Inference of regulatory genetic networks by neural networks

In this paper we provide a brief introduction to three computational methods involving non-linear neural networks, for inferring genetic regulatory interactions. These methods are applied to model regulatory interactions in rathippocampus-development temporal gene expression data. A testing procedure, alternating time points, is devised to evaluate the performance and generalizability of a trained neural network. A second testing method, referred to as reverse prediction, is also presented to ensure that the modeling methodologies capture regulatory interactions, that is, causal relationships, as opposed to simple correlations in the gene expression profiles. The trained neural networks predict the expression profile of a gene, utilizing a minimal set of input gene profiles, with high accuracy on the test data. Additionally, the results of the reverse prediction procedure indicate that the trained neural networks have not merely learned correlation relationships between the gene expression profiles.