Quadruplex Recognition System by Deep Learning Methods

At the present time the field of DNA secondary structures is actively studied. This work presents a study of deep learning methods for predicting the formation of one of these structures, G-quadruplexes. In this study, we developed models capable of predicting the formation of G-quadruplexes in vivo using both the nucleotide sequence alone and together with omics data. We used LSTM model that was trained on two of available data sets – ChIP-seq and G4-seq. Next we processed the omics data from the ChIP-Atlas, and the resulting features were used to train the model. The resulting model had a higher quality than the model using only nucleotide data. Then we performed interpretation of the model and identified features important for predicting G-quadruplex formation from omix data. With the trained model we generated whole-genome annotation maps of G-quadruplexes for human and mouse genomes. Finally we developed a method for transferring the resulting model to other organisms using the domain-adversarial neural network (DANN) method from the domain adaptation class.