[BioC] Problem with limma's topTable lfc filter when more than 1 contrast is output

[Gordon K Smyth]

Dear Vladimir,

You are correct that topTable() has been ignoring the lfc argument when 
the ranking is by F-statistic, i.e., when ranking by several contrasts 
simultaneously.

Alex is correct as to the reason for this -- it was just ambiguous what 
should be done.  However the behaviour should have been explained more 
fully in the documentation.

I now have committed an update to the topTable() function so that the lfc 
cutoff will now be applied to the minimum absolute logFC for the contrasts 
being tested when ranking by F-statistic.  The behaviour is now analogous 
to decideTests().  Genes will be kept in the table if any contrast 
satisfies the cutoff.

I committed this change to the developmental version of limma a few days 
ago.  I committed the change to the official release version of limma just 
a few minutes ago, so it should be able for download from Bioconductor in 
a day or so.

Normally decideTests() mimics the behavour of topTable() applied to 
individual contrasts.  However you can recover the same genes that appear 
in an F-statistic topTable() ranking by using:

   result <- decideTests(fit, method="nestedF", ...)
   which.genes <- apply(result, 1, any)

The effects of p-value and lfc cutoffs should now be the same in both 
topTable() and decideTests().

Best wishes
Gordon


> Date: Sat, 21 Jan 2012 04:52:10 +0000
> From: Vladimir Zhurov <vzhurov2 at uwo.ca>
> To: <bioconductor at stat.math.ethz.ch>
> Subject: Re: [BioC] Problem with limma's topTable lfc filter when more
> 	than	1 contrast is in output
>
> Alex Gutteridge <alexg at ...> writes:
>
>>
>> On 20.01.2012 05:11, Vladimir Zhurov wrote:
>>>> ...
>>>
>>> Dear Bioconductors,
>>>
>>> I am having the following problem which can be due to a 
>>> misunderstanding, or an actual problem with topTable function in limma 
>>> package.
>>>
>>> As far as I understand lfc and p filters should work together in 
>>> filtering topTable results. Am I correct in this regard?
>>>
>>> If it is an intended situation then the problem is the following: when 
>>> more than one contrasts is reported lfc filter does not affect the 
>>> output. Which is shown in the sample R session below.
>>>
>>> I would appreciate you help.
>>>
>>> Regards.
>>>
>>> Vladimir.
>>>
>>
>> Hi Vladimir,
>>
>> I think the issue maybe that it is not clear (to me anyway!) what
>> result one would expect with a lfc filter on a table with multiple
>> contrasts. Should rows where all contrasts are above the cutoff be
>> reported? Or where any are? It doesn't seem to be explicit in the docs,
>> but my assumption is that to avoid confusion topTable ignores lfc= in
>> the case you describe.
>>
>> Perhaps decideTests() is what you are really looking for? You can then
>> index a topTable generated data.frame with the results of decideTests.
>> E.g (pseudo-code not tested!)
>>
>> tt = topTable(fit,p.value=1,number=Inf,sort.by="none")
>> dt = decideTests(fit,p.value=0.05,lfc=1)
>>
>> tt[dt[,1] != 0 & dt[,2] != 0 & dt[,3] !=0,] #Gives rows where all
>> contrasts pass filter (assuming 3 contrasts)
>> tt[dt[,1] != 0 | dt[,2] != 0 | dt[,3] !=0,] #Gives rows where any
>> contrast passes filter
>>
>
> Alex,
>
> Thank you for your input. I was considering decideTests but wanted to check
> whether there is something that can be done with topTable. I think that it would
> be nice at very least to explain intended behavior in more details in the manual
> for a case of multiple contrasts comparison.
>
> Anyway, I ended up with the following workaround (lets say lfc=1, p=0.05 and no
> adjustment of p value):
>
> tt.all<-topTable(fit.eb, sort.by="none", number=Inf, adjust.method="none")
> dt<-decideTests(fit.eb, lfc=1, p=0.05, adjust.method="none", method="separate")
> rs<-rowSums(abs(dt))
> tt<-tt.all[rs >= 1,] # a single contrast above lfc will be sufficient
>
> One may think that you are done at this point. However, decideTests with
> method="separate" recapitulates topTable performed on individual contrasts. As a
> result there are extra probes which will be reported and will have p value above
> cut-off when topTable is done across multiple comparisons. They need to be
> further filtered out.
>
> tt.filt<-tt[tt$adj.P.Val < 0.05,]
>
> Regards.
>
> Vladimir.

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