Our paper (with Julie Cartier, Johanna Lagoas, Youmna Ayadi, Adeline Fermanian and @flomass) on the use of statistical knockoffs for the differential analysis of transcriptomics data just came out, very appropriately as it nicely illustrates my point:
https://academic.oup.com/bib/article/27/3/bbag148/8687371
Using simulated outcomes on real transcriptomics data, we've shown that KOs (and in particular, the KOPI approach) do retrieve important variables with better power than classical approaches (Wilcoxon, Lasso), while controlling FDR.
However, all methods perform poorly when the relationship between gene expressions and outcome is nonlinear.
On real outcomes, the method is overly conservative (having no discoveries is a surefire way of controlling your number of false discoveries), and we had to turn the false discovery rate threshold to 50% to select any gene at all.
#machineLearning #genomics #featureSelection #biomarkerDiscovery #transcriptomics