On 31 Oct 2014, at 23:43, Steve Lianoglou <lianoglou.steve at gene.com> wrote:
Hi,
On Fri, Oct 31, 2014 at 2:35 PM, Thomas Lin Pedersen
<thomasp85 at gmail.com <mailto:thomasp85 at gmail.com>> wrote:
With regards to abstraction - I would personally much rather read and write code that contained plotScores() and plotScree() etc. where the intend of the code is clearly communicated, instead of relying on a plot() function whose result is only known from experience. Trying to squeeze every kind of visual output into the same plot generic seems artificial and constrained to me. I totally agree on the plotPCA critique on the other hand...
If we've bought a ticket to ride on Kevin's and Michael's (and whoever
else) train of thought, wouldn't plot(pca(x), type='scree') or
plot(pca(x), type='scores') be the preferred way to go ... for some
definition of "preferable"?
-steve
On 31 Oct 2014, at 22:09, Michael Lawrence <lawrence.michael at gene.com> wrote:
I strongly agree with Kevin's position. plotPCA() represents two separate concerns in its very name: the computation and the rendering. Those need to be separated, at least behind the scenes. The syntax of plot(pca(x)) is preferable to plotPCA, because the structure of the operation is represented by in the expression itself, not just in a non-computable function name.
With regard to how a plot,PCA should behave: there is always a tension between high-level and low-level APIs. In the end, we need multiple levels of abstraction. While high-level APIs sacrifice flexibility, we need them because they communicate the high-level *intent* of the user in the code itself (self-documenting code), and they enable reusability, which not only reduces redudant effort but also ensures consistency. Once our brains no longer need to parse low-level code, we can focus our mental power on correctness and efficiency. To design a high-level API, one needs to carefully analyze user requirements, i.e., the use cases. To choose the default behavior, one needs to rate the use cases by their prevalance, and by how closely they match the intuition-based expectations of the user.
The fact that at least 9 packages are performing such a similar task seems to indicate that a common abstraction is warranted, but I am not sure if BiocGenerics is the appropriate place.
Michael
On Tue, Oct 21, 2014 at 12:54 AM, Thomas Dybdal Pedersen <thomasp85 at gmail.com <mailto:thomasp85 at gmail.com> <mailto:thomasp85 at gmail.com <mailto:thomasp85 at gmail.com>>> wrote:
While I tend to agree with you that PCA is too big an operation to be hidden within a plotting function (MDS is an edge-case I would say), I can't see how we can ever reach a point where there is only one generic plot function. In the case of PCA there is a number of different plot-types that can all lay claim to the plot function of a PCA class, for instance scoreplot, scatterplot matrix of all scores, biplot, screeplot, accumulated R^2 barplot, leverage vs. distance-to-model... (you get the idea). So while having some very well-thought out classes for very common result types such as PCA, this class would still need a lot of different plot methods such as plotScores, plotScree etc (or plot(..., type='score'), but I don't find that very appealing). Expanding beyond PCA only muddles the water even more - there are very few interesting data structures that only have one visual representation to-rule-them-all...
just my 2c
best
Thomas
Date: Mon, 20 Oct 2014 18:50:48 -0400
From: Kevin Coombes <kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com>>>
Well. I have two responses to that.
First, I think it would be a lot better/easier for users if (most)
developers could make use of the same plot function for "basic" classes
like PCA.
Second, if you think the basic PCA plotting routine needs enhancements,
you still have two options. On the one hand, you could (as you said)
try to convince the maintainer of PCA to add what you want. If it's
generally valuable, then he'd probably do it --- and other classes that
use it would benefit. On the other hand, if it really is a special
enhancement that only makes sense for your class, then you can derive a
class from the basic PCA class
setClass("mySpecialPCA", contains=c("PCA"), *other stuff here*)
and implement your own version of the "plot" generic for this class.
And you could tweak the "as.PCA" function so it returns an object of the
mySpecialPCA class. And the user could still just "plot" the result
without hacving to care what's happening behind the scenes.
On 10/20/2014 5:59 PM, Michael Love wrote:
Ah, I see now. Personally, I don't think Bioconductor developers
should have to agree on single plotting functions for basic classes
like 'PCA' (because this logic applies equally to the situation of all
Bioconductor developers agreeing on single MA-plot, a single
variance-mean plot, etc). I think letting developers define their
plotPCA makes contributions easier (I don't have to ask the owner of
plot.PCA to incorporate something), even though it means we have a
growing list of generics.
Still you have a good point about splitting computation and plotting.
In practice, we subset the rows so PCA is not laborious.
On Mon, Oct 20, 2014 at 5:38 PM, Kevin Coombes
<kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com>> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com>>>> wrote:
Hi,
I don't see how it needs more functions (as long as you can get
developers to agree). Suppose that someone can define a reusable
PCA class. This will contain a single "plot" generic function,
defined once and reused by other classes. The existing "plotPCA"
interface can also be implemented just once, in this class, as
plotPCA <- function(object, ...) plot(as.PCA(object), ...)
This can be exposed to users of your class through namespaces.
Then the only thing a developer needs to implement in his own
class is the single "as.PCA" function. And he/she would have
already been rquired to implement this as part of the old
"plotPCA" function. So it can be extracted from that, and the
developer doesn't have to reimplement the visualization code from
the PCA class.
Best,
Kevin
On 10/20/2014 5:15 PM, davide risso wrote:
Hi Kevin,
I see your points and I agree (especially for the specific case
of plotPCA that involves some non trivial computations).
On the other hand, having a wrapper function that starting from
the "raw" data gives you a pretty picture (with virtually zero
effort by the user) using a sensible choice of parameters that
are more or less OK for RNA-seq data is useful for practitioners
that just want to look for patterns in the data.
