Recurrent Neural Network for Gene Regulation Network Construction on Time Series Expression Data

We propose a new way of exploring potential transcription factor targets in which the Recurrent Neural Network (RNN) is used to model time series gene expression data. Once the training of the RNN is completed, inference is performed through feeding the RNN artificially constructed signals. These artificial signals emulate the original gene expression data and the transcriptional factor of interest is set to be zero constantly to model the knockout state of the transcription factor. The predicted expression patterns of the other genes from the RNN are then used to measure the likelihood that the gene is regulated by the knocked out transcriptional factor. After repeating the same process for each gene as Transcription Factor in the dataset, we construct a gene regulation network with edge weights assigned. We demonstrate the effectiveness of our model by comparing our method with existing popular approaches. The result shows that our RNN method can identify transcription factor targets with higher accuracies than most of existing approaches. Overall, our RNN model trained on time series gene expression data can be useful for discovering transcription factor targets as well as building a gene regulation network.