[BioC] outlier removal from gene chip

[Sean Davis]

On 9/19/06 1:02 PM, "Weiwei Shi" <helprhelp at gmail.com> wrote:

> dear listers:
> 
> I have a question on whether bioconductor has some tool-kit to detect
> outliers and remove them.
> 
> my original dataset looks like this:
>             V1       V51       V53        V55       V57
> 1   -493249600  1.459459 -3.069444  -1.300000  1.935484
> 2  -1613096495 -1.139269 -5.525281 -16.592593 -1.831978
> 3   1626196571 -3.500000 -1.011662   2.223881  3.921053
> 4  -1397009217 -3.571429  1.685714  -1.180297 -6.807692
> 5   1428659728 -1.405405 -1.469004  -4.779754 -1.033708
> 6    459853658 -2.158879 -7.510823  -1.085581 -9.382979
> 7    530182506 -1.431677 -1.336343  -3.126437  4.878788
> 8   1173842263  1.215385  1.856410  -2.059794 -6.020833
> 9        28847  2.407895 -2.048889  -1.730337 -1.178947
> 10 -1961875610  2.864159 -2.301234  -4.733264 -1.172058
> 
> V1: internal probe id
> the rests are different samples. the cells are fold-change of disease/normal.
> 
> summary of the sample columns( V51, ... V57) gives the following:
>       V51                V53                 V55                V57
>  Min.   :-482.000   Min.   : -55.7342   Min.   :-122.074   Min.   :-14086.750
>  1st Qu.:  -2.159   1st Qu.:  -1.7312   1st Qu.:  -2.125   1st Qu.:    -1.831
>  Median :  -1.199   Median :  -1.0416   Median :  -1.200   Median :    -1.080
>  Mean   :  -0.918   Mean   :   0.1662   Mean   :  -1.027   Mean   :    -1.874
>  3rd Qu.:   1.441   3rd Qu.:   1.5721   3rd Qu.:   1.419   3rd Qu.:     1.521
>  Max.   : 198.434   Max.   :1478.1639   Max.   :  95.768   Max.   :   683.519
> 
> 
> My question is, is there any package which can detect those outliers
> (like -14086.750)and remove them and get an "average" for each gene
> (instead of each probe)?

Hi, Weiwei.  

The better option, probably, is to remove datapoints that are questionable
BEFORE making a ratio using good quality control, plots, etc.  Extreme
ratios may be biologically very important, so simply removing them is
probably not the best option.  I would suggest looking at the two data
values that went into making the ratios that you think are in question and
see if there is an explanation (for example, one probe of the two failed,
for example).  Simply removing ratios because they look like outliers is
potentially removing your most interesting data.

Sean