I guess it would be the same to have a PCA method for each of the
objects and then using the plot method on those new objects, but
that would just create a lot more objects and functions than the
current approach (like Mike was saying).
Your "as.pca" or "performPCA" approach would be definitely better
if all the different methods would create objects of the *same*
PCA class, but since we are talking about different packages, I
don't know how easy it would be to coordinate. But perhaps this
is the way we should go.
Best,
davide
On Mon, Oct 20, 2014 at 1:26 PM, Kevin Coombes
<kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com>> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com>>>> wrote:
Hi,
It depends.
The "traditional" R approach to these matters is that you (a)
first perform some sort of an analysis and save the results
as an object and then (b) show or plot what you got. It is
part (b) that tends to be really generic, and (in my opinion)
should have really generic names -- like "show" or "plot" or
"hist" or "image".
With PCA in particular, you usually have to perform a bunch
of computations in order to get the principal components from
some part of the data. As I understand it now, these
computations are performed along the way as part of the
various "plotPCA" functions. The "R way" to do this would be
something like
pca <- performPCA(mySpecialObject) # or
as.PCA(mySpecialObject)
plot(pca) # to get the scatter plot
This apporach has the user-friendly advantage that you can
tweak the plot (in terms of colors, symbols, ranges, titles,
etc) without having to recompute the principal components
every time. (I often find myself re-plotting the same PCA
several times, with different colors or symbols for different
factrors associated with the samples.) In addition, you could
then also do something like
screeplot(pca)
to get a plot of the percentages of variance explained.
My own feeling is that if the object doesn't know what to do
when you tell it to "plot" itself, then you haven't got the
right abstraction.
You may still end up needing generics for each kind of
computation you want to perform (PCA, RLE, MA, etc), which is
why I suggested an "as.PCA" function. After all, "as" is
already pretty generic. In the long run, l this would herlp
BioConductor developers, since they wouldn't all have to
reimplement the visualization code; they would just have to
figure out how to convert their own object into a PCA or RLE
or MA object.
And I know that this "plotWhatever" approach is used
elsewhere in BioConductor, and it has always bothered me. It
just seemed that a post suggesting a new generic function
provided a reasonable opportunity to point out that there
might be a better way.
Best,
Kevin
PS: My own "ClassDicsovery" package, which is available from
RForge via
**|install.packages("ClassDiscovery",
repos="http://R-Forge.R-project.org <http://r-forge.r-project.org/> <http://r-forge.r-project.org/ <http://r-forge.r-project.org/>>"
<http://R-Forge.R-project.org <http://r-forge.r-project.org/> <http://r-forge.r-project.org/ <http://r-forge.r-project.org/>>>)|**
includes a "SamplePCA" class that does something roughly
similar to this for microarrays.
PPS (off-topic): The worst offender in base R -- because it
doesn't use this "typical" approch -- is the "heatmap"
function. Having tried to teach this function in several
different classes, I have come to the conclusion that it is
basically unusable by mortals. And I think the problem is
that it tries to combine too many steps -- clustering rows,
clustering columns, scaling, visualization -- all in a single
fiunction
On 10/20/2014 3:47 PM, davide risso wrote:
Hi Kevin,
I don't agree. In the case of EDASeq (as I suppose it is the
case for DESeq/DESeq2) plotting the principal components of
the count matrix is only one of possible exploratory plots
(RLE plots, MA plots, etc.).
So, in my opinion, it makes more sense from an object
oriented point of view to have multiple plotting methods for
a single "RNA-seq experiment" object.
In addition, this is the same strategy adopted elsewhere in
Bioconductor, e.g., for the plotMA method.
Just my two cents.
Best,
davide
On Mon, Oct 20, 2014 at 11:30 AM, Kevin Coombes
<kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com>>
<mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com>>>> wrote:
I understand that breaking code is a problem, and that
is admittedly the main reason not to immediately adopt
my suggestion.
But as a purely logical exercise, creating a "PCA"
object X or something similar and using either
plot(X)
or
plot(as.PCA(mySpecialObject))
is a much more sensible use of object-oriented
programming/design. This requires no new generics (to
write or to learn).
And you could use it to transition away from the current
system by convincing the various package maintainers to
re-implement plotPCA as follows:
plotPCA <- function(object, ...) {
plot(as.PCA(object), ...)
}
This would be relatively easy to eventually deprecate
and teach users to switch to the alternative.
On 10/20/2014 1:07 PM, Michael Love wrote:
hi Kevin,
that would imply there is only one way to plot an
object of a given class. Additionally, it would break a
lot of code.?
best,
Mike
On Mon, Oct 20, 2014 at 12:50 PM, Kevin Coombes
<kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com>>
<mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com> <mailto:kevin.r.coombes at gmail.com <mailto:kevin.r.coombes at gmail.com>>>> wrote:
But shouldn't they all really just be named "plot"
for the appropriate objects? In which case, there
would already be a perfectly good generic....
On Oct 20, 2014 10:27 AM, "Michael Love"
<michaelisaiahlove at gmail.com <mailto:michaelisaiahlove at gmail.com> <mailto:michaelisaiahlove at gmail.com <mailto:michaelisaiahlove at gmail.com>>
<mailto:michaelisaiahlove at gmail.com <mailto:michaelisaiahlove at gmail.com> <mailto:michaelisaiahlove at gmail.com <mailto:michaelisaiahlove at gmail.com>>>> wrote:
I noticed that 'plotPCA' functions are defined
in EDASeq, DESeq2, DESeq,
affycoretools, Rcade, facopy, CopyNumber450k,
netresponse, MAIT (maybe
more).
Sounds like a case for BiocGenerics.
best,
Mike
[[alternative HTML version deleted]]