[Bioc-devel] eSet questions

[Vincent Carey 525-2265]

> Hi,
>
> First, according to the manual pages for the "annotatedDataSet" class

small "s" for Dataset ... threw me for a minute

> (in BioConductor 1.9) is a "virtual superset for 'exprSet' , 'eSet',
> etc".  While this seems to be the case for the soon-to-be-deprecated
> exprSet, it seems not to be the case for an eSet.  Is that
> interpretation correct?

I am not sure annotatedDataset is going anywhere.  It probably should
be removed.

>
>
> Now to the real question.  While trying to think about how to handle a
> couple of data sets, I've started to become convinced that the current
> design of an eSet could be improved.  As it stands now, the design makes
> some assumptions about how the data should be stored and interpreted
> that I think are unnecessary and make it harder to generalize to other
> data types.
>
> I have three use cases in mind:
> [1] A vanilla two-color mRNA microarray expression data set, but one
> that is not quantified with a software package currently recognized by
> either limma or marray.
> [2] A MINiML format file containing glass array data from GEO
> [3] Reverse phase protein array (RPPA) data
>
> In the first two cases, I'd like to be able to get all the raw data
> files into R as quickly as possible, and work there to figure out which
> columns represent red and green foreground and background, after which I
> can convert from the input format to something where I can use limma or
> marray.
>
> In the RPPA case, the notions of "featureData" and "phenoData" are
> reversed.  Lysates of individual patient samples are spotted on the
> array, which is then probed with a mono-specific antibody targeting one
> protein. (See, for example, Tibes et al., Mol Cancer Ther 2006; 5:2512-21.)
>
> One way to handle all three cases would be in something I'm tentatively
> calling an "ArrayCube", which should correspond fairly closely to a set
> of files on a hard drive.  Each file holds a two-dimensional table,
> where the rows correspond to spots on an array and the columns
> correspond to various things measured by a quantification software
> package.  An ArrayCube can be thought of conceptually as a list of these
> two-dimensional objects, where this third (list) dimension corresponds
> to whatever label-producing stuff was hybridized or incubated on the array.
>
> Given this description, one might attempt a design something like
>
> setClass("ArrayCube", representation=list(
> 	rawData = "AssayData",
> 	experimentData = "MIAME",
> 	featureData = "AnnotatedDataFrame",
> 	hybridizationData = "AnnotatedDataFrame",
> 	measurementData = "AnnotatedDataFrame"
> ))
>
> This obviously looks a lot like an eSet.  The differences are

It seems to me that you don't want to adopt the "AssayData"-"phenoData"
relationship documented in the eSet man page.  So the above is
not like an eSet, and the conflict is mostly with the AssayData
structure.

> [1] I am thinking about the rawData entry as a list of data frames (or
> data matrices), with each one corresponding to a unique file on the hard
> disk.  These would be easy to read into R in the use cases above, but
> violate one of the validity constraints on the assayData object in the
> current eSet. (Specifically, the contraint that the columns in any
> matrix in the assayData object must correspond to rows of the phenoData
> object.)

This constraint is quite important for all the applications of eSet
in use, so abandoning it suggests designing another class.

> [2] The featureData slot would describe the rows in each of those data
> matrices.  In order to accommodate the RPPA data, however, featureData
> might refer to patient samples instead of the genes that it would refer
> to in the eSet design.

I have not had time to think at length about the RPPA data structure.
It seems possible to use the eSet design to represent it, but there
is substantial reorganization of the data relative to its physical
origins.  There are costs and benefits to shoehorning the data into
an ExpressionSet-like structure and I don't know how to weigh
them at the moment.  The real question seems to me to be whether it
is valuable to request X[G, S] for any of these data structures, where
X is the basic container, G is a predicate identifying a gene selection
and S is a predicate identifying a sample selection.  If you want that
AND you want to inherit the infrastructure available for ExpressionSets
to get that, then it makes sense for you to try to extend what we have in Biobase
to cover what you are dealing with.  It seems to me that you might
want to combine AssayData and AnnotatedDataFrame components in a
structure that does not extend eSet to get what you want.

> [3] Similarly, hybridizationData would replace the phenoData slot, and
> it also could refer to samples or to genes/proteins depending on the
> data type.  Also, the phenoData object has to describe the "list"
> dimension of the rawData instead of the "column" dimension".
> [4] The measurementData slot would describe the measurement columns from
> the software.  For already known packages, it would then be easy to
> convert an ArrayCube into, for example, an RGList by slicing along the
> desired measurement columns.  For novel quantification packages, one
> could make an interface that lets the user specify which measurements
> have which interpretation, and then make an RGList after they have had a
> chance to load the data easily and start exploring it.
>
> In any event, before I head further down this road, I'd like to get some
> feedback on whether it would be [a] feasible or [b] desirable either to
> create such a thing or to change the design of an eSet into such a thing.

My reaction, based on very brief contemplation, is that you'll be designing
a structure that does not extend eSet but shares some components and some
functionalities.  If the ability to represent, e.g.,
RPPA and Expression arrays in a single container type becomes important
we'll consider how the eSet constraints need to evolve.  Thus far they
seem to be effective for the most common types of high-throughput data
encountered.

The folks who actually designed the key Biobase containers may well have
different views of this situation.  This is just my personal reaction.