Bimodal distribution of normalised values

[nbartonicek]

Good day

I'm working with sparse data from a single-cell-like dataset (attached).
When I try to extract normalised counts with zinbwave for a set of controls, I'm getting a bimodial distribution centred on zero, but with no values at zero.

load("data.Rdata")
filtered_sce <- subset(data_sce,rowSums(data_sce) > 10)
zinb <- zinbwave(filtered_sce, K = 2,
                                 epsilon=1000,
                                 BPPARAM = p,
                                 normalizedValues=TRUE,
                                 residuals = TRUE)
normalised_values <- zinb@assays@data$normalizedValues
hist(normalised_values)

Despite the raw counts being fairly close, the normalised counts are in different parts of the binomial distribution, centred on zero. Am I doing something wrong?

filtered_sce@assays@data$counts["ISG15",1:5]
AACCAGCCAG AACCAGTTGA AACTCTACAC CCGGCCAATT CCGTCAGAAC
8 9 7 6 14
normalised_values["ISG15",1:5]
AACCAGCCAG AACCAGTTGA AACTCTACAC CCGGCCAATT CCGTCAGAAC
-2.125217 2.014329 -2.202011 -2.171598 2.425445

Cheers,

Nenad

data.Rdata.zip

[nbartonicek]

Hi @drisso,
This still presents a problem... Normalised values are bimodal compared to edgeR, and I have no idea what is causing thie artefact.