[BioC] Subtraction with NA values

[Gordon Smyth]

At 03:28 PM 11/02/2004, A.J. Rossini wrote:
>Gordon --
>
>Is it really standard?

It is for us. We try to balance for dye-effects but seldom find them very 
important in themselves.

>  Mathematically, I understand it, but
>statistically/scientifically, how does that jibe for trusting the
>order of calls with/without those genes?   We've been seeing some
>really interesting dye effects, though it isn't clear how much is
>background and cross-hyb vs. low-copy results (2-channel agilent
>chips).

I'm not entirely sure what you mean here - maybe send me more details 
off-line and we can discuss?

Cheers
Gordon

>   Of course, restricting to the genes above median expression,
>or some slightly more advanced filtering, seems to solve this
>problem, but I'm still worrying about the use of averaging for dye
>effects.
>
>best,
>-tony
>
>
>Gordon Smyth <smyth at wehi.edu.au> writes:
>
> > This is pretty standard analysis with linear models (limma package):
> >
> > dat <- cbind(M', M)
> > fit <- lmFit(dat, design=c(1,-1))
> >
> > Then fit$coef contains the combined log-ratios you want.
> >
> > Gordon
> >
> > At 02:54 PM 11/02/2004, A.J. Rossini wrote:
> >>"Daniel F. Simola" <simola at mail.med.upenn.edu> writes:
> >>
> >> > Hello,
> >> >
> >> > I have a microarray experiment using dye-swapped slides. I am trying
> >> > to combine (average) the intensities of a gene from a slide and its
> >> > dye-swapped pair, but just discovered that the subtraction operator in
> >> > R does not work the way I would like it to for missing (NA) values.
> >> >
> >> > I am doing: ( M - M' ) / 2, where M is an array of intensities for
> >> > genes, and M' is the same, except dye-swapped.
> >> >
> >> > Say I want the result of " 5 - NA ", where 5 is the intensity of one
> >> > spot and NA is that of the same spot on the dye-swapped slide, then I
> >> > get NA for an answer. Because I want to average the values ( 5 - NA /
> >> > 2 ), then I would like my average value to be 5, instead of NA. Thus
> >> > it's better to make use of the available data than disregard a gene
> >> > completely.
> >> >
> >> > So, does anyone know either of a workaround for this, or of a function
> >> > that I can use to perform element-wise subtraction over a matrix that
> >> > will work how I want (or that will let me define my own function to be
> >> > applied on an element wise basis)?
> >>
> >>Comments:
> >>
> >>1. you could impute 5 by replacing the missing values with the values
> >>from the other slide.  Judicious application of is.na and assignment
> >>will help with this.
> >>
> >>2. Do you really want to do this?  The end result will be some
> >>dye-corrected genes, and some dye-non-corrected genes (don't get me
> >>started on whether this is a reasonable thing to do, I still think the
> >>jury is out).  and then you want to compare how extreme they are
> >>(so you've got some genes with more inherent variance in the measure,
> >>not being averages, and if there is a gene by dye effect...
> >>
> >>Now, I wish I had a positive suggestion, and would appreciate hearing
> >>any (rather than the negative one I have above!).
> >>
> >>best,
> >>-tony