Generative Adversarial Network for DNA Secondary Structures Prediction

In addition to the primary structure, DNA molecule can form various secondary structures, such as G-quadruplexes, Z-DNA and others. One of the main problems in annotating genomes with DNA secondary structures is the lack of genome-wide experimental data. On the other hand, the existing computer methods for predicting the formation of DNA secondary structures are not accurate enough, and their improvement is an actual task. Recently, generative-adversarial neural networks (GAN) are increasingly used for research purposes, including in solving genomics problems. Their goal is to generate new data from distribution of real data. Therefore, such neural networks make it possible to expand experimentally obtained data with artificially generated data, which can be used to increase the accuracy of models predicting locations of secondary structures. This thesis demonstrates the possibility of using GAN to generate new data on DNA secondary structures and increase the accuracy of prediction of machinelearning models using these generated data.