Prediction of Loops in the Chromatin of Dictyostelium Discoideum

Chromatin is a dynamic structure consisting of DNA, RNA and DNA-binding proteins that compact the polymer of the DNA molecule inside the cell nucleus, as a result of which various structures are observed that affect the important cellular processes. A specialized molecular biology technique called Hi-C (high-throughput chromosomes conformation capture) exists for studying the three-dimensional structure of DNA. In this work, we present the UNBLoC (U-Net-Based Loop Caller) algorithm for searching for the properties of bright points (or “loops”) of chromatin in Hi-C maps of humans, the social amoeba. UNBLoC Models are based on a convolutional neural network U-Net architecture that was trained using manual loop marking for yeast and amoeba and algorithmic for human models and predicted masks with loop areas. After that, the coordinates of the loops predicted by the network were refined by evaluating the position of the center of the mask. Three models based on amoeba, yeast, and human data were obtained. Also, models with primary learning on human maps and additional training on yeast or amoeba data were obtained (transfer learning). All models were compared with basic neural network metrics, however in view of additional features of the task-specific metrics were developed like the number of matched/not matched loop coordinates, the profile of the convergence score measure in the loop coordinates. For the assessment of the quality of the results, classic graphs have been built to assess the quality of the loops found on Hi-C maps. These include average loops for all labeled source data and all the results of all models on all organisms, graphs of the dependence of the probability of contacts on distance (scaling plots). It is noteworthy that with the help of UNBLoC with transfer learning, the property of loop nesting not previously observed for amoebas was discovered.