Best Practices for >2 treatments about datest

[Russel88]

Hi jjscarpa,

There is no google group, or similar, but it might be a good idea to start one.

The intention with DAtest was to choose one method to use. I don't think that using an ensemble approach with allDA is wrong, it's just harder to interpret. Also, running testDA gives you an idea of the specificity and sensitivity of single methods, not a combination. allDA was meant as a sanity check; do the top methods find nearly the same significant features. My advice would therefore be to simply use ds2.

Regarding DESeq2, the default with multi-class predictors is for ds2(x) to run a likelihood ratio test. While you only get one p-value, you can get log fold-changes for all comparisons:

Assuming your species are called "mouse" and "cat":
res <- DA.ds2(data.new, predictor = "Sp", allResults = TRUE)
res_mouse.vs.cat <- results(res, contrast = c("predictor","mouse","cat")@listData

Alternatively, if you want a p-value for each comparison you can set out.all = FALSE in both testDA, DA.ds2 and so on. ds2 will then run Wald tests and you will get a p-value for each comparison between species, but I think this is intended for two-class predictors or multi-class predictors with a common reference you wish to compare to. But you could in principle run it for each comparison:

res <- DA.ds2(data.new, predictor = "Sp", allResults = TRUE, out.all = FALSE)
res_mouse.vs.cat <- results(res, contrast = c("predictor","mouse","cat")@listData

But note that p-values are then adjusted for each comparison independently (unless you correct them yourself), which might inflate the false positive rate.