Velocyto, dropEst and STARsolo

CellRanger (v3.0.2) [4] and velocyto (v0.17) [9] were run with default settings to generate exonic and intronic counts based on the CellRanger index. Also dropEst (v0.8.6) [20] was run on the BAM files output by CellRanger, specifying the -V flag to instruct dropEst to return exonic, intronic and exon-intron spanning UMI count matrices. The intronic and exon-intron spanning UMI counts were summed and used to represent the unspliced UMI count. STARsolo (v2.7.3a) [17] was run using the STAR index, specifying the SOLOfeatures argument to generate ‘Velocity’ (exonic and intronic), ‘Gene’ (regular exonic gene expression) and ‘GeneFull’ (reads with any overlap with the gene locus) counts. Based on these count matrices, we obtained exonic and intronic count matrices in two different ways. First, we directly used the ‘Velocity’ count matrices as exonic and intronic counts (below referred to as starsolo). Second, we used the ‘Gene’ count matrix as the exonic counts, and the difference between the ‘GeneFull’ and ‘Gene’ counts as the intronic counts (below referred to as starsolo_diff). For genes where the ‘Gene’ counts were higher than the ‘GeneFull’ counts, the intronic count was set to zero. This can happen, for example, for a gene located in the intron of another gene. In the ‘GeneFull’ quantification, reads mapping to such a gene are considered ambiguous and therefore discarded. However, they may be assigned in the ‘Gene’ quantification, if they are compatible with the annotated gene model. An overview of the evaluated quantification approaches is provided in Table 3.