Prediction of Functional Genome Regions Using Deep-learning Methods

The use of large data sets and machine learning methods show outstanding results in bioinformatics problems. Previously developed and published by us, the DeepZ approach showed a high predictive ability in the task of predicting Z-DNA regions for the human genome. In this paper, the DeepZ approach was transferred to the mouse genome. To train the model, full-genome ChIP-seq experiments were used to predict the Z-DNA formation sites in the mouse genome under various cellular conditions. The aggregated results of more than 60,000 genome-wide experiments were used as omix data. Various architectures of convolutional, recurrent, and hybrid models were tested to select the best model. The best model from the RNN family of architectures generated the genome-wide prediction of Z-DNA regions in the mouse genome, and the statistically reliable results on the interpretation of the models were obtained. The results of model interpretation show already known relationships between Z-DNA and epigenetic markers, but also reveal previously unknown relationships. In general, an interpretation method is proposed that can work for a wide class of tasks to determine the relationship of the epigenetic and regulatory code with the functional elements of a genome.