Program Model for Prediction of Loops for Genomic Sequence in Chromatin of Dictyostelium Discoideum

This work is dedicated to the development of the model for prediction of the DNA loops from the genomic sequence and its properties, such as nucleotide content, k-mers, the presence of transcription factors’ motifs and the presence of G-quadruplexes. The paper reviewed the problem of binary and multiclass classification. For binary classification, we considered the Random Forest model and Convolutional Neural Network, and achieved ROC-AUC of 0.67 and 0.7, respectively. During the research, we observed loops with an increased number of contacts (extrusion track) and developed an algorithm for its detection. Therefore, the problem of binary classification has been modified to multiclass: the classification of the genomic region to one of the four classes: out-loop, loop, inter-loop or loop with extrusion track region. The Random Forest model was considered for multiclass classification. Accuracy of 0.69 was achieved for this model. We interpreted models by obtaining feature importances from Random Forest and compared convolutional filters of the 1st layer of Convolutional Neural Network with transcription factors motifs.