It would be useful to have a method similar to most_correlated but accepting categorical variables as a reference (for instance to identify features associated the most with the experimental condition). Finding good metrics is however less straightforward (AUROC, cross-entropy, regression, ...).