The figure illustrates the intuition behind Mimosa. Consider a TF gene X and a potential target gene Y. The expression values of X and Y for all expression samples are shown as a heat plot and as a scatter plot. We presume that X and Y expression are correlated only in an unknown subset of samples (depicted by "+") and not in the remaining samples (denoted by "-"). Mimosa computes the maximum likelihood partition of samples. Then given the sample partition, a third gene Z with differential expression between the two partitions may represent a potential modifier. To be precise, we assign a partition probability to each sample as opposed to a binary partition.