Fig. 6.

Nanog motifs in H1-hESCs derived using TF-MoDISco with GkmExplain scores, HOMER, MEME and the method of Gandhi et al. Letter heights are proportional to the information content of the probabilities across the different bases at each position in the motif. ‘Seqlets’ are subsequences of high importance that are used by TF-MoDISco to create motifs (Shrikumar et al., 2018). The number of seqlets contained within each TF-MoDISco motif is indicated. The TF-MoDISco motifs were derived using only importance scores from the test set, while HOMER and MEME used the full set of training and test sequences. We find that TF-MoDISco run on GkmExplain importance scores tends to produce consolidated motifs that have a strong resemblance to the canonical motifs. Note that the underlying GkmExplain model used 11-mers, which might be why the Oct4-Sox2-Nanog motif, which is wider than 11 bp, is learned as two separate motifs by TF-MoDISco. Also note that the Sox2 motif, while present in the MEME results, did not meet the E-value threshold for significance and was 9th in the list of 10 motifs, ranking below several non-significant motifs that had very few supporting instances