[GenABEL-dev] Interesting idea by Gulnara Svischeva

[Astle, William J]

Hi all,

This might be of interest,

best wishes

Will

>
>
> Dear William,
>
> My name is Christoph Lippert. I am a PhD student from the Max Planck Institute for Developmental Biology in Tuebingen, currently working on Mixed Models for GWAS.
>
> Recently, we had a paper in Nature Methods, where among other results we presented a way to optimize the ratio of variance parameters for genome-wide SNPs in a mixed model with only a single cubic operation.
>Last week, I heard that you had a similar result in your thesis. I would be highly interested in your derivation of that result.
> Could you perhaps send me a copy of your thesis or a writeup of that result? This would be highly appreciated.
>
> Thank you,
> Christoph
>
> P.S.: If you are interested in our paper, here is the link:
> http://www.nature.com/nmeth/journal/vaop/ncurrent/abs/nmeth.1681.html


On 8 Oct 2011, at 04:37, L.C. Karssen wrote:

Dear list,

I was going through the changelog of GenABEL and saw that Gulya's
poly_eigen is the default method for polygenic() since version 1.6-8. I
did a quick comparison on the computation speed of the three methods
(polygenic_hglm, polygenic with polylik and with poly_eigen) on our
population and even though I knew what to expect from Yurii's results, I
am still impressed at the speedup!

I was wondering, is there a publication on this method/implementation?
Has there been any comparison of poly_eigen vs. the old method (in terms
of variance, lambda etc.? In February/March this year there was a
discussion on this list where Yurii compared the polylik method with
polygenic_hglm, but I haven't seen anything similar on the poly_eigen
method.


Enjoy the weekend,

Lennart.


On vr, 2011-05-20 at 21:16 +0200, Yurii Aulchenko wrote:
In attached PDF I show the timing for running 'polygenic' (red line;
ML operating on Gulya's idea), polygenic_hglm (black; Xia/Lars), and
REML.rotate (green; Lars's implementation/extension of Gulya's idea)
as a function of sample size (X axes).

While polygenic and hglm show exponential trend, REMLrotate line is
not only the lowest, but also looks almost linear to me (also looking
at the code, it should be roughly linear on time with no of subjects).
So, with REMLrotate we can get arbitrary big speed-up (x100, x1000,
x10000, ... you name it!) cf other methods, just by adding more people
to the data :) Not a bad result!

Lars and Gulya, you are the champions!

best wishes,
Yurii

2011/5/19 Yurii Aulchenko <yurii.aulchenko at gmail.com<mailto:yurii.aulchenko at gmail.com>>:
Correction: as Lars spotted out, I did placed my 'tags' wrongly
concerning time performance results, so actually REML.rotate is the
fastest.

Here are updated results:

With conv=1e-8 mean time for estimation on 300 IDs was:

17.81 sec for polygenic_hglm
2.27 sec for 'polygenic'-ML
1.33 sec for REML.rotate

best wishes, and sorry for inconvenience,
Yurii

_______________________________________________
genabel-devel mailing list
genabel-devel at lists.r-forge.r-project.org<mailto:genabel-devel at lists.r-forge.r-project.org>
https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/genabel-devel

--
-----------------------------------------------
L.C. Karssen
Erasmus MC
Department of Epidemiology
Room Ee-2224

Postbus 2040
3000 CA Rotterdam
The Netherlands

phone: +31-10-7044217
fax: +31-10-7044657
e-mail: l.karssen at erasmusmc.nl<mailto:l.karssen at erasmusmc.nl>
GPG key ID: 0E1D39E3
-----------------------------------------------
_______________________________________________
genabel-devel mailing list
genabel-devel at lists.r-forge.r-project.org<mailto:genabel-devel at lists.r-forge.r-project.org>
https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/genabel-devel

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/genabel-devel/attachments/20111011/b20a9647/attachment.htm>