Slashdot Mirror

Ok no "ifs", I am going to explain exactly what the article did, and why they way you characterized the principal components ("they're components that let you get a visually distinct separation between groups that you've already clustered.") is wrong.

The figure I linked to is figure S3B. The supplementary material describes the construction of this figure as follows

We computed the Identity-by-State (IBS) matrix among the 938 individuals by using PLINK (12), producing a 938-by-938 matrix. We then performed a Principal Component Analysis on this IBS matrix for all samples and for seven regions separately, and used the top components to illustrate the genetic relatedness among individuals. We show the PC1-PC2 plots for all samples in Fig. S3B and six of the seven regions in Figs. 2 and S4.

Now what you need to know about the IBS matrix, is that it is computed without reference to any prior clustering of the data. In case someone else decides to argue this point without doing their own research, here are the definitions:

[The IBS matrix is a] square, symmetric matrix of the IBS distances for all pairs of individuals
and further

The [IBS] distance between two samples is returned as the proportion of allele comparisons which are not IBS.

and finally

For each pair of subjects the [number of allele comparisons which are IBS is] the total number of alleles which are IBS. For autosomal SNPs, each locus contributes 4 comparisons, since each subject carries two copies. For SNPs on the X chromosome, the number of comparisons is also 4 for female:female comparisons, but is 2 for female:male and 1 for male:male comparisons.

source

What is important here is that at no point as we trace through the definitions needed to understand how the principal components were constructed, is any data other than the DNA of the individuals used. So from these quotes, even the most reluctant reader can see that the graph I originally linked to was constructed without any prior knowledge of the races of the 938 individuals (except of course for the color of the dots).

Now you are of course free to argue what the principal componets do and do not represent. But you are no longer in a position to argue that they were constructed in order to visualize a clustering that we had already assumed. They were not.

Don't forget R/Bioconductor! Not only is R free/free, but there are thousands of available Bioconductor packages ready for out-of-the-box use. Also consider Cytoscape and or EGAN for graph visualization of established and experimental bio-knowledge. http://www.bioconductor.org/ http://www.cytoscape.org/ http://akt.ucsf.edu/EGAN/ (full disclosure - I work on EGAN)

Like many posters have already said, the syntax of R is terribly outdated, and that's a first problem with me (I started programming with Python, go figure). But the main problem I have with R is the performance. A lot of functions and packages are dead slow or quite memory hungry (compared to a, say, C++ equivalent - for the initiated, check out the performance of rma from Bioconductor with RMAExpress, which is written in C++).

Another issue I have is not with R itself, but with its most popular add-on, the Bioconductor suite, widely used in bioinformatics. The packages' quality varies a great deal, and there's no way to file bug reports (unlike R itself, which has a bug tracker) short of emailing the authors, who, being academics, may not even have the time/will to reply to you. I'd love to see stuff like Bioconductor in a more recent programming language, but I doubt it - doing this kind of stuff doesn't give you any funding.

Regarding interfacing with R note that R itself can do minimal symbolic
differentiation out-of-the-box as shown by this sample R session:

> deriv(expression(x^2))
2 * x

and has a partially developed interface with yacas via the addon package Ryacas.
After installing Ryacas and yacas this R code works:

> library(Ryacas)
> x = Sym("x")
> deriv(x^2)
expression(2 * x)

This Ryacas interface includes a partial recursive decent R parser that translates
R code to yacas code and an XML-based OpenMath connection in the other direction.
Communication is via sockets. See

http://ryacas.googlecode.com/

Unlike Sage, symbolic computation is not really the focus of R but R does have
1000+ free addon packages including interfaces to numerous other free and
commercial systems. The addon packages are listed in these repositories which
focus on general items, interfacing and biology, respectively:

http://cran.r-project.org/
http://www.omegahat.org/
http://www.bioconductor.org/

Also there is a graphics gallery with sample R graphics:

http://addictedtor.free.fr/graphiques/

The R home page can be found by entering the single letter R into google.

Two other wonderful, and interrelated, open source packages are R, the open source statistical software project, and Biocondutor , basically a set of R functions.

pubmed

golden path

bioconducter

public library of science

gnumeric

cluster analysis

etc. etc. etc.

What's the BFD ??? A lot of scientists are on the open source bandwagon and have been for years. Walmart's coming to town and the Ivory Towers are falling.

In my field(s) of study (bioinformatics, genetics, evolution, organismal and molecular biology), I can say that the open source movement has NOT been adopted by the majority. However, I have recently come across the R project (particularly Bioconductor.org ) which is HIGHLY respected by many MS-oriented biologists from a wide scope of fields. Many computational tasks which are limited to those who can fork out $6000 per year for a license, can be performed on R (for free). Although the learning curve is somewhat steep, I've found the documentation to be exceptional and have been able to execute my needs within a day or two. It has also come to my attention that many graduate students and professors across the country are actively learning R. Granted, many of the users of R are probably "sophisticated"- familar with SAS, C, or UNIX- but it seems to me that open-source is congruent with the peer-review process that so many academics are familiar with and value. I think, if we can introduce it to enough people, and explain to them WHY the open-source method (as opposed to proprietary software) is invaluable, it will become mainstream. R seems to be leading the way in academia at least...