Skip to content

[Bioc-devel] assay dimnames in SingleCellExperiment / SummarizedExperiment

5 messages · Kevin RUE, Hervé Pagès, Aaron Lun

Original
Kevin RUE
# 
Dear all,

I cc-ed to this email individual package maintainer to directly 'notify'
them of this thread and have their respective opinions, but I thought the
common use of SummarizedExperiment was worth involving the community as
well.

Background: I was updating one of my workflow from SCESet to the
SingleCellExperiment class recently introduced on the development branch.

1)
One thing leading to another, I ended up noticing that there is no validity
check on dimnames of the various assays in SummarizedExperiment. In other
words, the different assays can have different `dimnames` (or some assays
can have NULL dimnames). Using the example code from SummarizedExperiment:

nrows <- 200; ncols <- 6
counts3 <- counts2 <- counts <-
  matrix(runif(nrows * ncols, 1, 1e4), nrows)

rnames <- paste0("F_", sprintf("%03.f", seq_len(nrows)))
cnames <- LETTERS[1:6]

dimnames(counts) <- list(rnames, cnames)
dimnames(counts2) <- list(Tags = rnames, Samples = cnames)
dimnames(counts3) <- list(Features = rnames, Cells = cnames)

colData <- DataFrame(row.names=cnames)

rse <- SummarizedExperiment(assays=SimpleList(c1=counts, c2=counts2,
c3=counts3), colData=colData)

assayNames(rse)
names(dimnames(assay(rse, "c1"))) # NULL
names(dimnames(assay(rse, "c2"))) # [1] "Tags"    "Samples"
names(dimnames(assay(rse, "c3"))) # [1] "Features" "Cells"

Although not critical, it'd probably be best practice to have a validity
check on identical dimnames across all assay, so that one does not have to
worry later about `melt` calls returning different column names whether
each assay has proper dimnames or not.


2)
The initial glitch that prompted this email related to the `reshape2::melt`
method that extracts dimnames, if available, in the
`scater::plotHighestExprs` function. Anyway, Davis has already prepared a
fix to deal with the scenario whereby the assay does have dimnames (e.g.
counts in the edgeR::DGEList class that I generally use to import counts).
Somehow that wasn't an issue with the SCESet that I was using previously
(probably a side-effect of ExpressionSet).

The point is, the glitch prompted me to think whether a potential
standardisation of names(dimnames) could be beneficial, perhaps more
specifically in the new `SingleCellExperiment` class (as
SummarizedExperiment has a much more general purpose). Considering the
fairly specific purpose of the former, I was wondering whether it would be
worth:

   - enforcing names(dimnames(x)) to "Features" and "Cells", (bearing in
   mind that features could still be genes, transcripts, ...)
   - or maybe dropping dimnames altogether, storing them only once
   elsewhere (although a slot for that seems overkill)

There may be other possibilities that I haven't thought of yet, but I
thought I'd get the ball rolling.
Having well-defined dimnames sounds good practice, with the added benefit
of generating aesthetically pleasing column names in melted data-frame as a
by-product.
However, I can't tell whether the handling of dimnames is something that
needs to be handle by individual downstream package developers, or whether
standards should be set in parent classes.


Thanks for your time!

Best,
Kevin
1 day later
Aaron Lun
# 
I'll leave the first point to the SummarizedExperiment maintainers, though I  note that your code seems to be about the names of the dimnames rather than the dimnames themselves. (I'm under the impression that consistency in the actual dimnames is enforced somehow by the SE constructor.)


As for the second point; I suppose we could set the second name for the dimnames as "Cells" in SingleCellExperiment, though the choice for the first name is more ambiguous. This request has come up before, and I've never been entirely convinced by its necessity. It seems mostly aesthetic to me, and honestly, if a user doesn't already know that rows are genes and columns are cells, I can't see them flailing away at the keyboard until they call dim() to tell them what the dimensions correspond to.


But I guess other people like aesthetics, so if you want, you can put in a PR to override dim() and dimnames() for SingleCellExperiment to put some names on the returned vectors or lists. If I had to choose, I would go with "Features" and "Cells" for the rows and columns, respectively. (We already use a RSE so we're already implicitly assuming genomic features.)


