Fig. 6

Clustering mouse brain cells based on transcript-compatibility counts. a The TCC distribution and gene expression matrices for the 3005 mouse brain cells are visualized using t-SNE (based on Jensen-Shannon distances between TCC distributions and gene expression distributions of cells, respectively) and colored with the cell type determined by Zeisel et al. [7]. We note that the transcript-compatibility based t-SNE also visually maintains the cluster structure of the nine major clusters, even though it can be computed two orders of magnitude faster than the gene expression matrix. b Cells from each of two cell types determined by Zeisel et al. were randomly selected, and then the clustering accuracy of multiple methods was tested. The clustering accuracy was measured as the error rate of the clustering. First, we note that the 3’-end bias in this dataset significantly affects the accuracy of kallisto and eXpress that have been chosen here as representative methods for model-based quantification (see Methods). For both methods clustering was performed on gene expression profiles obtained by summing the corresponding transcript abundances. For each point in the eXpress and kallisto curves, we took the minimum of the error rates obtained with bias modeling turned on and off. By avoiding estimation of the read model, transcript-compatibility based methods were indeed more accurate. We see that transcript-compatibility based clustering achieves similar accuracy to the gene-level UMI counting method implemented by the authors for this dataset without explicitly accounting for PCR biases. Refining transcript-compatibility counting to correct for PCR biases (by counting only the distinct UMIs of reads in each equivalence class) leads to a marginal improvement of our method. c Running affinity propagation on the TCC distribution matrix (using negative Jensen-Shannon distance as similarity metric) produced a cluster of 28 cells, 24 of which were labeled Oligo1. Zeisel et al. [7] classified 45 of the 3005 cells as this new class of cells. The bar plot compares the mean expression of selected oligodendrocyte marker genes in the TCC cluster to their mean expression in Zeisel et al.’s Oligo1. As reported in [7], Oligo1 cells are characterized by their distinct expression of genes such as Itpr2, Rnf122, Idh1, and Gpr17. The similarity of the bars seems to suggest that clustering on TCC can capture this fine-grained information. Note that although single-cell clustering was entirely performed based on transcript-compatibility counts, the gene expression data used to evaluate this figure were obtained from Zeisel et al