[Bioc-devel] avoiding clashes of different S4 methods with the same generic

On 04/27/2016 04:24 AM, Michael Lawrence wrote:

On Tue, Apr 26, 2016 at 11:12 PM, Herv? Pag?s <hpages at fredhutch.org>
wrote:

Hi,

I would not discard defining a SummarizedExperiment subclass so quickly.
SummarizedExperiment is very generic and can contain any kind of data.
IIUC the csaw package uses SummarizedExperiment to store a particular
kind of data (ChIP-seq data) and I believe specialization is a
legitimate situation for defining a subclass, even if the subclass is
a "straight" subclass i.e. a subclass that doesn't add new slots or
doesn't touch the existing slots.

OTOH introducing a "straight" subclass only to define one specialized
method on it (the "normalize" method in this case) might not be worth
it since there is a cost for such class, even if that cost is minimal:
a cost for the user (one new container/constructor to deal with) and a
cost for the developer (e.g. multiplication of coerce methods).

If the data are more specialized, specialize the data structure,

Isn't it what I'm doing when I define a "straight" subclass? The fact that I
don't need to alter the internal representation is an
implementation detail (and it could change at some point) but what's
important is that from a user point of view my container is now
tagged/specialized. I might only have one specialized method for it
at the moment but I might have more in the future, and/or other package
developers might build on top of my specialized container and add
specialized methods in the future (and I cross my fingers that since
I own the specialized container and already implemented a "normalize"
method for it, nobody will redefine that method in their package).

H.

but
the fact that the specialization solves the normalize() ambiguity is a
mere coincidence. There are two different concerns.

Changing the signature of the normalize() generic in BiocGenerics and
introducing dual dispath is of course doable but that means the
maintainers of the packages that define methods on this generic are
ok with the dual dispatch game and are willing to make the required
modifications to their packages. It's an important change and I don't
see an easy way to make it happen smoothly (i.e. thru a
deprecated/defunct cycle).

In conjunction with what Martin said, you could define a
"ANY","missing" method that emits a deprecation warning, and then
recall the generic using NULL or something for the second argument so
that it falls through. Packages would only need to fix the formals of
their method definition.

Here is the list of packages that currently define methods for
BiocGenerics::normalize():

   affyPLM
   Cardinal
   codelink
   CopyNumber450k
   csaw
   diffHic
   EBImage
   epigenomix
   MSnbase
   oligo
   qpcrNorm
   scran

[Interestingly the scran package defines a default "normalize" method
(i.e. a normalize,ANY method)].

Whether we make the second argument lightweight or parameterized (which
is something that would need to be decided at the level of the generic)
these packages will break as soon as we change the signature of the
generic. So we'll need to wait after the release before this happens.

Personally I find the lightweight second argument not particularly
intuitive, elegant, or user-friendly. I'd rather type
normalizeSwing(se, ...) or normalize(se, SwingParam(...)) than
normalize(se, WithSwing(), ...).

Sure, WithSwing() could hold arguments as well, but I agree that the
Param suffix is more consistent. The Param naming is not great for
autocompletion. Though I guess the interface could provide hints based
on the defined methods.

Last thing: In case of a parameterized second argument, do we really
need a virtual normalizeParam class as parent of all the concrete
normalizeParam* classes? If so then I guess we would need to have it
defined in BiocGenerics but I think we should try hard to not start
defining classes in this package (that could take us too far...)

I would say no, no real need for a base class.

H.


On 04/26/2016 03:03 PM, Aaron Lun wrote:

Yes, but "monkeyBars" doesn't have quite the same pithiness for a
package name.

Anyway, the dual dispatch mechanism sounds most interesting. I assume
that means we'd have to define some sort of base "normalizeParam" class,
and then derive "csawNormParam" and "swingsNormParam" subclasses, so
that specific methods can be defined for each signature.

- Aaron

Martin Morgan wrote:

On 04/26/2016 05:28 PM, Michael Lawrence wrote:

On Tue, Apr 26, 2016 at 2:16 PM, Martin Morgan
<martin.morgan at roswellpark.org>  wrote:

On 04/26/2016 04:47 PM, Michael Lawrence wrote:

On Tue, Apr 26, 2016 at 11:00 AM, Aaron Lun<alun at wehi.edu.au>

wrote:
...

BiocGenerics. However, if some other hypothetical package

(I'll call it

"swings", for argument's sake) were to define a normalize()

method with a
...

I like the dual dispatch method quite a bit (but wonder why we get

several

swings but only one csaw? Maybe a csaw implies two participants

[though I

think I once in a while csaw-ed alone], so a singular csaw and a

pair of

swings balance out?), partly because it's very easy to extend

(write another

method) and the second argument can be either lightweight or

parameterized.

I could go along with the dual dispatch. "Swings" is short for "Set of
swings". Usually, there are several swings in a row, but only one
see-saw.

Googling for "how many swings per see-saw" took me to

    https://www.cpsc.gov//PageFiles/108601/playgrnd.pdf

where it is apparent that swings are much more dangerous than see-saws
(e.g., 51 matches for "swing" versus 4 for "see-saw"; "Swings ... were
involved in about 19 ... percent of injuries ... See-saws accounted
for about three percent"; "Homemade rope, tire, or tree swings were
also involved in a number of hanging deaths" [no mention of death by
see-saw]).

I think for the sake of our users, especially our younger users, we do
not want to consider swings, or even methods on swings, further.

Martin


This email message may contain legally privileged and/or confidential
information.  If you are not the intended recipient(s), or the
employee or agent responsible for the delivery of this message to the
intended recipient(s), you are hereby notified that any disclosure,
copying, distribution, or use of this email message is prohibited.  If
you have received this message in error, please notify the sender
immediately by e-mail and delete this email message from your
computer. Thank you.

--
Herv? Pag?s

Program in Computational Biology
Division of Public Health Sciences
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N, M1-B514
P.O. Box 19024
Seattle, WA 98109-1024

E-mail: hpages at fredhutch.org
Phone:  (206) 667-5791
Fax:    (206) 667-1319

--
Herv? Pag?s

Program in Computational Biology
Division of Public Health Sciences
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N, M1-B514
P.O. Box 19024
Seattle, WA 98109-1024

E-mail: hpages at fredhutch.org
Phone:  (206) 667-5791
Fax:    (206) 667-1319