R help - Jun 2003 - speeding up 1000s of coxph regression?

I have a gene expression matrix n (genes) X p (cases), where n = 8000 and p
= 100. I want to fit each gene as univariate in a coxph model, i.e., fitting
8000 models. I do something like this:

res <- apply(data, 1, coxph.func)

which takes about 4 min, not bad. But I need to do large numbers of
permutations of the data (permuting the columns), for example, 2000, which
would take 5 days. I would like to know if there is way to speed this up?

Any help appreciated.

Xiao-Jun

On Tue, 10 Jun 2003, Xiao-Jun Ma wrote:
> I have a gene expression matrix n (genes) X p (cases), where n = 8000 and p
> = 100. I want to fit each gene as univariate in a coxph model, i.e.,
fitting
> 8000 models. I do something like this:
>
> res <- apply(data, 1, coxph.func)
>
> which takes about 4 min, not bad. But I need to do large numbers of
> permutations of the data (permuting the columns), for example, 2000, which
> would take 5 days. I would like to know if there is way to speed this up?
>
Calling coxph.fit directly would likely be faster.

Also, you probably don't need to do 2000 permutations on all 8000 genes: a
few hundred permutations is probably enough to decide that most of the
genes aren't interesting.

If you are going to be doing a lot of this sort of thing it might be worth
looking at the parallel processing facilities in the `snow' package.
There's a description of their use in another gene expression problem in
the new R Newsletter.


	-thomas