[BioC] Limma: background correction. Use or ignore?

[J.delasHeras at ed.ac.uk]

Dear Gordon,

many thanks for your helpful reply and this link:

> For some brief but relevant comments see See Section 6.1: Background
> Correction in the Limma User's Guide, and Section 3 of
> http://www.statsci.org/smyth/pubs/mareview.pdf


> Whether background subtraction is a good idea depends entirely on the
> background estimation used. You do not mention what image analysis
> program you used or which background estimation method was chosen,
> but everything depends on this.

It varied.
Initially I used TIGR Spotfinder (SF), using its Otsu algorithm (of 
which I don't know much, but I understand it's a variation of the 
histogram method that takes into consideration the physical 
distribution of the pixels, so that only pixels that group into 
something resembling a spot are considered -proximity to one another, 
proximity to the centre of the grid, and also considering estimates of 
spot sizes given by the user). I used to simply substract the 
background estimated by SF, which also turns negative values into zero.

This always gave me decent results, although now I realise that not 
doing background substraction may improve them.

More recently, coinciding with the use of a deifferent set of arrays, I 
started using a different scanner (Axon 4200AL, prior to that I used an 
ArraywoRx one), and Genepix 6.0. I used its default background 
estimation (local background median, 3 feature diameters wide area, 
except for the area surrounding the spots, 2 pixels wide). The spots 
are located using the "irregular features" option. Here's where the 
trouble started, when I substracted teh background estimated in Genepix 
from Limma. Simple substraction.

It took me a while to notice there was a problem because I was using 
different arrays, and the experiments were also different and more 
prone to variation. But when I re-did some simple experiments whose 
results I could expect with reasonable certainty, it showed me very 
poor B values... and my worries started!

I just did not expect that the background correction would have such a 
profound effect. When I compared quantitation made by SF and Genepix, I 
only looked at the raw intensities, not the background... the 
intensities looked reasonably comparable.

Right now I am re-quantitating some images using Genepix and Spotfinder 
to compare the differences estimating foreground and background, but I 
still don't have that data.

> Firstly, can you get away with ignoring the background entirely? I
> agree with Jim and Naomi's general remarks, and I agree with Jim that
> not background correcting can lead to cleaner results for some data
> sets, especially for good quality arrays with low background. The

I think this is where I am leaning towards, at the moment. My slides 
are quite clean and the background images as displayed by Limma show 
the background is quite uniform too.
It certainly seems like the simplest option. Although you point at some 
reasons not to do that...


> UCSF microarray center has made the same argument for their own
> arrays. But in my lab, we always background correct. There are a lot
> of reasons for this. For one thing, foreground-background plots
> almost always show that background correcting does remove some
> systematic bias.

what do you mean exactly by this? what's teh source of this bias?
dye-specific bias?

> The most critical reason though is to achieve
> comparability between experimental conditions. Not background
> correction is a lot like adding an offset to your data (see the
> backgroundCorrect function in limma), and the size of the offset
> depends on the level of the background. In my lab we see data from
> lots of different labs, platforms, image analysis programs, species
> etc, and the background levels can vary wildly. For example, I
> analysed one important experiment when the scanner changed from Axon
> to Agilent halfway through, and the overall background levels
> increased 10-fold. I prefer to background correct and to add the
> offset explicitly, rather than to allow it to vary with the data in
> an uncontrolled way. Had I not background corrected the Axon-Agilent
> experiment, the results would have been far more damped in the second
> half of the experiment and not comparable to the first.

I understand how in a case like this it would be very important to 
account for the variability of background measurements.

Usually my experiment are performed ina  relatively short period of 
time, so teh scanner/analysis program is the same for each individual 
experiment. I do observe that the background levels vary between 
slides, but usually not a lot (although, what is "a lot"?)... which is 
the reason I am thinking I will probably end up not background 
correcting. But since it seems to have such a marked effect, I don't 
want to make a decision like this without investigating a little more.


> But background correcting doesn't mean that we simply *subtract* the
> background. We subtract if we have
> 1. morph background from SPOT
> 2. morphological opening background from GenePix 6, or
> 3. background from AgilentFE
> and in no other cases. In most other cases, subtracting is so bad
> that you would indeed probably be better off ignoring the background 
> entirely.


In what I am testing now, I included background measurements using the 
morphological option in GenePix, as well as the default local bkg, and 
using negative spots as controls, for comparison. I haven't analysed 
the data yet.


> In most other cases we currently use 'normexp' background correction
> with an offset. This is an adaptive background method which is a
> modification of the background correction method used by the RMA
> algorithm for affy data. It is adaptive in that it adapts to the
> overall level of background on each array. It avoids the negative
> intensities which so often arise from naive background subtraction.

I imagined you'd be a fan of this method ;-)
I must say I haven't tried this one yet, but I will. My "problem" to 
try this method was how to determine the offset. I am guessing I just 
have to try different values and check the effect...


> The Bioconductor book has an example of a data set which was analysed
> both with no background correction (Chapter 4) and with normexp
> background correction (Chapter 23).

I'll be sure to read that.

Many thanks, Gordon. Very helpful.

Jose

-- 
Dr. Jose I. de las Heras                      Email: J.delasHeras at ed.ac.uk
The Wellcome Trust Centre for Cell Biology    Phone: +44 (0)131 6513374
Institute for Cell & Molecular Biology        Fax:   +44 (0)131 6507360
Swann Building, Mayfield Road
University of Edinburgh
Edinburgh EH9 3JR
UK