Fig. 6

Figure 6a shows the DNN architecture. It takes 24 histone modifications (each has 20 features) as input and predicts whether a genomic location is a distal regulatory site or not. There are three hidden layers and one output layer. Between each hidden layer, we used PReLU as activation function and dropout with rate 0.3 between each hidden layer, to prevent overfitting. Figure 6b gives an illustrative example of row-wise stacking of histone modifications used as inputs to our CNN model. As shown in Figure 6b, each location has various histone modification signals, represented by zigzag lines with di.erent colors in the figure. For illustration purposes, we only represent four histone modification signals. By stacking these signals row-wise, these signals are captured as snapshots of informative features of the genome at each location. Similar to standard RGB images where channels provide di.erent color features, each type of histone modification signal provides unique information to the model. Since the patterns of those signals are quite di.erent across di.erent types of histone modifications, removing any subset of them could result in information loss. With the proper design of the convolution kernel, where the height can cover all signals, the convolution kernel can extract local features to the next layer of the designed CNN. The width of the kernel should not be too large. Too wide a kernel would result in the kernel convolving remote features that are irrelevant to characterizing the local information. Figure 6c shows the CNN architecture. The input is in 2D form with each row representing one histone modification feature. After each convolutional layer, it has PReLu layer (due to the space constraint, we skipped showing them in the Figure). After Max-Pooling for down sampling, CNN connects two layers of fully connected neurons, each layer has 300 neurons, and finally connects with output. To prevent overfitting, we also add dropout with rate 0.5 between Max-Pooling and first fully connected layer and between first and second fully connected layer, and dropout with rate 0.3 between the second fully connected layer and output layer