[Bioc-devel] Can I analyze with bioconductor a microarray experiement where the distribution of probes intesisties follow a bimodel distribution?

Sat, Jun 8, 2013 6:46 PM

Dear Miguel,

No, this does not follow at all, nor it is needed.  Standard microarray 
analysis methods (eg limma package) make no assumptions at all about the 
distribution of intensities over all probes.

You should however be checking whether you are using the best possible 
background correction and preprocessing methods.

No, there is no such implication.  The within probe (over samples for each 
probe) and between probe (over probes for each sample) distributions are 
almost unrelated.

No they don't.  The popular RMA algorithm for Affymetrix data, for 
example, assumes that the distribution of intensities for each probe 
approximately follows an exponential distribution, and that is not 
normal-like.

Best wishes
Gordon

BTW. This is not the right list for this sort of question.  You should be 
mailing to the main Bioconductor list.

Thank you again,

Miguel

-----Original Message-----
From: James W. MacDonald [mailto:jmacdon at uw.edu]
Sent: Friday, June 07, 2013 5:27 PM
To: Miguel Moreno-Risueno
Cc: bioc-devel at r-project.org
Subject: Re: [Bioc-devel] Can I analyze with bioconductor a microarray
experiement where the distribution of probes intesisties follow a bi modal
distribution?

Hi Miguel,

On 6/7/2013 5:11 AM, Miguel Moreno-Risueno wrote:

Hello all,

We have recently received a microarray experiment in the Nimblegen
platform where the intensity of the probe sets follow a bi-modal
distribution. We have been said from the facility that this is because
of the dynamic range of the Agilent scanner they use. We are concerned
about the statistical analysis with bioconductor as it is our
understanding that these statistical analyses are developed for normal
or normal-like distribution. We appreciate any information on this regard.

If I understand your question correctly, you are noting that the overall
distribution of probes within a sample has a bi-modal distribution. This
doesn't really have anything to do with any statistical tests you might be
computing, as you are not doing any statistics within a sample (e.g., one
usually doesn't test to see if probe X is differentially expressed as
compared to probe Z in sample Q).

Instead, what you should be concerned with are the distributions of the
individual probes across samples. With microarray data we usually don't have
enough data to even begin to assess the across-sample, within probe
distributions (e.g., if you have three replicates for two sample types, good
luck trying to discern if those probes follow a normal distribution, or are
even 'hump-shaped'). In addition, there are usually tens of thousands of
probes on a given chip. I have never heard of anybody looking at each probe,
trying to assess if it follows a reasonable distribution across samples. I
suppose you could do it, but to what point?

Instead we simply assume that the data follow a reasonable distribution and
then do the test. This is one of the reasons that it is imperative to follow
up promising leads with confirmatory testing, preferably with new samples.

Best,

Jim

Thank you in advance for your help,

Miguel

______________________________________________________________________
The information in this email is confidential and intend...{{dropped:4}}