Generics and methods already in DelayedArray

[PeteHaitch]

There are several generics and methods already implemented in DelayedArray that 'belong' in DelayedMatrixStats; see https://github.com/PeteHaitch/DelayedMatrixStats/blob/master/README.md and note that I'm using 'belong' loosely here.

S4 generics

The simple ones

It makes sense to me to move the following generics (and methods) from DelayedArray to DelayedMatrixStats:

• colMaxs()
• colMins()
• colRanges()
• rowMaxs()
• rowMins()

The obvious complication of moving these is that any users of these generics and methods would have to update their code accordingly (but I doubt there are many). It would, however, tidy things up somewhat and slightly simplify ongoing development.

The complicated one

The final generic is:

• rowRanges()

This is more complicated because it is a generic that is serving 2 very distinct use cases:

1. Get the row-ranges of a SummarizedExperiment-based object
2. Compute the row-wise ranges of a matrix-like object

This is something we're stuck with. The S4 generic should remain where it is to avoid having SummarizedExperiment depend upon DelayedMatrixStats . It also must have a slightly different signature to the rest of the methods due to this dual role; compare rowRanges() to colRanges()

suppressPackageStartupMessages(library(DelayedMatrixStats))

rowRanges
#> standardGeneric for "rowRanges" defined from package "DelayedArray"
#> 
#> function (x, ...) 
#> standardGeneric("rowRanges")
#> <environment: 0x7fe2e9354958>
#> Methods may be defined for arguments: x
#> Use  showMethods("rowRanges")  for currently available ones.
colRanges
#> standardGeneric for "colRanges" defined from package "DelayedArray"
#> 
#> function (x, rows = NULL, cols = NULL, na.rm = FALSE, dim. = dim(x), 
#>     ...) 
#> standardGeneric("colRanges")
#> <environment: 0x7fe2e7e3ee00>
#> Methods may be defined for arguments: x
#> Use  showMethods("colRanges")  for currently available ones.

S4 methods

Each of these S4 generics from DelayedArray have a ANY-method implemented. This is in contrast to what I have elected to do for the generics defined in DelayedMatrixStats: I specifically implemented a matrix-method but no ANY-method. E.g., compare the following

suppressPackageStartupMessages(library(DelayedMatrixStats))

# Has an ANY-method
showMethods("colRanges")
#> Function: colRanges (package DelayedArray)
#> x="ANY"
#> x="DelayedMatrix"

# Has a matrix-method instead of an ANY-method
showMethods("colMedians")
#> Function: colMedians (package DelayedMatrixStats)
#> x="DelayedMatrix"
#> x="matrix"

The ANY-method of these generics are the correspondingly named matrixStats functions. Now, the matrixStats functions of the same names do check their inputs, so will error if given, say, a data frame. However, the real reason I took great pains to avoid introducing an ANY-method for the S4 generics defined in DelayedMatrixStats is because of the way I have implemented S4 dispatch based on the seed of the DelayedMatrix object - when an appropriate method does not exist for the seed, then the default 'block-processing' method is used. If an ANY-method exists, however, then there always exists a method for the seed but it is (erroneously) the corresponding function from matrixStats, which obviously only works with matrix objects.

Summary

I need to discuss these issues with members of the Bioconductor core team.