and the maximum correlation of all offset comparisons was recorded. To select a representative subset of motifs for each TF regulator, we first found the motif correlated with the most other motifs at R>0.9. Treating that motif and all of the correlated motifs (R>0.9) as a group, we next found the motif with the greatest mean correlation to the other members of the group, and kept that motif as a representative motif for the TF. Motifs highly correlated with that chosen motif (R>0.9) were then discarded from further analysis, and the process was iterated until no motifs remained. We repeated the process using the medium and low-quality databases for TF regulators with no associated motifs in the high-quality database. The final curated motif database contains 1,764 human motifs and 1,346 mouse motifs representing 870 human and 850 mouse regulators. The resulting names are formatted as follows:: “ensemble ID”_”unique line number”_”common TF name”_”direct (D) or inferred (I)”_”number of similar motifs grouped”. These position frequency matrices were then converted into Position Weight Matrices (PWMS) by taking the log of the frequency after adding a 0.008 pseudocount and dividing by 0.25.
