Figure 4.

A standard N-to-1 Neural Network, as used in the sequence-based experiments in this article and in (Mooney et al. 2011 ). All motifs from the sequence are input to replicas of input-to-feature neural networks (grey boxes at the bottom of the image), the feature vectors output are added up, to produce a global feature vector (f). f is then mapped to the property of interest through a feature-to-output neural network (light grey box at the top of the image).