Fig. 8

Coverage requirements for clustering based on transcript-compatibility counts. As an intermediate between raw reads and quantified transcript abundances, transcript-compatibility counts intuitively have more information than the reconstructed transcripts and less noise than the raw reads. a As the read coverage of a cell of the dataset of [7] decreases from approximately 627k mapped reads, different methods have varying robustness to the loss of coverage. Each method was evaluated on its ability to cluster 200 randomly selected neurons mixed with 200 randomly selected non-neurons into the two cell types (the clustering of [7] being considered as the ground truth). Among methods which do not explicitly account for PCR bias, TCC based clustering performed much better than kallisto and eXpress and was quite close in performance to the UMI counting method of [7]. For both kallisto and eXpress, 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 counting the number of unique UMIs rather than reads (TCC with UMI in the plot), transcript-compatibility based clustering was adapted to account for PCR bias, resulting in similar performance to that of gene-level UMI counting used in [7]. b Even at significantly decreased coverage depths, our method maintains clusters corresponding to the nine major cell types identified by Zeisel et al. The transcript-compatibility distribution matrices at varying coverage depths are visualized using t-SNE. c At various coverage depths, transcript-compatibility counting with UMIs disagrees slightly with the cells the authors labeled as Oligo1 (cyan cell IDs). As the coverage decreases, transcript-compatibility based affinity propagation still identifies a cluster that captures the vast majority of Oligo1 cells in the 3005-cell population