Ensemble of Elman neural networks and support vector machines for reverse engineering of gene regulatory networks

In this study, recurrent Elman neural networks (ENNs) and support vector machines (SVMs) have been used for temporal modeling of microarray continuous time series data. An ensemble of ENN and SVM models is proposed to further improve the prediction accuracy of the individual models. The prediction results on the simulated non-stationary datasets and the real biological datasets outperform the results of the other existing approaches. In order to provide the neural networks with explanation capabilities, a pedagogical rule extraction technique has been proposed to infer the output of our proposed ensemble system. The proposed pedagogical rule extraction technique is a two-step test of causality and Pearson correlation for the network inference between the causal gene expression inputs and their predicted outputs. The results of the network inference demonstrate that the gene regulatory network can be reconstructed satisfactorily with the proposed approach.