Fig. 4: Lower-dimensional UMAP representation of the transcriptome data restricted to the communities and conservation of global structure.
a UMAP generates embedded structures through a low-dimensional projection of the submatrix consisting of only the genes that belong to a community in at least one tissue (n = 3259). This subset of genes (17.7% of total) contains sufficient information to recover the tissue clusters. In addition, known relationships between tissues, based on organ membership and, separately, on shared function, are reflected in the UMAP projection. b Using bootstrapped manifolds (see “Methods” section), we estimated the persistence of the global structure and pairwise relationships across tissue clusters. Here, we show the upper-triangular matrix of the average pairwise distances across the bootstrapped manifolds. We found consistent clustering of known related tissues, including the 13 brain regions, the colonic and esophageal tissues, and various artery tissues. Additional patterns were observed. For example, as reflected in the heatmap, we found a highly correlated relationship, i.e., high "clustering conservation coefficient'' (\({C}_{({{i}}_{0},{{i}}_{1})}\approx 0.62\), p = 3.4 × 10−5) (see “Methods” section), of the two skin tissues to all the other tissues.
