[BioC] RMA normalization,which samples should be normalized together

[Naomi Altman]

Dear Johannes,
Actually, technical replication is of little interest when you have 
biological replication.  If I understand your experiment, you have 40 
patients, each measured at 2 times.

Because of the pairing, you have several options for appropriate 
normalization and analysis:

1) Normalize the before and after for each patient together, and analyze M.

You could use either RMA, or a simpler M vs A loess for this.

2) Normalize all the arrays together and then compute M for each patient.

I would use RMA or gcRMA for this.

In either case, I would simply use limma with the 
contrast   rep(1,npatients) since this gives the t-test for before-after 
which seems to be of most interest.  Limma has an advantage over ordinary 
t-tests in that it combines some information across genes.  However, I 
expect it to be very similar to ordinary t-tests (or Wilcoxon tests) 
because you have a fairly large sample size.  Any of these methods are 
appropriate.

Incidentally, the technical reps are interesting for quality control, but 
should not be included in this analysis.

--Naomi

At 09:48 AM 2/7/2005, Dipl.-Ing. Johannes Rainer wrote:
>thanks arne
>
>i have no replicates, affymetrix is still a little bit expensive ;) . all 
>our chips were made by ourself and by looking at the histograms of the raw 
>values there are no differences at all. in the whole experiment we made 
>also two replicates, one with the same RNA, but different amount before 
>amplification (one time 5 mug, the second time 1 mug) and the second 
>replicate is RNA from the same patient, same time point, but the RNA was 
>extracted by two different people  not at the same time. if i normalize 
>only those replicated chips i see nearly no differences between them (with 
>a M (log2 regulation value) cut off of M=1 i get about 30 probe sets that 
>differ), but when i normalize all 80 chips of all patients together the 
>replicated chips show more differences... in my opinion i have to 
>normalize all patient chips together, exspecially if i want to do for 
>example a wilcox between all 0 hour and 6 hours chips.
>can you tell me a little bit more about the linear model you have used to 
>merge the results?
>
>regards, jo
>
>
>Quoting Arne.Muller at sanofi-aventis.com:
>
>>Dear Johannes,
>>
>>I've a study with 84 affy chip to characterize a dose effect of a drug. 
>>The study was conducted in 3 different laboratories. There are strong 
>>differences betweent the laboratories and I've RMA normalized per 
>>laboratory and then merged the results in a single linear moel including 
>>the laboratory as an additional factor. Maybe you can make the patient or 
>>source of RNA a random factor in a mixe effects model - if you've 
>>replication per patient.
>>
>>Just looking at those genes with a significant dose effect I did not find 
>>much differences between normalizing all chips together and
>>normalizing per laboratory.
>>
>>         regards,
>>
>>         Arne
>>
>>
>>>-----Original Message-----
>>>From: bioconductor-bounces at stat.math.ethz.ch
>>>[mailto:bioconductor-bounces at stat.math.ethz.ch]On Behalf Of Dipl.-Ing.
>>>Johannes Rainer
>>>Sent: 07 February 2005 10:13
>>>To: bioconductor at stat.math.ethz.ch
>>>Subject: [BioC] RMA normalization,which samples should be normalized
>>>together
>>>
>>>
>>>hi,
>>>we are interested in the response of patients to a special treatment,
>>>so we have patient samples before and after treatment. i have
>>>normalized this samples in different ways using RMA. As RMA tries to
>>>detect and correct probe effects by looking at the expresison
>>>levels of
>>>the probes across all chips it is not surprising that the outcome of
>>>the analysis differs depending on which chips i normalize together.
>>>It is clear that i have to normalize all patient samples
>>>together if i
>>>want to compare the expression values of the genes (lets say using
>>>statistical tests). i am also analyzing the chips using the 'old
>>>fashioned way' by using M and A values and i suppose it is not
>>>problematic at all to compare M values of lets say patient 1, 6 hours
>>>sample against 0 hours sample with those from patient 2, also 6 hours
>>>versus 0 hours where the chips from the two patients were NOT
>>>normalized together.
>>>
>>>-now my question is if someone else has experience in what samples
>>>could and should be normalized together with RMA. I saw that ther are
>>>(big) differences in the regulation (M) values if i normalize two
>>>different patients together compared with the values that i
>>>get when i
>>>normalize only samples from the same patients together.
>>>
>>>thanks in advance
>>>
>>>_______________________________________________
>>>Bioconductor mailing list
>>>Bioconductor at stat.math.ethz.ch
>>>https://stat.ethz.ch/mailman/listinfo/bioconductor
>
>_______________________________________________
>Bioconductor mailing list
>Bioconductor at stat.math.ethz.ch
>https://stat.ethz.ch/mailman/listinfo/bioconductor

Naomi S. Altman                                814-865-3791 (voice)
Associate Professor
Bioinformatics Consulting Center
Dept. of Statistics                              814-863-7114 (fax)
Penn State University                         814-865-1348 (Statistics)
University Park, PA 16802-2111