Prediction of Functionality of DNA Secondary Structures with Deep Learning Methods

G-quadruplex is a DNA secondary structure which can be formed by sequences rich in guanine. There are more than 700 000 sites that can potentially form G-quadruplex structure in the human genome. G-quadruplexes have been located in promoter regions of many genes. They took part in transcriptional regulation. However, there is no method to predict quadruplex formation in different types of cells. In this master thesis deep convolutional neural network has been trained to predict histone modifications in different types of tissues (median AUC ROC 0.83) based only on information about DNA sequence. The analysis of first convolutional layer filters identified motifs, some of them appear to be known transcriptional factor binding sites. Method to predict quadruplex functional significance in different tissues was suggested, based on evaluation of neural network prediction difference for original and mutated sequences near potential quadruplex sites. The efficiency of this method was demonstrated on the known G-to-A mutation in the promoter region of c-MYC. Suggested method was applied to quadruplex dataset for myelogenous leukemia cell line K562. It was shown that significant differences on mutation maps are associated with sites that form quadruplex structures in this cell line. This thesis demonstrated the effectiveness of deep learning methods to recognize relationships between patterns of DNA secondary structures and epigenetic code.