[BioC] Do my Limma results look "normal"?

[Gordon K Smyth]

Dear Paul,

> Date: Thu, 5 Jun 2008 13:42:40 +0100
> From: "Paul Geeleher" <paulgeeleher at gmail.com>
> Subject: [BioC] Do my Limma results look "normal"?
> To: Bioconductor <bioconductor at stat.math.ethz.ch>
>
> Hi,
>
> This is the first time I've ever analyzed a microarray experiment
> using Limma (or anything else for that matter) and I was hoping that
> somebody could look at my results and tell me if they look normal.

You're asking a question that doesn't really have an answer, because all 
experiments are different and give different results.  Your results 
suggest a lot of probes are strongly DE, with a predominance of down over 
up regulated results.  You're the only one who knows the background to 
your experiment, so you're the only one who knows whether this makes 
sense from a biological point of view.

> The experiment is measuring differential expression between miRNAs of
> HER2+ and HER2- breast cancer tissue. There are 3 HER2+ arrays and 4
> HER2- arrays and each of the 399 miRNAs is replicated 4 times in each
> array.
>
> TopTable() reveals the following miRNAs with a fold change above 1.5,
> which I thought was a reasonable cutoff:

If you want a fold change of 1.5, you need lfc=log2(1.5) not lfc=1.5.

>                  ID     logFC         t      P.Value    adj.P.Val          B
> 273      hsa-miR-451 -4.645060 -8.226854 4.510441e-09 9.246404e-07 10.8484797
> 128      hsa-miR-205  3.551495  7.574564 2.370061e-08 3.239083e-06  9.2222865
> 13       hsa-miR-101 -2.310652 -6.569497 3.374177e-07 2.567796e-05  6.6146751
> 282      hsa-miR-486 -2.686910 -6.542808 3.626060e-07 2.567796e-05  6.5439656
> 55       hsa-miR-144 -2.890719 -5.889594 2.152998e-06 1.261042e-04  4.7952480
> 387      mmu-miR-463 -2.609257 -5.764143 3.042120e-06 1.559086e-04  4.4561920
> 388      mmu-miR-464 -2.080402 -5.696976 3.662006e-06 1.668247e-04  4.2743601
> 151      hsa-miR-223 -1.722956 -5.637290 4.318942e-06 1.770766e-04  4.1126276
> 51    hsa-miR-142-3p -3.262824 -5.397809 8.386312e-06 3.125807e-04  3.4626378
> 14   hsa-miR-101_MM1 -1.922710 -5.224075 1.358743e-05 4.175776e-04  2.9905370
> 159  hsa-miR-26b_MM2 -2.221853 -5.206724 1.425875e-05 4.175776e-04  2.9433849
> 236 hsa-miR-376a_MM1 -1.633555 -4.653220 6.637043e-05 1.700742e-03  1.4433277
> 266     hsa-miR-432*  1.512622  4.627293 7.131510e-05 1.719952e-03  1.3734422
> 168      hsa-miR-29b -1.954087 -4.198854 2.323860e-04 4.763912e-03  0.2280262
> 31  hsa-miR-126*_MM2 -1.537988 -3.209957 3.233842e-03 5.099520e-02 -2.2888897
> 52    hsa-miR-142-5p -1.881192 -2.831493 8.332384e-03 9.002153e-02 -3.1731794
>
>
> Another person is sanity testing this data using GeneSpring and they
> are getting much higher p-values compared to mine.

This is not surprising considering that GeneSpring has chosen not to use 
any statistical test invented since 1947.  In particular, it has not taken 
any advantage of the last 8 years' intensive research on differential 
expression for microarray data.

> They are also taking the step of excluding quite a few of the miRNAs 
> from the experiment based on their standard deviation across the arrays 
> of each group. Should I be doing this also or is this taken into account 
> by the eBayes() function or lmFit()?

You could choose to filter on raw standard deviation across all arrays. 
Some authors recommend this.  Or you could filter on mean intensity. 
With only 399 probes on your arrays, I doubt either of these things would 
make much difference, but they might.

It does not make sense to filter on standard deviation computed within 
groups.

Best wishes
Gordon

> If you are interested the script I wrote to do the analysis is here:
> http://article.gmane.org/gmane.science.biology.informatics.conductor/18032/match=miRNA
>
> Thanks for any advice,
>
> -Paul.