Fig. 3
Cell-type-specific topic inference and deconvolution of pancreatic tissue. a Gene signatures of cell-type-specific topics in pancreas. We trained GTM-decon on the PI-Segerstolpe scRNA-seq dataset of pancreas tissue. We used 5 topics per cell type, allowing sub-cell-type inference. For the inferred genes-by-cell-type matrix \(\Phi\), we took the top 20 genes under each topic and visualized their topic distributions in heatmap. Whenever available from CellMarkerDB and PanglaoDB, cell-type marker genes are indicated on the left. For the cell types where marker genes are not available, “NA” were indicated on the left. The number of statistically significantly different genes in each cell type based on their topic scores (p-value < 0.05; permutation test) is shown below. b Gene set enrichment analysis (GSEA) of inferred topics based on known marker genes. Cell-type-specific topics for acinar, alpha, and beta were evaluated based on whether the top genes are enriched for the known marker genes under that cell type. The bar plots show the − log10 (p.adj values of enrichment score) for the gene set enrichment analysis for each of the 5 topics. The leading-edge plot for the topic with the best adj. p-value for that cell type is shown on the right. In each of the leading-edge plots, genes were ordered in decreasing order from left to right. The green curves indicate the running scores of enrichments. The barcode bars indicate cell-type marker genes. Adjusted p-values based on the GSEA enrichment scores are indicated in each panel. The three large panels display the most significantly enriched topic of among the five topics for each cell type and the 12 small panels display the remaining topics. c, e Deconvolution of bulk RNA-seq samples of 89 human pancreatic islet donors. The GTM-decon models separately trained on the Segerstolpe pancreas islet dataset (i.e., panel c) and Baron pancreas islet data (i.e., panel e) reference datasets were used to deconvolve the 89 bulk transcriptomes. As indicated by the legend, the 89 subjects consist of 51 normal, 15 impaired glucose tolerance, and 12 T2D individuals. In the heatmap, the rows represent subjects, and the columns represent cell types; the color intensity are proportional to the inferred cell-type proportions. d, f Deconvolved cell-type proportions as a function of Hemoglobin A1c (HbA1c) level. GTM-decon were trained on Segerstolpe (i.e., panel d) and Baron scRNA-seq (i.e., panel f) reference datasets. Each of the 10 panels displays a scatter plot of inferred cell-type proportion (y-axis) and HbA1c level (x-axis). The color legend indicates the 3 phenotypes. The heatmap on the right shows the deconvolved proportion of 3 most indicative cell types with subjects (rows) ordered on the basis of inferred beta cell-type proportions