Fig. 6: Singularity scores inform the selection of the perplexity parameter.

Comparative t-SNE embeddings and the corresponding singularity scores at two different perplexities in mouse embryonic cell differentiation data (a) and in mouse mammary epithelial cell data (c). The perplexity as a tuning parameter has a large impact on t-SNE visualization qualitatively. At a small perplexity, there are many spurious sub-clusters. Embeddings with high singular scores appear in random locations, which indicates the presence of such spurious structures and severe FI discontinuity. Plots of the degree of FI discontinuity and neighborhood preservation versus perplexity are shown for mouse embryonic cell differentiation data (b) and for mouse mammary epithelial cell data (d). We recommend choosing a perplexity no smaller than the elbow point, as this ensures that randomly positioned points with high singularity scores largely disappear, remaining only at cluster peripheries. Consequently, the neighborhoods of most points are embedded more faithfully, resulting in better neighborhood preservation score. Source data are provided as a Source Data file.