Fig. 1

Schematic of the forward model convolutional neural network (CNN). This neural network is trained to predict chromatin contacts maps (Hi-C, top) from various chromatin-sequence factors (Chip-seq profiles, bottom). The Hi-C data to be predicted as an output is the upper diagonal of the Hi-C matrix of a w-wide genomic window. The input to the CNN is a 3w-long sequence that includes the w-long region of the Hi-C matrix (inner sequence) as well as two w-long sequences on each side (flanking sequences). The CNN is made of two parts. First, a sigmoid-activated convolutional layer reduces the M chromatin profiles to a 1D sequence profile. Then, the 1D sequence profile is fed to a ReLu-activated dense neural network (DNN) that predicts the Hi-C contact maps