Study On Computational Modeling Of Histone Modification

Histones are components of eukaryotic nucleosomes and are the main protein components of cellular chromatin.As a biological process that occurs on histones,its modification is one of the key factors influencing epigenetic inheritance.The detection and identification of modification-related information provides an importance prerequisite for understanding the gene regulation mechanisms of histone modifications.In this paper,a computational modelling study of histone modifications had been carried out,with the following main findings:A deep neural network model based on multi-label classification and multi-task learning was developed to predict the type of histone modifications.iHMnBS is able to predict histone modifications bound to any part of DNA.Evaluation results on a classical dataset show that iHMnBS has excellent performance advantages for the problem of predicting seven histone modification types,as well as providing accurate binding sites for modified histones to DNA for biological experiments.FusionSite was developed to predict the site of action of histone crotonylation modifications based on a deep learning model of multimodal feature fusion.FusionSite explores in depth the local environment of histone lysine using multimodal feature representations,including protein structure information learned by graph neural networks,and sequence features extracted from protein pre-trained language models,specifically for one modification.Evaluation results on a range of modification site prediction datasets show that FusionSite not only effectively fuses multiple representations of proteins,but also captures information specific to the representation of the local environment of the crotonylation site.Based on a deep graph neural network model,UniMut was developed to predict the pathogenicity of protein missense mutations and to explore the relationship between missense mutations and pathogenicity under the constraints of histone modifications.UniMut integrates the sequence features of genes and proteins before and after mutations and further learns the structural information of proteins through graph networks.Competitive performance has been achieved on several independent test sets and modified histone datasets.Deep learning methods build efficient,robust and interpretable predictive models that provide unique insights into the mechanisms of action behind epigenetics from a computational perspective and extend their important applications in cross-disciplinary fields.