Normalize based on the 99.99th percentile

[cr2106]

Hi!

After arcsinh normalization of flow data using prepData, is it possible to normalize based on the 99.99th percentile to have everything between 0 and 1?

Thank you very much for your work!

[HelenaLC]

Jup, should be easy using base R commands... Luckily, there's a little helper in CATALYST that is used for some heatmap visualizations, which you could use via CATALYST:::.scale_exprs (here, x is a matrix of expression values, margin specifies whether to scale rows (1) or columns (2), and q defines the quantiles). You'd then assign the output of this to a new assay, e.g., assay(sce, "scaled") <- CATALYST:::.scale_exprs(assay(sce, "exprs"), ...).

# CATALYST:::.scale_exprs
function (x, margin = 1, q = 0.01) {
    if (!is(x, "matrix")) 
        x <- as.matrix(x)
    qs <- c(rowQuantiles, colQuantiles)[[margin]]
    qs <- qs(x, probs = c(q, 1 - q))
    qs <- matrix(qs, ncol = 2)
    x <- switch(margin, `1` = (x - qs[, 1])/(qs[, 2] - qs[, 1]), 
        `2` = t((t(x) - qs[, 1])/(qs[, 2] - qs[, 1])))
    x[x < 0 | is.na(x)] <- 0
    x[x > 1] <- 1
    return(x)
}

[Ivy-ops]

Hi @HelenaLC , Thanks for the nice package, which is pretty helpful to the community. I have searched the issue related to .scale_exprs() and found this post, thus am asking here. Do you have any specific reason for the last 2 lines of code for covert values greater than 1 into 1 and less than 0 or is NA into 0? x[x < 0 | is.na(x)] <- 0 x[x > 1] <- 1

Thank you!

[HelenaLC]

I'd say that

• we want to scale between 0 and 1 using q and 1-q *100% quantiles as boundaries, say these are A and B, so per definition
• any x < A will become < 0, and we set that to 0
• any x > B will become > 1, and we set that to 1
• the is.na(x) is a fail-safe/for security, so that missing values become 0, which could probably be argued about, but shouldn't happen much in CyTOF data in any case (as far as I know, no missing values per se in the measurement data)

[SamGG]

I fully agree. Say differently, we want the color range to reflect the intensity range for each marker and accept to ignore the lower and higher percentiles (q being 1% as default). Ignoring values outside the [0..1] range could be achieved via thresholding the data as performed in .scale_exprs(). Alternatively, pheatmap allows to assign limits to the color scale (ComplexHeatmap may provide this also), which results in the same display. Best.