Entering edit mode
Bogdan Tanasa
▴
10
@bogdan-tanasa-2475
Last seen 9.6 years ago
Thanks everyone for the precious input ! I appreciate it !
Bogdan
--
Bogdan Tanasa, MD
Kellogg School of Science and Technology,
The Scripps Research Institute,
10550 North Torrey Pines Road,
La Jolla, California 92037
On 11/2/07, Robert Gentleman <rgentlem at="" fhcrc.org=""> wrote:
Hi,
If they were assayed at approximately the same time, using
approximately the same protocols then yes, one normalization is
likely
to be better than two. I think that there may also be issues if
the set
of genes that are expressed is very different in the different
tissue
types (as them being the same is one of the basic assumptions in
most
normalization methods). But if very much is different, then it is
better
not to try and normalize, but rather to adjust after
normalization.
best wishes
Robert
James W. MacDonald wrote:
> Yes but if I am not mistaken, the OP had a situation in which
the
> samples were simply different cell or tissue types, rather than
> different batches. I this case I would favor normalizing all
together
> rather than doing things in batches.
>
> Best,
>
> Jim
>
>
> Robert Gentleman wrote:
>>
>> Naomi Altman wrote:
>>> Dear Bogdan,
>>> Any normalization method that uses a set of arrays, reduces
the
>>> variability among those arrays.
>>>
>>> So, if you have 2 sets of arrays and normalize separately, you
will
>>> find that the within set variability is smaller than the
between set
>>> variability - i.e. you induce significant differential
expression
>>> simply by the normalization. To avoid this effect, when you
are
>>> doing differential expression analysis (or sample clustering)
you
>>> must either use methods that normalize each array separately
(MAS) or
>>> normalize all together.
>>
>> An alternative (and the one that I prefer) is to do separate
>> normalizations, and to then use some sort of batch effect term
in the
>> model used to assess differentially expressed genes.
>>
>> Normalization is intended to clean up the relatively minor
issues
>> that arise due to slightly different conditions etc. for arrays
that
>> are essentially the same. As far as I can see it is not
intended to
>> adjust for batch effects, and in my experience generally does a
bad
>> job of that. Just because you can normalize (or fit any
statistical
>> model) does not mean that you should.
>>
>> best wishes
>> Robert
>>
>>
>>> --Naomi
>>>
>>> At 12:01 PM 11/2/2007, Bogdan Tanasa wrote:
>>>> Greetings Naomi,
>>>>
>>>> thanks for reply. To generalize my question: when dealing
with 2
>>>> sets of
>>>> samples, let's say X1, X2, ...., Xn and Y1, Y2, ..., Yn,
>>>> I could run the normalization in 2 ways: A. only X(1,n) and
only
>>>> Y(1,n), or
>>>> B. both X(1,n),Y(1,n). Are there any a priori statistical
>>>> criteria that favors a way or the other ? If I would take
into
>>>> consideration biological criteria (the things I am interested
in), the
>>>> results
>>> >from A may sometimes look better than B', or vice versa.
Thanks !
>>>> Bogdan
>>>>
>>>>
>>>>
>>>> On 11/2/07, Naomi Altman <naomi at="" stat.psu.edu=""> wrote:
>>>>> Dear Bogdan,
>>>>> I do not have an opinion on gcRMA versus RMA. But if you
are doing
>>>>> differential expression analysis comparing the cell samples
with
the
>>>>> organ samples, you need to normalize
>>>>> all the samples together.
>>>>>
>>>>> --Naomi
>>>>>
>>>>> At 11:31 AM 11/1/2007, Bogdan Tanasa wrote:
>>>>>> Hi folks,
>>>>>>
>>>>>> I would like to ask for your opinions on the following:
>>>>>>
>>>>>> I have 60 expression profiles of 60 samples (cells and
organs in
>>>>>> resting conditions).
>>>>>> I normalized these arrays in many ways, including RMA.
>>>>>>
>>>>>> Considering the biological arguments (cells samples vs
organs
>>>>>> samples), I am planning to do the normalization separately,
on the
>>>>>> group of cell samples, and on the group of organ samples.
>>>>>>
>>>>>> My questions are:
>>>>>>
>>>>>> - after RMA normalization on separate groups of samples
(cells vs
>>>>>> organs), the results are different, but are these better ?
GO
>>>>>> analysis
>>>>>> do not display major differences.
>>>>>>
>>>>>> - would gcRMA work better than RMA ? The majority of
opinions in
>>>>>> SoCal
>>>>>> are pro-RMA.
>>>>>>
>>>>>> thanks,
>>>>>>
>>>>>> Bogdan