Fig. 6
From: Visualizing complex feature interactions and feature sharing in genomic deep neural networks
Visualization of sequence features captured by the 480 channels in 2nd convolutional layer of DeepSEA. The sequences are generated using gradient ascent (see section 1). The matrix represents -log10 of TOMTOM Q-values for each pair of channel and its top matching motifs. Each row represents a known JASPAR motif which has been ranked as top 1 matching motif for at least one of the channels. Only pairs that achieve less than 0.005 Q-value are represented with actual Q-value, and the dark blue region represents default value for low Q-values. In the right panel, the left column shows the SeqLogo visualizations of representative gradient ascent outputs of 5 of the channels, and the top matching motifs are shown in the right column. Channel 116 and 451 captures single motif of Alx4 and MafG. Channel 280 captures 3 consecutive motifs (GATA1,Myod1 and GATA2), while channel 77 captures consecutive NFYB/YA motif and its reverse compliment. Channel 179 captures either REST or its reverse compliment depending on the input sequences used for initialization