Extracting Activity Patterns from Altering Biological Networks: A Sparse Autoencoder Approach

In this paper we address the problem of extracting activity patterns from biological networks that cannot be characterized in the form of a static graph, such as gene regulatory networks wherein the co-expression patterns of different genes may vary over time. A simple autoencoder architecture is applied in order to detect activity patterns appearing in several observational samples of the data. We apply Kullback-Leibler divergence based sparsity constraint on the hidden layer neurons, such a way the activity of a hidden layer unit can be associated to a particular activity pattern which can be extracted using the trained decoder function.