Biological evaluation of the learned regulatory program.

We evaluated our learned regulatory programs relative to a reference set of regulatory interactions collected from various datasets that were not used by the Lirnet method (see text for more details). A prediction that a regulator r regulates a module m was considered as validated if there was significant overlap (hypergeometric p<0.01) between the members of m and the putative targets of r in the reference set above. For each method, we counted the number of validated interactions (column named # interactions) for module m containing ≥10 genes, where each entry shows: a/b (c%), where a is the number of significant regulators, b is the total number of predicted regulators that appear at least once in the reference dataset, and c is the proportion (a/b×100). We similarly counted the number of modules that have at least one validated regulator (column named #modules), relative to the total number of modules having a predicted regulator in the reference set. We also considered two-step regulatory cascades, as described in the main text. Table S4 shows this analysis for expression regulator and genetic marker regulators separately.