Living Neural Networks: Dynamic Network Analysis of Developing Neural Progenitor Cells

The architecture of the mammalian brain has been characterized through decades of innovation in the field of network neuroscience. However, the assembly of the brain from progenitor cells is an immensely complex process, and a quantitative understanding of how neural progenitor cells (NPCs) form neural networks has proven elusive. Here, we introduce a method that integrates graph-theory with long-term imaging of differentiating human NPCs to characterize the evolution of spatial and functional network features in NPCs during the formation of neural networks in vitro. We find that the rise and fall in spatial network efficiency is a characteristic feature of the transition from immature NPC networks to mature neural networks. Furthermore, networks at intermediate stages of differentiation that display high spatial network efficiency also show high levels of network-wide spontaneous electrical activity. These results support the view that network-wide signaling in immature progenitor cells gives way to a hierarchical form of communication in mature neural networks. We also leverage graph theory to study the spatial features of individual cell types in developing cultures, uncovering spatial features of polarized neuroepithelium. Finally, we employ our method to uncover aberrant network features in a neurodevelopmental disorder using induced pluripotent stem cell (iPSC) models. The “Living Neural Networks” method bridges the gap between developmental neurobiology and network neuroscience, and offers insight into the relationship between developing and mature neural networks.


INTRODUCTION 30
The study of complex, multiscale brain networks using concepts from graph theory and network science 31 -an approach collectively termed network neuroscience -has enabled significant insight into the 32 structural and functional organization of the brain 1,2 . Micro-connectomics -the study of organizational 33 principles of neuronal networks at the cellular scale 3,4 , is an important subset of complex brain networks 34 that has yielded insight into architectural features of the nervous system at the level of their basic building

RESULTS 82
Functional characterization and spatial network representation of differentiating NPC cultures.

83
In the first part of this study, we used primary hNP1 neural progenitor cells derived from H9 human 84 embryonic stem cells. These cells were maintained as undifferentiated, mitotic progenitor cells in the 85 Figure 2. Functional characterization and spatial network representation of differentiating NPCs. (a) hNP1 cells at day 0 stain positively for Nestin. (b) Cells at day 14 stain positive for MAP2. In (a-b) nuclei are labeled by Hoescht; scale bar = 100μm. (c) Peak inward and outward currents determined through whole-cell patch clamp electrophysiology. Sample sizes: n=17, n=25, n=33 cells recorded for day 0-2, day 4-8 and day 10-14 respectively. Error bars represent SEM; *p < 0.05 from two-sample t-test. (d) Voltagegated inward and outward currents seen in a cell at day 14. Voltage steps applied were from -60mV to +90mV in 10mV increments. (e) Weak action potentials evoked from the same cell through current injection. Magnitudes of current injected are -30pA, +20pA and +120pA from holding. (f) Representative phase contrast image of hNP1 cells, shown at day 3; scale bar = 50 μm. (g) First derivative of the pixel intensity histogram, with a linear fit to the ascending portion shown as a red line. The point where this line met the x-axis was used as a threshold for segmentation. (h) Binary image obtained upon thresholding the grayscale image. (i) Separation of linear features through morphological opening of the binary image yields cell bodies (blue) and neurites (red). (j) Phase contrast image from (f) with soma boundaries overlaid in red, and proximity edges shown in yellow. Inset shows six soma, of which two pairs (1, 2) and (4, 5) are connected by proximity edges; the intercellular distance for these two pairs are smaller than their average diameter multiplied by a scaling factor S = 2; Soma 3 and 6 are isolated nodes since they are not sufficiently close to any other soma. All microscope images are displayed with enhanced contrast for easy visualization. In order to uncover topological changes in differentiating hNP1 cells, we combined long-term imaging of 104 differentiating cultures with a graph-based approach for quantifying cell community structure.

