[Bioc-devel] Changes to the SummarizedExperiment Class

sessionInfo()

On Mar 9, 2015 12:36 PM, "Martin Morgan" <mtmorgan at fredhutch.org> wrote:

On 03/09/2015 08:07 AM, Michael Love wrote:

Some guidance on how to avoid duplication of the matrix for developers
would be greatly appreciated.

It's unsatisfactory, but using withDimnames=FALSE avoids duplication on extraction of assays (but obviously you don't have dimnames on the matrix). Row or column subsetting necessarily causes the subsetted assay data to be duplicated. There should not be any duplication when rowRanges() or colData() are changed without changing their dimension / ordering.

Thanks Martin for checking into the regression.

Sorry, I should have been more specific earlier, I meant more guidance/documentation in the man page for SE. I scanned the 'Extension' section but didn't find a note on withDimnames for extracting the matrix or this example of renaming the assays (it seems like this could easily be relevant for other package authors).

A prominent note there might help devs write more memory efficient packages.

The argument section mentions speed but I'd explicitly mention memory given that we're often storing big matrices:

"Setting withDimnames=FALSE  increases the speed with which assays are extracted."

(its entirely possible the info is there but i missed it)

Best,

Mike

Another example of a trouble point, is that if I am given an SE with
an unnamed assay and I need to give the assay a name, this also can
expand the memory used. I had found a solution (which works with
GenomicRanges 1.18 / current release) with:

names(assays(se, withDimnames=FALSE))[1] <- "foo"

But now I'm looking in devel and this appears to no longer work. The
memory used expands, equivalent to:

names(assays(se))[1] <- "foo"

Here's some code to try this:

m <- matrix(1:1e7,ncol=10,dimnames=list(1:1e6,1:10))
se <- SummarizedExperiment(m)
names(assays(se, withDimnames=FALSE))[1] <- "foo"
names(assays(se))[1] <- "foo"

while running gc() in between steps.

I think this is a regression of some sort, and I'll look into it. Thanks for the heads-up.

Martin

On Mon, Mar 9, 2015 at 10:36 AM, Kasper Daniel Hansen
<kasperdanielhansen at gmail.com> wrote:

On Mon, Mar 9, 2015 at 10:30 AM, Vincent Carey <stvjc at channing.harvard.edu>
wrote:

I am glad you are keeping this discussion alive Kasper.

On Mon, Mar 9, 2015 at 10:06 AM, Kasper Daniel Hansen <
kasperdanielhansen at gmail.com> wrote:

It sounds like the proposed changes are already made.  However (like
others) I am still a bit mystified why this was necessary.  The old
version
did allow for a GRanges inside the DataFrame of the rowData, as far as I
recall.  So I assume this is for efficiency.  But why?  What kind of
data/use cases is this for?

I am happy to hear that SummarizedExperiment is going to be spun out into
its own package.  When that happens, I have some comments, which I'll
include here in anticipation
   1) I now very strongly believe it was a design mistake to not have
colnames on the assays.  The advantage of this choice is that sampleNames
are only stored one place.  The extreme disadvantage is the high
ineffeciency when you want colnames on an extracted assay.

after example(SummarizedExperiment)

colnames(assays(se1)[[1]])

[1] "A" "B" "C" "D" "E" "F"

so this seems to be optional.  But attempts to set rownames will fail
silently

rownames(assays(se1)[[1]]) = as.character(1:200)

rownames(assays(se1)[[1]])

NULL
seems we could issue a warning there

Vince, you need to be careful here.

The assays are stored without colnames (unless something has recently
changed).  The default is to - upon extraction - set the colnames of the
matrix.  This however requires a copy of the entire matrix.  So
essentially, upon extraction, each assay is needlessly duplicated to add
the colnames.  This is what I mean by inefficient. I would prefer to store
the assays with colnames.  This means that changing sampleNames of the
object will be inefficient (as it is for eSets) since it would require a
complete copy of everything.  But I would rather - much rather - copy when
setting sampleNames than copy when extracting an assay.

Best,
Kasper

         [[alternative HTML version deleted]]

--
Computational Biology / Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N.
PO Box 19024 Seattle, WA 98109

Location: Arnold Building M1 B861
Phone: (206) 667-2793