I am willing to analyse a fraction (63 controls, 19 diseased) of samples gathered in the microRNA profiling study - GEO GSE31568 http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE31568). The data are summarized (across each probe for specific miRNA, I believe) and normalized (quantile) by the submitter. There are no control probes, nor raw data containing background values.
Since roughly 15% of all reads are zero intensities, I would like to perform some filtering before applying limma. The point is NOT to get more significance, but to improve the data quality. Of note, applying limma to this data would give me ~300 DEGs with p < .05. limma manual advises to use kOverA approach (which in this case would be 19 over 0), but it excludes only a few miRNAs from the total of ~850.
What are the possible alternatives or diagnostics? I would like to avoid filtering by variance. Can I perform quantile filtering by mean intensities?