111
We built network representations of spatial topology by denoting cell soma as nodes and using spatial 112 proximity between soma to assign edges (Figure 2j, k)

117
Structure and information flow in NPC spatial graphs

118
In order to describe the structure and topology of hNP1 spatial graphs, we evaluated 17 metrics derived 119 from graph theory that were computed and normalized appropriately to account for network size (Table   120 1) 24 . The network metrics provide information on various aspects of the graph structure such as 121 information flow, connectivity and abundance of motifs (repeating patterns of cell arrangements).

122
Through analyses of the covariance matrix of 17 metrics via hierarchical clustering, we were able to 123 identify several strong positive correlations among degree-related metrics including average degree, 124 average neighbor degree and degree variances (Figure 3a). We also identified negative metric correlations 125 including those between network efficiency and number of connected components, as well as between 126 clustering coefficient and all degree-related metrics. In the following section, we focus on metrics that

165
We performed calcium imaging using the fluorescent calcium indicator Fluo-4 to record spontaneous 166 activity in differentiating ReNcell VM cultures at days 1, 3 and 5, and employed cross-correlation analysis 167 to infer functional connectivity in the networks (Figure 4a, Supplementary Figure 7). Analysis of functional 168 networks revealed that cultures at day 3 had significantly more activity than those at days 1 and 5, as 169 measured by the fraction of active cells (Figure 4b, Supplementary Videos 3-6). Interestingly, the 170 functional network was not restricted to cells with short intercellular distances, with cells in the whole field of view (832μm x 702μm) having highly correlated calcium activity (Figure 4c) Spatial networks overlaid on immunofluorescence images of nuclei stained with Hoescht dye; scaling factor = 3. The nucleus images correspond to the images shown in (a). (e) Network efficiency of spatial networks peaks at day 3. Number of connected components shows the inverse trend. Sample sizes: Day 1 (n=5); Day 3 (n=8); Day 5 (n=5) for all plots. Red notches show mean and standard deviation; *p < 0.00029 from two-sample t-test (significance threshold adjusted using Bonferroni correction for 17 statistical tests to 0.005/17 = 0.00029). All microscope images are displayed with enhanced contrast for easy visualization.
We next built spatial graphs using nucleus images from the same cultures in which calcium imaging was connected components (Figure 4e) (Figure 1b, Supplementary Video 8-10). This analysis revealed that high-spiking cells had a 204 greater proportion of Tuj1-/Ki67-cells and a lower degree than low-spiking cells (Figure 5e,f) (Figure 6a-d, Table 1). Our analyses both confirmed accelerated neural specification in CWI 4F2 cultures through day 3 and 220 revealed that Tuj1+ neurons in day 3 CWI 4F2 cultures did not have a high clustering coefficient compared 221 to Tuj1-cells, as was the case in NCRM-5 cultures (Figure 6e, f). This was due to the presence of many 222 more 'lone' neurons with higher neurite extensions in the CWI 4F2 cultures (Figure 6g)

445
Creation of spatial graphs. Spatial graphs were created from microscope images using cytoNet, software 446 developed in-house 43 . For each pair of objects (soma/nuclei), a threshold distance for proximity was 447 defined as the average of the two object diameters, multiplied by a scaling factor (S). If the Euclidean 448 distance between the object centroids was lower than the threshold distance computed, then the pair of 449 objects was connected with a "proximity edge" (Figure 2j, k). We chose a scaling factor of 2 for phase 450 contrast images and 3 for nucleus immunofluorescence images based on similarity in network density for 451 the resulting networks (Supplementary Figure 2).

452
Due to the high density of NCRM-5 cultures, quantification of global network metrics proved unfeasible 453 (Supplementary Figure 8). However, qualitatively we observed the prevalence of highly clustered cell 454 bodies at late stages of differentiation. and rich-club metric (Supplementary Figure 4)

471
Single-cell analysis. Consolidated multi-parametric datasets were obtained by performing calcium 472 imaging followed by immunocytochemistry (Supplementary Videos 8-10). Functional data obtained 473 through calcium imaging (e.g., number of spikes) was combined with cell identity information obtained 474 through immunostaining (e.g., Ki67, Tuj1 status), and spatial features extracted using nuclei as described

Supplementary Figure 7
Cross-correlation analysis to infer functional connectivity from calcium imaging data