Prediction of Triplex Structures by Deep Learning Methods

Triplex structures are a special class of secondary DNA structures, in which small pieces of RNA attached to double-stranded DNA according to the principle of complementarity. There is experimental evidence that triplex structures are formed by non-coding regulatory RNAs and affect the expression of nearby genes. Recently whole-genome experiments determining the position of triplex structures have become approachable. The aim of this work was to use machine learning models to predict triplex structures using data from a whole-genome experiment for mapping these structures. This study examined the applicability of deep learning methods for solving this kind of problems. The CNN and RNN architectures were applied. A comparative analysis of the predictive power of deep neural networks and classical machine learning models was also carried out. It was found that convolutional neural networks achieve a higher quality of predicting triplexes in DNA sequences than classical machine learning models. However, the low quality of the prediction may also indicate that information stored in the primary DNA sequence is not enough for an effective model construction, so additional information, for example, from epigenetics and chromatin organization, should be used.