Skip to main content
Fig. 2 | BMC Bioinformatics

Fig. 2

From: CoSTA: unsupervised convolutional neural network learning for spatial transcriptomics analysis

Fig. 2

Analyzing MERFISH data with the CoSTA approach. A and B Visualization of the spatial feature vectors obtained for each gene, blank controls, and reference cell type patterns from MERFISH data in a 2D UMAP layout. Filled circles represent true genes, filled triangles represent cell-types, and filled squares are blank controls. A features extracted from a randomly initialized ConvNet with no training. Each dot is a gene, blank control, or cell type pattern. Colors indicate cluster labels obtained from clustering on the full CoSTA-derived feature vectors; B features extracted by trained ConvNet. Each dot is colored with the original clustering labels from a to show how some cluster memberships rearrange. C Local intrinsic dimensionalities of spatial representations by CoSTA without and after training (10 independent runs of CoSTA). D CoSTA-detected spatial correlations of genes identified as SE uniquely by SPARK. Top row displays known cell type specific expression patterns for 3 cell types. Lower rows display expression patterns of particular genes. Dotted lines indicate the CoSTA-identified similarity between a pair of genes or a gene with a cell type pattern. Raw count values for each image are scaled from 0 to 1 to normalize the visual comparison. E Euclidean distances between each gene shown in D and Ependymal or Mature OD patterns. Euclidean distances are measured using the CoSTA spatial representation

Back to article page