Genome Functional Annotation using Deep Convolutional Neural Networks

Deep neural network application is today a skyrocketing field in almost all disciplinary domains. In genomics, which deals with DNA sequences, the development of deep neural networks is expected to revolutionize current practice, from fundamental issues such as understanding the evolution of genomes to more specific applications such as the development of personalized medicine. Several approaches have been developed relying on convolution neural networks (CNN) to identify the functional role of sequences such as promoters, enhancers or protein binding sites along genomes. These approaches rely on the generation of sequences batches with known annotations for learning purpose. While they show good performance to predict annotations from a test subset of these batches, they usually work less well when applied genome-wide (i.e., for whole genome annotation). In this paper, we address this issue and propose an optimal strategy to train CNN for this specific application. We use as a case study gene Transcription Start Sites (TSS) and show that a model trained on one organism (e.g., human) can be used to predict TSS in a different species (e.g., mouse). Author summary We propose a method to use deep convolution neural networks in order to label genomes with functional annotations. Functional annotations cover any relevant features which can be associated with specific positions on the genome (e.g., promoters, enhancers, conserved regions). This method is based on a optimized generation of the examples used to train the network in order to deal with the well-known problem of using unbalanced data. When these annotations are known in one species, the trained neural network can be used to predict these annotations in a different species if the mechanisms used to interpret the genomes are conserved in the two species. We use as a case study gene transcription start sites (TSS) in human and show that the model trained on human TSS can be used to recover a similar information on the mouse genome.