Fig. 1

Epiphany employs long short-term memory and adversarial loss to predict the Hi-C contact map. A Architecture of Epiphany. Epigenomic signal tracks are first presented to the model in a sliding window fashion, with window size of 1.4 Mb and step size of 10 kb. During training, we take a total length of 3.4 Mb of the input (200 windows) in one pass. In the generator, the processed input data is first featurized by convolution modules, followed by a Bi-LSTM layer to capture the dependencies between nearby bins. After a fully connected layer, the predicted contact map is generated. An MSE loss between the predicted map and the ground truth is calculated in order to train the generator to predict correct structures. To mitigate the overly-smoothed predictions by the pixel-wise losses, we further introduced a discriminator and adversarial loss. The discriminator consists of several convolution modules, and an adversarial loss was calculated to enable the model to generate highly realistic samples. We trained Epiphany with a combined loss of these two components. B An illustration of prediction scheme. The first window of input data (blue horizontal line, 1.4 Mb) is used to predict a vector on the Hi-C contact map that is orthogonal to the diagonal (blue bin vector, covers 1 Mb from the diagonal). Note that an extra .2 Mb of input is added to either side of each input window (a total length of 1.4 Mb instead of 1 Mb) in order to provide the model with additional context. During training, 3.4-Mb input tracks are processed using sliding windows (200 windows) in one pass, and 200 consecutive vectors are being predicted. C An example region of input epigenomic tracks (bottom), target Hi-C map (top row), and predicted Hi-C map (second row)