIKV NS3 helicase and explore the allosteric modulation of this function by non-hydrolyzable ATP analogues. The discovery of three super-helicases motivate a kinetic model of RNA unwinding that differentiates the wild-type NS3 and mutants. These assays enhance our understanding of allostery from the perspective of sensors and effector. We use our experimental protein activity data from ZIKV and existing literature for related members of the NS3 helicase family to form a structure-activity relationship dataset. We formulate explainable machine learning based methodologies to study allostery in the NS3. Residue level propagation of allostery in NS3 helicase mutants are modeled using residue interaction networks. Using a classification model, we demonstrate that networks-level changes are predictive for protein function. A linkage metric, resi-SHAP is introduced to identify allosteric linkages within the NS3 helicase family. In the last part, we explore whether molecular simulations-based sampling and coarse-grained residue properties at the binding site are predictive for ATPase activity and RNA binding affinity of the NS3 helicase mutants. The metric, grid-SHAP allows us to visualize residue property changes at the binding site that enhances mutant function. Our method captures the effect of binding site and off-site mutations on NS3 activity. We assess how “universal” such a model is by reporting prediction performance of a model trained on Hepatitis C virus NS3 helicase on the Dengue virus, ZIKV and the Vasa helicase. The studies collectively demonstrate that explainable machine learning models can be built using features from networks-based analysis and coarse-grained simulations. These can bridge our understanding of protein structure and function as derived from experimental research, as demonstrated in the NS3 family of helicases.