[BioC] Finding similarities and differences for more than one catergory

[Daniel Brewer]

Sean Davis wrote:
> On Thursday 14 September 2006 07:11, Daniel Brewer wrote:
>> Hi,
>>
>> I have results from a series of 2-colour microarray experiements that
>> compare reference RNA to RNA from cells that fall into 4 catergories:
>> Cancer CD133+
>> Normal CD133+
>> Cancer CD133-
>> Normal CD133-
>>
>> What I would like to find genes that are:
>> 1) Significantly different from the reference RNA
>> AND
>> 2) either (in both CD133+/- seperately)
>> 	i) significantly different between cancer  and normal
>> 	or ii) significantly _similar_ between cancer and normal
>>
>> I have been thinking of using the following strategy:
>> 1) Treat CD133+ and - results separately
>> 2) Use results from lmFit to filter out genes that are not significantly
>>  different from reference RNA in BOTH Cancer and normal
>> 3) Perform a t-test between cancer and normal results and take genes
>> with p>0.05 as significantly different and p>0.95 as significantly similar.
> 
> There is not a good test to show that a gene is "unchanged" between two 
> groups, so point 2.ii doesn't really make sense.  In hypothesis testing 
> terms, using t-tests or the like allow you to "regect the null hypothesis" 
> with a given amount of certainty.  However, NOT rejecting the null hypothesis 
> (of differential expression) is NOT the same as proving the null hypothesis, 
> no matter how non-significant the p-values are.
> 
>> Is this a reasonable approach or would it be better to use ANOVA or
>> regression analysis.  To add to the complexity at some point I would
>> also like to compare the CD133+/- samples
> 
> Using all the data simultaneously is the better way to go, so use limma (or 
> some other package) to treat the data as the two-factor experiment that it 
> is.  
> 
> Sean

Thanks for the quick reply.  I must admit that I am slightly confused by
 the fact that there is no way to test that a gene is unchanged between
two groups.  Can't you just set the null hypothesis that the two groups
are different?  Or is that not possible.  Is there any other statistical
similarity metrics that I could use?

Many thanks again.

Dan

-- 
**************************************************************

Daniel Brewer, Ph.D.

Institute of Cancer Research
Molecular Carcinogenesis
MUCRC
15 Cotswold Road
Sutton, Surrey SM2 5NG
United Kingdom

Tel: +44 (0) 20 8722 4109
Fax: +44 (0) 20 8722 4141

Email: daniel.brewer at icr.ac.uk