Fig. 2.

Current normalizations are inconsistent with UMI counts and do not remove distributional variation across biological replicates and batches. a Kernel densities of nonzero read count values for 20 random cells from individual NA19098, replicate r3 of the Tung dataset. b UMI counts from the same cells as a. c Quasi-UMI counts (Poisson-lognormal with shape 2.0) computed from read counts of the same cells. d As a but normalized to counts per million. e As a but normalized to census counts. f As a but normalized with scran. g Kernel densities of nonzero read count values for 20 random cells from each batch of tumor MGH26 in the Patel dataset. Color indicates batch. h As g but normalized to transcripts per million. i As g but normalized to census counts