Why does colnames(mae) return a character list?

[LTLA]

This is not expected:

library(MultiAssayExperiment)

# Making up some data.
mat <- matrix(rpois(10000, 5), nrow=10)
rownames(mat) <- paste0("CD", seq_len(nrow(mat))) # made-up markers
colnames(mat) <- paste0("BARCODE_", seq_len(ncol(mat))) # made-up barcodes

# First building the SE. Assuming that barcodes come from a range of donors.
donor <- rep(LETTERS[1:10], each=100)
se <- SummarizedExperiment(list(counts=mat))
se$Donor <- donor

# Now building the MAE. Each 'column' of the MAE corresponds to a donor;
# this is specified by the sample mapping from donors (primary) to barcodes (colname).
mae <- MultiAssayExperiment(list(markers=se),
    sampleMap=data.frame(assay="markers", primary=se$Donor, colname=colnames(se)),
    colData=DataFrame(row.names=LETTERS[1:10])
)
colnames(mae)
## CharacterList of length 1
## [["markers"]] BARCODE_1 BARCODE_2 BARCODE_3 ... BARCODE_999 BARCODE_1000

What I really want requires an unintuitive extra step:

rownames(colData(mae))
## [1] "A" "B" "C" "D" "E" "F" "G" "H" "I" "J"

I don't understand why these two calls are different. If the colData of an object has row names, those should be the colnames of the object by definition. If I wanted the column names of the individual SEs, I'd look them up individually.

[lwaldron]

This went through quite a bit of discussion back in the day, because there are opposing perspectives on what rownames and colnames should return and it's still not clear to me that there's a right answer. As a guide we turned to SummarizedExperiment as much as possible, and SummarizedExperiment does require rownames(colData(mae)) to return the rownames of the colData (below). It was from these behaviors of SummarizedExperiment that we decided that the CharacterLists returned were most conceptually similar to behavior expected from SummarizedExperiment. That said, there is also the usual contract that rownames and colnames return atomic vectors or NULL, and from that perspective perhaps they should just return an error or NULL for MultiAssayExperiment, but that seems like taking away functionality to make nobody happy.

Like you point out, for a SE, rownames(colData(se)) is identical to colnames(se), but that is a naming convenience possible because of the 1:1 relationship between colData and assay data for SummarizedExperiment. It breaks down for MultiAssayExperiment because each assay can have different samples that do not need match 1:1 to rownames(colData(se)). So that unintuitive difference does arise from a fundamental difference in the kinds of experiments these objects can represent. I'm still unconvinced that either of these solutions is "correct", and the most "correct" solution still may be to take away rownames and colnames with an error message pointing to two related functions with these two different behaviors. I'll move this issue onto Discussions for now...

suppressPackageStartupMessages(library(SummarizedExperiment))
mat <- matrix(rpois(10000, 5), nrow=10)
rownames(mat) <- paste0("CD", seq_len(nrow(mat))) # made-up markers
colnames(mat) <- paste0("BARCODE_", seq_len(ncol(mat))) # made-up barcodes

# First building the SE. Assuming that barcodes come from a range of donors.
donor <- rep(LETTERS[1:10], each=100)
se <- SummarizedExperiment(list(counts=mat))
se$Donor <- donor

rownames(se)
#>  [1] "CD1"  "CD2"  "CD3"  "CD4"  "CD5"  "CD6"  "CD7"  "CD8"  "CD9"  "CD10"
rownames(colData(se))
#>    [1] "BARCODE_1"    "BARCODE_2"    "BARCODE_3"    "BARCODE_4"   
#>    [5] "BARCODE_5"    "BARCODE_6"    "BARCODE_7"    "BARCODE_8"   
## (cut)

[LTLA]

I have always imagined the "columns" of an MAE to represent the samples; I don't expect the columns of the MAE to have a simple relationship to the columns of the internal SEs. (Which is kind of the point of the class.) So when I think of colnames(mae), I'm expecting to get the names of the samples. In fact, as currently implemented, colData(mae) returns the sample-level metadata, which already implies that the columns of the MAE are samples. So why not make colnames() consistent with this paradigm? We can argue about whether this is the most right or wrong approach, but at least it'll be self-consistent, and that's better than the current situation.

[LTLA]

Similarly, ncol(mae) should return nrow(colData(mae)) to be consistent.