Fig. 4

Schema identifies a gene set in granule neurons whose expression covaries with spatial cellular density. a Rodriques et al. [10] simultaneously assayed spatial and transcriptomic modalities in mouse cerebellum tissue (data from puck 180430_1 is shown here). In addition, they labeled beads (each corresponding to a transcriptome) with a putative cell type by comparing gene expression profiles with known cell-type markers. b Spatial distribution of the most common cell types in the tissue: granule cells, Purkinje cells, interneurons, and oligodendrocytes. Note the variation in spatial density for granule cells. c We quantified this spatial density variation by computing a two-dimensional Gaussian kernel density estimate, with cells in dense regions assigned a higher score. d Schema is able to identify a set of genes that are highly expressed only in densely packed granule cells. The four figures here show mutually disjoint sets of cells: granule cells with high expression of the gene set, granule cells with low expression of the gene set, other cells with high expression, and other cells with low expression. Here, a cell is said to have high expression of the gene set if the cell’s loading on this gene set ranks in the top quartile. e Schema’s results are robust across biological replicates. Across three replicates, we evaluated the consistency of gene rankings computed by Schema, canonical correlation analysis (CCA), SpatialDE, and Trendsceek. The black points indicate the Spearman rank correlation of gene scores across pairs of replicates. We needed to adapt SpatialDE and Trendsceek for this task by first applying them separately on granule and non-granule cells and then combining the results (Methods); here, the black and gray points indicate the cross-replicate correlations of the final and intermediate gene rankings, respectively