Fig. 3.

The visualization of our low-level Motifs Activation Map network. Take the example of one 9-mer peptide, converting to the feature matrix with the shape of 9 × 16 (16 is the kernel size of the first 1-D convolutional layer) out from the first convolutional layer. The representation of the site, 9, is preserved. Then using the W1 matrix to add each feature from the low-level weight and collect together. Then the feature matrix of size 9 × 1. As the low-level feature with respect to sites, to get the final site rank, we give a weight for low level and then merge all levels feature matrices together. the final result’s shape is still 9*1, we preserve the length through the calculating of sites contribution vector and it provides intuitive information for us to compare the contribution to the binding probability of each site