Fig. 7.

The UMAP plots comparing clustering performance in the Zhengmix8eq data set [29] using different data representations and clustering methods. Panel a displays the FACS labels we used as a benchmark to measure clustering performance. Both the Seurat pipeline (panel b and our Hclust-Departure pipeline (panel c) correctly identify the distinct cell types but fail to distinguish the subtypes within the T cells. Panel d uses the DIPD-based data matrix as data representation combined with Louvain clustering, which is a more direct comparison with panel b since 15 PCs and a resolution of 0.8 are used in both cases. It improves the original Seurat clustering performance by better distinguishing T memory cells from T helper/regulatory cells