Supplementary Figure 14: Dynamic filtering strategy for scRNA-seq quality control.

We selected two marker genes expected to be exclusively expressed in each of the four cell types: PDGFRA and ISLR for fibroblasts, CD2 and CD3D for T cells, CD79A and RALGPS2 for B cells, and CD14 and C1QA for monocytes. We counted nonzero expression of these genes in the correct cell type as a true positive and nonzero expression in the incorrect cell type as a false positive. a, Estimated optimal thresholds for reads per unique molecular identifier (UMI) shown for three example quadrants (Q2, Q1, Q4) of three scRNA-seq plates (S006, S008, S011). The reader per UMI threshold determines which UMIs we discard. By discarding UMIs with few supporting reads, we can reduce the false positive rate (FPR), but this comes at the cost of also reducing the true positive rate (TPR). b, After selecting the optimal threshold that maximizes the ratio of TPR to FPR for each quadrant, we visualize the FPR and TPR for each quadrant. We noticed that some quadrants had poor performance even after optimizing the reads per UMI, so we discarded the cells from the red outlier quadrants. c, An example of gene expression for MMP2 across all cells before and after filtering the reads per UMI threshold. This matrix metalloproteinase is known to be expressed by fibroblasts, but not by other cell types, so we did not expect to see MMP2 expression in cell types other than fibroblasts. After applying the reads per UMI threshold to discard UMIs that are probably contaminants, MMP2 gene expression is closer to our a priori expectation: very few B cells, monocytes, and T cells express this gene. Note that we did not use MMP2 to determine the optimal reads per UMI threshold.