Fig. 1.

Using Differential Peak Prediction to Identify Motifs of Different DNA Binding Domains. To identify the motif of a DBD, we train a deep convolutional neural network to predict whether a TF ChIP-seq peak is preserved or significantly weaker in a dataset from that TF with a mutated DBD relative to a dataset from the wild-type TF. We then use DeepLIFT followed by TF-MoDISco to identify the motifs that the neural network learned. The browser tracks in this figure are pooled replicate fold-change bigWigs from wild-type CTCF and CTCF with ZF1 mutated from [11] and were visualized using the WashU Epigenome Browser [35] with assembly mm10 [36]. The motif logo in this figure is the motif we discovered when interpreting the model for the ZF 1 mutant