Fig. 3

HIPPO framework applied to Zhengmix8eq data. a Computing time for each method using LAMBDA QUAD workstation with Intel Xeon W-2175 processor sequentially (non-parallel). b Computing time for HIPPO using different k. c HIPPO’s sequential clustering results for K=3,⋯,8. d t-SNE plots for clustering results from three methods: HIPPO, Seurat, and SCTransform, compared to true labels. Seurat and SCTransform cannot differentiate helper T/regulatory T and memory T cells. e Clustering results comparisons using Adjusted Rand Index. f Comparisons of features selected by different methods for their gene mean, CV, and variance. Seurat and SCTransform use CV as the selection criteria, and hence, their features weigh heavily on genes with small mean expression and variance