-Aaron
Kevin RUE
# 
Hi Aaron,

Yes - sorry, I meant the names of dimnames. Dimnames are indeed checked,
but my code was meant to demonstrate that names of dimnames aren't.
Obviously, it's not the end of the world, but just something I noticed
while I was investigating the glitch.

My second point is not that much about calling dim or dimnames, but rather
about the side-effects of having names(dimnames(x)) not NULL, such as the
case of `reshape2::melt`.
I think it'd be one worry less for downstream methods to 'know' the
colnames of a melted assay(x, 1) instead of having "Var1, Var2, value" if
names(dimnames) is NULL, and "something else" if not NULL.

Beyond aesthetics, it's really just semantics, but I do think small stuff
like that, if handled at a higher class level, can encourage downstream
developers to work off a more consistent mental and computational model (my
take from Michael Lawrence's BOF at Bioc2017). In other words, it has a
small cost to implement in the parent class, instead of if-else statements
in each child class.

It could be something as simple as :

   - c("Feature", "Sample") at the `SummarizedExperiment` level
   - overriden by c("Feature", "Cell") in `SingleCellExperiment`
   - overriden by developer's choice in other dependent packages.


All the best,
Kevin
On Sat, Sep 16, 2017 at 6:43 AM, Aaron Lun <alun at wehi.edu.au> wrote:

            

      
      
      
    

        

  
  


  


      

    

    
6 days later

    

      
      

        


  

        

      Hervé Pagès
    

    

      
      #
    

  


  

      
Hi guys,

        
On 09/16/2017 03:49 AM, Kevin RUE wrote:

            

      
      
      
    

        
Sounds like a good idea to me to check for consistent names of
dimnames across assays. I'll add this to the validity method of 
SummarizedExperiment objects.

            

      
      
      
    

        
I'm not too keen on enforcing this at the SummarizedExperiment level.
The rows of a SummarizedExperiment object sometimes correspond to
bins or to a running window. You could even imagine use cases where
they correspond to reads or groups of reads or protein IDs. As general
as "Feature" might sound, it would feel a little bit like a misnomer
for these use cases. This could slightly hurt the re-usability appeal
of SummarizedExperiment objects.

I could see the same argument being made about enforcing
names(dimnames(x))[1] to "Feature" for a SingleCellExperiment
object. However enforcing names(dimnames(x))[2] to "Cell" is
probably fine and "Cell" seems like the natural choice given
that this is hardcoded in the name of the class.

Note that technically you cannot have SummarizedExperiment
enforce c("Feature", "Sample") and SingleCellExperiment enforce
something else. That's because S4 doesn't let you override validity
criteria defined by an ancestor class. And that in turn is because
in S4 validation is *incremental*. This means that the validity
method for a subclass only needs to worry about validating what's
not already covered by the validity methods of all the ancestors
class. When one calls validObject(x), first the validity methods
for all 'x' ancestor classes are called (from the most distant
ancestor to the direct parent), and the validity method for the
class of 'x' is finally called. This means that you cannot write
a validity method for 'x' that contradicts what the validity
methods for the ancestor classes expect. In other words, if B
extends A, an object of class B must be a valid A object (remember
that is(x, "A") is TRUE) before it can be considered to be a
valid B object. In the (almost) real world this is just saying
that before a cat can be considered to be a valid red cat it must
first be considered to be a valid cat. (Don't ask me what a valid
cat is.)

Cheers,
H.

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Aaron Lun
    

    

      
      #
    

  


  

      
Well, I guess it probably wouldn't be too bad to have "Feature" for the SingleCellExperiment "names(dimnames(...))[1]". The SCE inherits from an RSE anyway so we do imply that the rows represent some kind of genomic feature. Maybe this is too presumptive, but it doesn't seem to clash with any of  the current  applications of SCE. Of course, we could always just leave the first name empty, but I'm not sure whether this would cause Kevin's use cases to do funny things.


On another note, what is the easiest way to enforce dimnames names at the SCE level? Do I need to implement new methods for assay<- or assays<-, to overwrite the names(dimnames()) of incoming matrices? That seems like a pain. If the base SE is going to check/enforce consistent names of dimnames, perhaps this could be made into a method that SCE can specialize to coerce the names to something else.


-Aaron