3.8 years ago by
Cambridge, United Kingdom
It should be okay to use ROAST in this application, as the definition of the gene set (from one experiment) is independent of the application of the gene set in ROAST (in the second experiment). This means that you're not doing any inappropriate data snooping, e.g., by running ROAST on the same data that you used to define the gene set.
In practice, I would define my DGE list from the first experiment, and then split it into two separate DGE lists based on the sign of the log-fold change. I would then use each of the two sublists in ROAST for the comparison in the second experiment. If the experiments have similar DE profiles, then the "up" sublist should have a consistent direction of "up" in the ROAST output, and similarly for the "down" sublist.
Besides intersection, no "sophisticated approaches" come to mind. One (visual) alternative might be to simply plot the log-fold changes of one experiment against the log-fold change of the other experiment for all genes. If you have similar DE profiles between experiments, then you should see a more-or-less straight line through the origin. You can also compute the correlation coefficient to get some quantitative measure of similarity. You don't have to worry about technical correlations as you've got independent experiments.
roast can now accept a vector of
gene.weights. This is intended to accept log-fold changes associated with the gene set, and can have both positive or negative entries. Thus, you don't have to split the DGE list from the first experiment to define your up/down sets. Rather, you can use the entire set in ROAST, along with the vector of log-fold changes for all genes in the set.
modified 3.8 years ago
3.8 years ago by
Aaron Lun • 23k