PyMINEr Finds Gene and Autocrine-Paracrine Networks from Human Islet scRNA-Seq

Abstract

SUMMARY Toolsets available for in-depth analysis of scRNA-seq datasets by biologists with little informatics experience is limited. Here, we describe an informatics tool (PyMINEr) that fully automates cell type identification, cell type-specific pathway analyses, graph theory-based analysis of gene regulation, and detection of autocrine-paracrine signaling networks in silico . We applied PyMINEr to interrogate human pancreatic islet scRNA-seq datasets and discovered several features of co-expression graphs, including concordance of scRNA-seq-graph structure with both protein-protein interactions and 3D genomic architecture, association of high-connectivity and low-expression genes with cell type enrichment, and potential for the graph structure to clarify potential etiologies of enigmatic disease-associated variants. We further created a consensus co-expression network and autocrine-paracrine signaling networks within and across islet cell types from seven datasets. PyMINEr correctly identified changes in BMP-WNT signaling associated with cystic fibrosis pancreatic acinar cell loss. This proof-of-principle study demonstrates that the PyMINEr framework will be a valuable resource for scRNA-seq analyses.

talklr uncovers ligand-receptor mediated intercellular crosstalk

“…Ligand-receptor interactions play a critical role in inter-cellular communication. Many single cell RNA-seq studies examine the expression levels of ligand-receptor pairs among different cell types in a tissue to gain biological insights: uncover cell type(s) that are hubs of inter-cellular communications and reveal paracrine/autocrine signaling between cell types [6][7][8][9][10][11] , many of which are then experimentally validated. For example, ligand-receptor interactions in the NF-κB, FGFR, IGF1R, and JAK pathways were validated as regulators of liver bud development after analyzing the expression patterns of over 2,500 ligand-receptor pairs 6 among three major cell types.…”

“…The KL divergence between the observed and reference interaction score distributions is 0.417. The right panel of Figure 1A would be missed by differentially-expressed gene based approaches 7,11 , because the receptor is highly expressed in two cell types, instead of just one cell type. However, the KL divergence for the right panel is 1.956.…”

“…To demonstrate the performance of talklr in the single condition scenario, we applied it on three datasets: bulk RNA-seq of three FACS-sorted cell types (podocytes, endothelial cells, mesangial cells) in the kidney glomerulus 15 , single cell RNA-seq of adult mouse heart from the Tabula Muris project 13 , and bulk RNA-seq of FACS-sorted neurons, astrocytes and microglia from mouse cortex 16 . We compared the performance of talklr to a commonly used workflow, where both the ligand and receptor are required to be enriched only in one cell type ("DEG pairs") 7,11 . In each tissue, we first identified k ligand-receptor pairs where both genes are enriched in only one cell type.…”

“…This is followed by counting how many co-detected ligandreceptor gene pairs exist between each pair of cell types, and the number of cell-cell interactions each ligand-receptor gene pair participates in 6,10 . In the second approach, when both the ligand and the receptor are specifically enriched in one cell type (not necessarily the same cell type), a connection is drawn between this pair of cell types 7,8,11 . The third approach first quantifies ligand-receptor interactions by multiplying the expression level of the ligand in cell type i with the expression of the cognate receptor in cell type j to obtain an interaction score, and then prioritizes the ligand-receptor gene pairs with large interaction scores 9,12 .…”

“…In other words, the observed interaction scores are highly concentrated in a few of the n 2 possible cell-cell interactions. Compared to existing approaches that often only focus on cell type-specific genes 7,11 (the expression of a gene is enriched only in one cell type), or ignore cell type-specificity 6,9 , talklr accommodates all possible expression patterns of the ligand-receptor pair. Cell-type specific expression of individual ligand/receptor genes is sufficient but not necessary for high specificity of the interaction score distribution.…”