Modeling Research Of Nucleosome Positioning And Occupancy Based On Deep Learning

The precise positioning of nucleosomes can regulate the accessibility of DNA-binding proteins to the genome,thereby regulating gene expression,DNA replication,and repair.Currently,the main research methods for nucleosome positioning are biological experimental methods and machine learning-based computational methods.Since biological experiments require significant time and experimental costs,it is a more practical research need to develop efficient and accurate nucleosome positioning algorithms.The nucleosome positioning problem can be divided into two parts:nucleosome identification and nucleosome density prediction for each base position.In addition,because nucleosome formation depends on the highly curved DNA sequence,this article also studies the rules and mechanisms of nucleosome positioning from the perspective of DNA curvature.The specific research contents are as follows:This paper proposes DeepNup,a nucleosome positioning method using multi-scale convolutional network and gated recurrent network.It encodes DNA sequence with a hybrid scheme of One-hot and trinucleotide composition.DeepNup extracts and fuses local and abstract features from sequence matrix with multi-scale convolutional network,and learns long-range interactions between bases with gated recurrent network.The performance of DeepNup method is evaluated on two classic nucleosome positioning benchmark datasets and compared with existing advanced methods.The results show that DeepNup method can effectively improve the prediction performance of nucleosome positioning.In addition,this paper constructs a dataset based on the latest human genome,further verifies the robustness of DeepNup method,and analyzes the base preference of nucleosome sequences in human genome on this basis.This paper proposes a method for predicting nucleosome density,called DeepNDP,based on ResNet and Transformer.The method uses nucleotide chemical properties as features to better capture the chemical information and structural features in DNA sequences.The ResNet in DeepNDP can extract local information from DNA sequences,while the Transformer encoder can integrate local information and capture long-range interactions between bases.The performance of the DeepNDP method was evaluated on the yeast genome,and the results showed that it outperforms existing advanced methods.Furthermore,the article analyzes the model’s performance on the mouse genome data and demonstrates that the DeepNDP method has good cross-species generalization ability.In this paper,a DNA bending degree prediction method DBBP based on a dense convolutional network and a bidirectional gated loop network is proposed,and on this basis,the association between DNA bending degree and nucleosome density is verified.The method uses nucleotide chemical properties and nucleotide electron-ion interaction pseudopotentials as features.The dense convolutional neural network in DBBP can enhance the mobility of DNA sequence information between convolutional layers,and the bidirectional gated recurrent network can extract global information in the sequence.This paper evaluates the performance of the DBBP method on four data sets,and achieved better results than the current advanced methods.On this basis,the DNA bending degree was compared and analyzed with the predicted nucleosome density,and the correlation between DNA bending degree and nucleosome density was verified.This article proposes a method based on deep learning technology,incorporating physical sequence features to construct a nucleosome occupancy model that predicts nucleosome positions more accurately from different perspectives.This provides technical support for rapid exploration of the significance of nucleosome positioning in biological processes.