Authors did not use a formal hypothesis testing framework instead of fitting a ML model #137
Description
Relatedly, is there a reason that the authors did not use a formal hypothesis testing framework instead of fitting a ML model? It would be useful to report a version of Table 2 in which the authors examine the same classification tasks via a hypothesis testing framework with a null that can be rejected. PERMANOVA seems like a more straightforward way to test for associations between pairwise distances and groups, although it is possible that there are some subtleties of the data that we are missing. If PERMANOVA isn’t appropriate, it seems possible that even a permutation test on average pairwise distance within versus between sets of genes that share a property against randomly reshuffled labels could help contextualize both the magnitude and significance of the separation.