Machine learning approaches to evaluate correlation patterns in allosteric signaling: A case study of the PDZ2 domain

Abstract

Many proteins are regulated by dynamic allostery wherein regulator-induced changes in structure are comparable with thermal fluctuations. Consequently, understanding their mechanisms requires assessment of relationships between and within conformational ensembles of different states. Here we show how machine learning based approaches can be used to simplify this high-dimensional data mining task and also obtain mechanistic insight. In particular, we use these approaches to investigate two fundamental questions in dynamic allostery. First, how do regulators modify inter-site correlations in conformational fluctuations (C)? Second, how are regulator-induced shifts in conformational ensembles at two different sites in a protein related to each other? We address these questions in the context of the human protein tyrosine phosphatase 1E's PDZ2 domain, which is a model protein for studying dynamic allostery. We use molecular dynamics to generate conformational ensembles of the PDZ2 domain in both the regulator-bound and regulator-free states. The employed protocol reproduces methyl deuterium order parameters from NMR. Results from unsupervised clustering of C combined with flow analyses of weighted graphs of C show that regulator binding significantly alters the global signaling network in the protein; however, not by altering the spatial arrangement of strongly interacting amino acid clusters but by modifying the connectivity between clusters. Additionally, we find that regulator-induced shifts in conformational ensembles, which we evaluate by repartitioning ensembles using supervised learning, are, in fact, correlated. This correlation Δ is less extensive compared to C, but in contrast to C, Δ depends inversely on the distance from the regulator binding site. Assuming that Δ is an indicator of the transduction of the regulatory signal leads to the conclusion that the regulatory signal weakens with distance from the regulatory site. Overall, this work provides new approaches to analyze high-dimensional molecular simulation data and also presents applications that yield new insight into dynamic allostery.

Molecular basis for SARS-CoV-2 spike affinity for human ACE2 receptor

“…Given existing evidence on the general reliability of employed MD methods [13,[18][19][20], and our recent work [21] that yielded validated predictions on virus-host protein-protein interactions [22], we consider our qualitative conclusions to be robust. Nevertheless, from the perspective of intermolecular interaction theory, the underlying potential energy functions that we employ do rely on describing interactions using point charges, no polarization, and only pairwise vdW interactions.…”

“…The advantage of the affinity propagation algorithm over traditional clustering approaches is that it does not assume a priori the number of clusters or a cutoff value for delineating clusters. We adopted this unsupervised machine learning algorithm previously to cluster correlations in structural fluctuations [13].…”

“…Binding modes are identified by clustering conformations extracted every nanosecond from MD. Conformational clustering is performed using affinity propagation [11,13] in which we take RMSD as an index of similarity between conformations. (c) and (d) show these pairwise RMSDs.…”

“…We use Amber99sb-ILDN parameters to describe protein and ions [19], and SPC/E parameters to describe water molecules [31]. This force field has been demonstrated to perform well in reproducing structural data from X-ray diffraction and NMR spectroscopy [19,20], and also dynamics data from NMR spectroscopy [13].…”

Drug design targeting active posttranslational modification protein isoforms

“…ML is known for its capability of capturing the hidden pattern and therefore it will help uncover the structural changes induced by PTMs along with MD simulations or other computational methods. ML tools also enable the identification of the allosteric binding site 304–314 …”