Fig. 2

Implementation of BarCounter and BarMixer. A Trie data structures are used in BarCounter to efficiently tabulate barcode frequencies. B Diagram of HTO read structure. Read 1 contains the 16 bp cell barcode and the 12 bp UMI. Read 2 contains the 15 bp hashtag. C Overview of the BarCounter workflow. At runtime, the user provides a barcode whitelist which is loaded into a trie for rapid lookups, a taglist containing all valid hashtag sequences, and paired Read 1 and Read 2 FASTQ files. For each read, checks are performed to verify the cell barcode exists in the barcode trie, and the hashtag sequence is in the taglist. The UMI sequence is checked against a trie and if it is not present, the trie is updated and the counts for the barcode and hashtag combination are incremented. D–F Benchmarking comparisons of BarCounter and other available HTO counting algorithms as a function of increasing cell loading per 10x Genomics well: cellranger count (10x Genomics); CITE-seq-Count (with or without barcode correction, [15], KITE, [4] (single-threaded or with 8 threads). D Maximum memory usage, E Average CPU load, F Elapsed time. G Overview of the barcode cutoff determination method used by BarMixer: Raw counts generated by BarCounter are clipped to remove low values, then log transformed, and used as input to 2-cluster K-means. If cluster medians are separable, the cutoff is set to the lowest value in the positive cluster. Note broken y-axis in the first two panels. H–L Visualizations provided by BarMixer QC reporting notebooks. H HTO count histograms (green bars) with cutoff values (blue lines), I Fractions of barcodes and reads attributed to singlets (dark blue), doublets (light blue), and multiplets (purple), J Counts of cells in each hashing category per well in a batch, K Number of UMIs per cell in each HTO category, L Number of genes per cell in each HTO category