Computational pipeline to probe NaV1.7 gain-of-function variants in neuropathic painful syndromes

Abstract

Applications of machine learning and graph theory techniques to neuroscience have witnessed an increased interest in the last decade due to the large data availability and unprecedented technology developments. Their employment to investigate the effect of mutational changes in genes encoding for proteins modulating the membrane of excitable cells, whose biological correlates are assessed at electrophysiological level, could provide useful predictive clues. We apply this concept to the analysis of variants in sodium channel NaV1.7 subunit found in patients with chronic painful syndromes, by the implementation of a dedicated computational pipeline empowering different and complementary techniques including homology modeling, network theory, and machine learning. By testing three templates of different origin and sequence identities, we provide an optimal condition for its use. Our findings reveal the usefulness of our computational pipeline in supporting the selection of candidates for cell electrophysiology assay and with potential clinical applications.

Figure 1

Snapshot of the Nav1.7 wild type. (a) Side view reporting high details of the Domain I: transmembrane helical segments S1, S2, and S3 are colored in green, segment S4 in yellow, segments S5 and S6 constituting the Pore Domain are identified by blue color, and the P-loop between S5 and S6 that helps to form the Selectivity Filter is colored in red. Thick dotted lines labelled as Intra and Extra identify the transmembrane region; (b) top view, Domain I is displayed in full colors as in (a), and the coordination of the four Domains is clearly visible. Here shaded blue, green and orange identify Domains II, III, and IV, respectively.

Figure 2

The computational pipeline. Starting from a template and 85 genetic variants, homology models are used to identify the corresponding three-dimensional structures, followed by energy minimization and quality assessment to refine them. RINs are then implemented to map them into their representing graphs and machine learning techniques are used to analyze them and identify patterns.

Figure 3

(Top) Primary structure and domains positions; (Bottom) Schematic illustration of the poly-peptide chain structure and localization of pathogenic mutations (PAT) associated with pain conditions (left) and not pathogenic variants (NEUTRAL) (right). Colored numbers (left) highlight the four different pathologies, color-coded according to the list shown in Fig. 4.

Figure 4

[Left] (a) PAT and NEUTRAL genetic variants. PAT mutations are further divided by disease and highlighted with different colors according to Fig. 3. Among the NEUTRAL variants, the 21 known human variants are highlighted in bold and the 4 not causing biophysical abnormalities are also starred. [Right] Mutation I136V: the initial Isoleucine (b) in position 136 is turned into a Valine (c).

Figure 5

DCA analysis for templates 6A90 and 6J8J where green points are representing the contact map of the original template and the red points the corresponding DCA representations. (a) and (b) full sequence; (c) and (d): interval 1181–1851 .

Figure 6

Quality assessments for the 6A90 Wild Type. (a) Predicted local similarity along the sequence: regions with high similarity are circled and correspond to the domains are DI-DIV; (b) Ramachandran plots where each point represents an amino acid and where characteristic regions are highlighted in different colours. Regions colored in white are considered forbidden.

Figure 7

[Left] Similarity matrices of the Weisfeiler–Lehman (5 iterations) (a)–(c)–(e) and Vertex Histogram (b)–(d)–(f) kernels applied to RINs resulting from MOESM3, 6A90 and 6J8J templates. [Right] Dominant set classification for the WL similarity matrix (c) of 6A90 template: first (g) and second (h) iterations.

Figure 8

Average-linkage hierarchical clustering of the 6A90 template. The leaves of the dendrogram are labeled with the ids of the considered genetic variants. The scale on the left shows the distance among variants. The input distance matrix has been derived from the WL kernel similarity matrix of the 6A90 template. The big red box highlights the cluster of the pain PAT mutations. The small red boxes highlight mutations ids 11 and 30.

Figure 9

Relative frequency of nodes for pathogenic (PAT) and non pathogenic (NEUTRAL) mutations. Vertical dotted lines identify regions of different behaviour between the two groups, color-coded according to their specificity (see text).

Figure 10

Entropy profile of the full alignment: red dots highlight PAT mutations and blue dots the NEUTRAL ones. The horizontal line is the mean entropy value of the alignment in the mutated sites.