Prediction of chromatin spatial structure characteristics using machine learning methods

Development of chromosome conformation capture methods boosted progress in the study of the spatial organization of chromatin. Accumulation of large amounts of experimental data provides an opportunity to apply machine learning methods to examine the connection between epigenetics and the three-dimensional structure of chromatin. The aim of this study was to predict the characteristics of the chromatin structure, namely the transitional gamma, from ChIP-Seq experimental data by means of machine learning methods, and also to reveal the properties of epigenetic data influencing prediction. The neural network and the loss function designed for the prediction task are shown to perform with a sufficiently high accuracy. In addition, the genomic size of the chromatin context required for improving the quality of the prediction was assessed. Several neural network visualization techniques were tested as a means for improving interpretability of network, showing the possibility for using visualization to study interrelations in epigenetic data relevant for three-dimensional chromatin structure. To sum up, a close relationship between epigenetic factors and the structure of chromatin has been confirmed.