[BioC] Filtering gene list prior to statistical testing

[Johan van Heerden]

Dear All,

I have scoured the BioC mailing list in search of a clear answer regarding the filtering of a data sets prior to differential testing, in an attempt to circumvent the multiple testing problem. Although several opinions have been expressed over the last couple of years I have not yet found a convincing argument for or against this practice.  I would like to make a comment and would appreciate any constructive feedback, as I am not a Statistician but a Biologists.

As far as I can see the problem has been divided into 2 categories: (1) "Supervised" and (2) "Unsupervised" filtering, where (1) is based on some knowledge regarding the functional classes present in the data, as opposed to (2) which does not consider any such information.  Several criticism have been raised against the "Supervised" approach, with many people calling it flawed logic. My first comments are regarding the logic of "Supervised" filtering.

As an example: A data set consisting of two classes (Treatment 1 and Treatment 2) has been generated.  A fold-change is then used to enrich the data set for genes that show within class activity (i.e. select only genes that show a mean x-fold change between classes). This filtered data set is then used for differential testing.  

My first question is: How is this different (especially when working with "whole-genome" arrays) from having custom arrays constructed from genes known show a response to some treatment. I.e. Arrays will then be selectively printed with genes that are known to or expected to show a response. This is a type of "filtering" step that will yield arrays with highly reduced gene sets. This scenario can result from known knowledge about pathways or can arrise from a discovery based microarray experiment, where a researcher produces whole genome arrays and from there select "responsive" genes for the creation of targeted (or custom arrays). Surely this step-wise sample space reduction should be subject to the same criticism?

Secondly, the supervised fold-change filter should not affect the statistic of each individual gene, but will have profound effects on the adjusted p-values. I have checked this only for t-tests and am not sure what the effect on more complex statistical differential testing methods would be. If the only effect of the "supervised" filtering step is the enrichment of class-specific responsive gene and a reduction in the severity of the p-value ADJUSTMENT (without affecting the actual statistic), this could surely be a very useful way of filtering data?

Wrt the "unsupervised" approaches: These approaches define some overall variability threshold which can be used to filter out genes that don't show a minimum degree of variability regardless of class. As far as I can tell there are several issues wrt this approach. (1) Some genes will be naturally "noisy", i.e. will show high levels of fluctuation regardless of class. These genes are likely to be included in a filter based on degree of varilablity. (2) Some genes might show low levels of variability (with small changes between classes) and could be important, but will be excluded if a filter is based on degree of variability.

I would greatly appreciate some feedback on these comments, specifically some statistical substantiation as to why a "supervised" approach is "flawed", given the similar experimental strategies included in the paragraph on this approach.

Many Thanks!!
Johan van Heerden