Reveal the Cognitive Process of Deep Learning during Identifying Nucleosome Occupancy and Histone Modification

Nucleosome occupancy and histone modifications are among the most significant factors that influence gene expression. Although Chromatin Immunoprecipitation-chip (ChIP-chip) and Chromatin Immunoprecipitation-sequencing (ChIP-seq) have significantly accelerated revealing the nucleosome regulation mechanism in the course of organism life, these experimental techniques require to take plenty of material resources and may produce noise. To this end, Computer-based information processing algorithms, such as Support Vector Machines (SVMs), have given different insights into the above deficiencies. Recent researches have demonstrated that deep learning can not only conquer the above unfavorable factors in many kinds of gene regulation tasks, but also exceed the prediction performance of SVM. However, there is no explicit understanding of why the so-called “Black box” performs so well. Here, we constructed deep learning model to identify the nucleosome occupancy and histone modifications, and evaluated the performance. Then from different perspectives (alignment-based and optimization-based approaches respectively), we reported the inside mechanism of deep learning model for revealing the distribution of nucleosome occupancy and histone modification states in Yeast. Eventually, our approaches not only achieved ascertain accuracy improvement on the published datasets compared with traditional machine learning, but also we found the sequence preference of nucleosome occupancy that is consistent with previous findings.