[BioC] limma question: direct two-color design & modeling individual subject effects
Paul Shannon
pshannon at systemsbiology.org
Wed May 2 01:44:28 CEST 2007
Hi Gordon,
Thanks very much for your reply. Your suggestion makes sense --
that I should explore the use of lm () on one gene's MA$M values, until I can
fit the experimental factors I care about, and only then return to
limma. When I do get that far, I'll probably be back with another
question or two. :}
You asked a couple of questions about some confusing aspects of my
plot at http://gaggle/pshannon/tmp/7346.png (which I have since
revised):
> ... you haven't explained exactly what is in your picture. The
> data values on the y-axis don't appear to be the M-values you used
> to fit the linear model, because we don't see the up-down pattern
> we'd expect to see from dye-swaps.
I failed to explain that I multiplied all of the measured dye-swaps
(cy5/cy3) by -1 so that their ratios would be easy to compare to the
(cy3/cy5) ratios.
> How have you obtained "fitted values"?
I multiplied the model by the fitted coefficients, summed the resultant vector,
and corrected for dye swap:
# corrector = 1 or -1
corrector * sum (model [slide,] * efit$coef [row,]))
> Note that M-values are already log-ratios, so it doesn't make sense
> to write "log2M".
Good point...
Thanks again for your reply! I'd reached the end of my own resources; now I can
get to work again.
Cheers,
- Paul
>
> lmFit() simply does least squares regression. It gives the same coefficients that you would get
> from lm() for each gene. I suggest that you extract the M-value data for one gene, and experiment
> with fitting the data using lm(), until you're satisfied that you understand the parametrization
> and fitted values.
More information about the Bioconductor
mailing list