Fig. 3

Performance of deMULTIplex2 on real datasets. A Heatmap summarizing the F-score of deMULTIplex2 and other methods on 9 real datasets. Stoeckius(C) and Stoeckius(P) are cell line and multi-donor PBMC datasets from Stoeckius et al. [3]. McGinnis(M) and McGinnis(S) are MULTI-seq and SCMK datasets from McGinnis et al. [28]. BAL1, 2, 3 are three batches of multi-donor bronchoalveolar lavage (BAL) datasets from Howitt et al. and Maksimovic et al. [29, 30]. The lung cell line dataset is also from Howitt et al. [30]. NA indicates the method cannot be run on the corresponding datasets due to the unavailability of mRNA count matrix or an error (i.e., demuxEM returns an error on the MULTI-seq PBMC dataset). B UMAP computed with deMULTIplex2-computed RQR for the MULTI-seq and SCMK datasets from McGinnis et al. [28], colored by donor ID predicted by deMULTIplex2. C Concordance between deMULTIplex2-predicted donor ID and the true donor ID based on SNP-based sample classification using souporcell [12]. D Performance of deMULTIplex2 and other methods on each sample. Mean values are highlighted with the diamond points. E Multiclass ROC curve of deMULTIplex2 and the two modes of demuxmix. False positive rate (FPR) and true positive rate (TPR) were computed for all samples using a one-vs-rest scheme and averaged to generate the ROC curve. F deMULTIplex and deMULTIplex2 recovered cells in the gene expression space. Percentage of correctly classified singlets are highlighted for each of the cell type. G Classification accuracy of each cell type across methods. H Performance of all methods on the MULTI-Seq dataset from McGinnis et al. [28] with down-sampled reads. demuxEM returns an error and is excluded from this analysis. I Performance of all methods on the SCMK dataset from McGinnis et al. [28] with down-sampled reads