Mining Key Regulators of Cell Reprogramming and Prediction Research Based on Deep Learning Neural Networks

Deciphering the dynamic changes of core factors at different reprogramming stages plays an important role in elucidating the reprogramming mechanism of induced pluripotent stem cells (iPSCs) and improving their induction efficiency. The use of transcription factors (TFs) in combination with histone modification is vital to understand the multiple regulatory of pioneer factor. However, existing studies are not enough to consider the classification of stage-specific gene clusters from the perspective of multi-omic in the process of cell reprogramming. In this study, three stage-specific gene clusters of reprogramming initiation, maturation and stabilization phase were identified by using differential expression analysis. Considering the effects of regional binding preference, we further constructed a quantitative model on different genome regions (promoter, enhancer and enhancer subdivision region) by integrating the DNA binding profiles of Oct4 and three histone modifications (HMs). For promoter and enhancer regions, the receiver operating characteristic curve (Roc curve) of support vector machine (SVM) model was above 0.75 and predictive with the accuracy (Acc) about 66~69%. But on enhancer subdivision region, the convolutional neural network (CNN) model we constructed showed more faithful predictive performance than the model on promoter and enhancer, which Roc curve area can reach 0.87. Taken together, our studies quantitatively reveal the cooperative effects of TFs and HMs on reprogramming stage-specific gene clusters, hoping to provide new sights in mining the key regulators of reprogramming.