Fig. 5

Impact of different types of hidden layers on the receptive field. The receptive field reflects the number of adjacent residues considered when making a prediction for a specific residue. Using a window size of three, the classic neural network architecture consisting of convolutional layers achieves a receptive field of size seven after three hidden layers. Using an increasing dilation rate for each layer, the receptive field can be expanded to 15 residues using the same window size and number of layers. For a bidirectional LSTM, the information from the complete sequence is available at each position and each layer