`aggregateAcrossCells.se` doesn't support all `scuttle::aggregateAcrossCells` `statistics`

[HelenaLC]

Hi Aaron. Trying to work towards getting rid of scran/scuttle soon-ish as GHA are getting angry already. Sorry if I'm missing sth in advance, but I can't seem to figure out how to get anything other than sum and detected for aggregation (see title)? I'm thinking averaging logcounts for heatmaps, etc. (i.e., not just DE on pseudobulk sum of counts).

[LTLA]

Oh right, I never really used the means much. Frankly the easiest solution is to just do the usual trick:

library(Matrix)
x <- round(abs(rsparsematrix(1000, 100, 0.1) * 100))
clusters <- sample(LETTERS, 100, replace=TRUE)

agg <- aggregateAcrossCells(x, list(cluster=clusters))
means <- t(t(agg$sums)/agg$counts)

The aggregated matrix is usually so small that performance shouldn't be a major consideration.

While you're here, I should mention that there are actually many replacements for this function that I couldn't fit into a single deprecation message. The recommendation for scrapper only exists for end-users who are probably doing the rest of the scRNA-seq analysis pipeline so they'd have to install scrapper anyway. But if you're a developer and you don't want to add scrapper into your dependencies, here are a few other solutions.

If you already depend on beachmat, you can use some of the lower-level tatami utilities:

library(Matrix)
x <- round(abs(rsparsematrix(1000, 100, 0.1) * 100))
clusters <- sample(LETTERS, 100, replace=TRUE)

library(beachmat)
clusters <- factor(clusters)
sums <- tatami.sums.by.group(initializeCpp(x), as.integer(clusters), row=TRUE, num.threads=1)
counts <- tabulate(as.integer(clusters))
means <- t(t(sums)/counts)

If you don't want to depend on anything, you can use a neat multiplication trick in alanocallaghan/scater#225:

library(Matrix)
x <- round(abs(rsparsematrix(1000, 100, 0.1) * 100))
clusters <- sample(LETTERS, 100, replace=TRUE)

clusters <- factor(clusters)
group.id <- as.integer(clusters)
num.groups <- nlevels(clusters)
multiplier <- Matrix::sparseMatrix(i = seq_along(group.id), j = group.id, x = rep(1, length(group.id)), dims=c(ncol(x), num.groups))
sums <- x %*% multiplier
group.sizes <- tabulate(group.id, nbins=num.groups)
means <- t(t(sums) / group.sizes)

[HelenaLC]

Sweet! I think a couple cases scrapper it is; but for one we don't need much else and the matrix multiplication will do with no added deps-- thanks a lot a lot!!

[LTLA]

(forgot to mention, credit to @plger, who proposed it in the first place)

[HelenaLC]

And I assume doing (x > 0) %*% multiplier gives detected, and doing the mean on that prop.detected, right?

[LTLA]

Yes, that's correct.

For a bit more nuance: in theory, aggregateAcrossCells() will be faster than two multiplications as the former passes through the matrix once and computes both sums and detected. By comparison, when you do two multiplications, you pass through the matrix twice, which is a little slower as you need to pull everything from RAM into the CPU twice.

In practice, I don't think it matters much. Maybe if you've got a file-backed matrix, but those introduce other problems.

[SchulzDan]

Hi Aron,

I came across a similar issue. I am inputting an sce object into the aggregateAcrossCells.se function to return an sce object. However, it seems to ignore the compute.median = TRUE argument for me. The output always contains the sums and detected slot but misses the median.
Running like this:

cur_df <- DataFrame(spot_id = object[[spot_id]],
bin = unlist(cur_split))
cur_out <- scrapper::aggregateAcrossCells.se(sce, assay.type = "counts",
factors = cur_df, more.aggr.args = list(compute.median = TRUE))

Any idea?
Thanks and best,
Daniel

[LTLA]

@SchulzDan Oops, I forgot to propagate the result of scrapper::aggregateAcrossCells up to aggregateAcrossCells.se. This should be fixed in 1.5.19.