aggregateAcrossCells: how to provide a custom function?

[SamGG]

Thanks for the package.
I want to compute the 1st and 3rd quantiles per gene for each sample by grouping their cells. How to perform it?
Best.

[LTLA]

You won't be able to do that via aggregateAcrossCells; part of it calls into C++ code, so the types of aggregation statistic are limited to what I decided to implement.

A less efficient but more general approach can be achieved by using DelayedMatrixStats:

X <- abs(Matrix::rsparsematrix(1000, 100, density=0.1))
f <- factor(sample(LETTERS, 100, replace=TRUE))

output1 <- output3 <- matrix(0, nrow(X), nlevels(f))
colnames(output1) <- colnames(output3) <- levels(f)

library(DelayedMatrixStats)
for (l in levels(f)) {
    res <- rowQuantiles(X, cols=(f == l), probs=c(0.25, 0.75))
    output1[,l] <- res[,"25%"]
    output3[,l] <- res[,"75%"]
}

In practice, for single-cell data, I find the quantiles to be pretty useless. At low densities, the majority of the quantiles are usually zero, which is not very informative. I don't think I've even used the statistics="median" mode in a real analysis.

In any case, the functionality of scuttle::aggregateAcrossCells() will eventually be moved to scrapper::aggregateAcrossCells(), so that's the place to go to request new features. You are free to propose that quantiles should be added to the latter, but I would be skeptical of the value of that proposition.

[SamGG]

My current insterest concerns cytometry and you know how distributions are different from seq analysis.

Thanks a lot for your feedback, and especially the code I would never have find.

I will validate my code on a few datasets before coming back. Best.

Here is the link to the issue/question I opened in CATALYST HelenaLC/CATALYST#424 @HelenaLC

Feel free to close the issue if you wish.