subsetting drops attributes from `reducedDim(., "PCA")`

[HelenaLC]

See title ... came across this working on some PCA-related things:

library(scater)
sce <- mockSCE()
sce <- logNormCounts(sce)
sce <- runPCA(sce)

pca <- reducedDim(sce, "PCA")
rot <- attr(pca, "rotation")
class(rot) # matrix

sub <- sce[, -1]
pcb <- reducedDim(sub, "PCA")
rot <- attr(pcb, "rotation")
class(rot) # NULL

Now, I'm aware this is not a SCE-specific thing, i.e., plain matrix subsetting does the trick:

names(attributes(pca))
# "dim" "dimnames" "varExplained" "percentVar" "rotation"
names(attributes(pca[1, ]))
# "dim" "dimnames"

However, I still wonder whether our "core" class for handling such data could somehow retain this information better? E.g., with rotations gone, it's impossible for me to tell what features underly the PCs in the object, etc.

Any thoughts, suggestions, further discussion appreciated!

[LTLA]

I usually just tuck the rotation matrix into the metadata() of the SCE itself. Not the "cleanest" solution as it separates the rotation matrix from its associated PCs, but that's not a real problem in many applications.

Another option is to use the LinearEmbeddingMatrix class, which is still in this package by virtue of having no better home. This is explicitly designed for your use case but it relies on downstream applications being aware of this class, at least enough to as.matrix() it before attempting to use it as an ordinary matrix.

There is another - currently unsupported - esoteric way to do this, which is to store the rotation vectors in the mcols() of the internal DataFrame that holds the reducedDims. This is a form of metadata that is "parallel" to the columns of the DataFrame, i.e., to the entries of the reducedDims. So, if you were to remove/duplicate/combine the PCA entries, you would automatically remove/duplicate/combine their rotation vectors as well. However, I think this is impossibly unintuitive.

As you've noticed, the attributes are generally too fragile for storage of anything meaningful. I only put stuff in there if I've programmed myself into a corner. For example, if I have a function that returns a matrix, and I need to return more information, but I want to preserve back-compatibility so I can't change the return value from a matrix into a list; I'll just stuff the extra bits and pieces into the matrix's attributes until I can muster up the courage to break back-compatibility.

[HelenaLC]

Okay, thank you! Agree attrs are flimsy at best. Will give both LinearEmbeddingMatrix and mcols a shot. Though the not-a-matrix part might be problematic for, e.g., scater::runUMAP or scran::buildSNNGraph? Let's see... Anything that let's me protect by PCs and leaves them ggplotable should be okay... So telling the user to run 1 extra line to assure that internally, that's acceptable enough. Thanks again!

(You can close this unless we see a bright future for more standardized handling of reduced dimensions... not sure how this would related to all things not PCA though tbh.)