[BioC] Limma: Very high logFC values

[J.delasHeras at ed.ac.uk]

Quoting "Ingrid H. G. Østensen" <Ingrid.H.G.Ostensen at rr-research.no>:

> Hi
>
> I am using limma to analyze Illumina expression data (two groups),   
> and this time I got some really high logFC values for some genes and  
>  "low" for others. Example:
>
> Probe       log2 Ratio(logFC)     Moderated t-statistic (t)  Raw   
> p-value    Adjusted p-value     B
> ILMN_27575	5443.972	         27.305	               1.81E-06	       
> 0.009621899	-2.29002
>
> ILMN_14823        42.5	                 19.077	                 
> 1.00E-05	     0.022251754	-2.32116
>
> The first gene has intensities in one group (3 samples) around 10   
> 000 and in the other group (3 samples) around 5000, and the second   
> gene has intensities around 110 in the first group and around 80 in   
> the second group.
>
> I have never seen so high logFC values before, are they realistic?   
> Does this values mean that there are big differences combined with   
> hight intensities?
>
> Regards,
> Ingrid

I've never seen log2 ratios that high. In fact, I don't thing those  
are log ratios at all. They *could* be fold-changes, 'though.
I am not familiar with Illumina, but look at the possible ranges:

If the scanned images are 16-bit, it means you can have raw  
intensities measuring anything between 1 to 2^16 (65536). The most  
extreme ratio possible would then be 65536/1 (essentially saturated in  
one channel and no signal on the other). The log2(65536) is 16.

The only way to get a logFC higher than 16 is if you scan at higher  
bit-depths... but the log2FC is still going to be limited by the bit  
depth in teh same way: 20-bit images can give you a max logFC of 20.

I don't think you're scanning with >5000-bit resolution.

Maybe the Illumina platform processes the data in some way to expand  
the dynamic range, which could then give rise to larger logFCs. But I  
am not familiar with Illumina. Is that possible?
Algorithms to combine scans at different PMTs also result in an  
expanded dynamic range. Are you using such algorithm? Even in that  
case, I can't imagine teh expansion will be large enough to allow a  
logFC of >5000.

Fold-changes of 5000 can occur, however, from standard quantitation  
from 16-bit images... it would come from a probe/spot with nice signal  
on one channel and negligible on the other, after background correction.

So, to cut it short... I'd really make sure I am feeding Limma the  
right type of data, as the numbers look almost impossibly large. I'm  
pretty sure Illumina scans are 16-bit.

Jose

-- 
Dr. Jose I. de las Heras                      Email: J.delasHeras at ed.ac.uk
The Wellcome Trust Centre for Cell Biology    Phone: +44 (0)131 6513374
Institute for Cell & Molecular Biology        Fax:   +44 (0)131 6507360
Swann Building, Mayfield Road
University of Edinburgh
Edinburgh EH9 3JR
UK

-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.