Activity regulation of many proteins involves small changes in structure that are comparable to thermal fluctuations. Consequently, their regulatory mechanisms cannot be modeled in terms of how their energy-minimum structures differ between states. Understanding their regulatory mechanisms requires assessment of conformational ensembles of different states. Toward this end, we presented previously methods to quantify and statistically analyze differences between conformational ensembles of states (JCTC 9:868-875 2013; BJ 111:1621-1630 2016). Here we present a new machine learning based method that enables determination of inter-site correlations in ensemble shifts. These pairwise correlations can then be combined with each other and with the spatial organization of sites to discern regulatory signaling networks. In this study, we employ a graph theory approach to combine pairwise correlations and spatial organization. We apply this development to study the regulation of hPTP1E's PDZ2 domain by RA-GEF2 peptide; a model system for studying dynamic allostery. We generate conformational ensembles of the apo and GEF2-bound states using molecular dynamics, and the employed protocol yields deuterium order parameters consistent with NMR. We find that the allosteric regulatory networks, which we determine using correlations in ensemble shifts, are strikingly different from the allosteric communication networks present in individual apo and GEF2-bound states. Consequently, a residue's role in regulation cannot be projected necessarily from its contribution to signaling in a given state. We also find that the regulatory network emerges from a combination of changes in structure and dynamics, and even small changes in structure and dynamics contribute significantly to regulation. Finally, we find that the ensembles of the apo and GEF2-bound states are asymmetric about their overlap region, which produces dissimilar GEF2-binding and unbinding response signaling networks. While additional studies will determine if these findings are broadly applicable, this work reports a method for addressing such key issues in dynamic allostery.

Machine learning enabled approach to incorporate multi-state information in molecular modeling of dynamic allostery: A case study of the PDZ2 domain

254th National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.