【作者】 陈伟；

【导师】 李维华；

【作者基本信息】 云南大学 ， 计算机系统结构， 2022， 硕士

【摘要】 核小体定位指DNA双螺旋结构相对于组蛋白的位置,并对细胞的生命活动起着重要的调节作用。预测核小体定位不仅有助于理解生物体的生命过程,而且有助于预防和治疗多种疾病。通过生物试验的方式测得核小体定位会消耗大量的时间和资源,伴随高通量技术的发展,DNA序列数据逐渐丰富。因此,利用DNA序列进行核小体定位预测成为了一个重要的研究方向。由于单一维度表征DNA限制了深度学习模型在提取DNA特征和预测上的优势,本文主要从以下两个方面进行改进:(1)基于经典的序列特征表示方法,设计了几种核小体DNA序列的表征方法。这些方法分别从不同的维度表征DNA序列,在保留原始DNA序列信息的同时突出了DNA序列中K核苷酸组的权重特性,考虑了K核苷酸组之间的上下文关系和全局联系;在此基础上,结合卷积神经网络、双向循环神经网络、门控机制和自注意力机制设计了四种模态的核小体定位预测模型;此外,通过多组对照实验全面分析了不同长度K核苷酸组、不同数量和类型的神经网络单元对核小体定位预测模型的影响。(2)利用权重分配和Stacking集成方法将核小体定位预测模型按照最优方式进行整合,得到端到端的多模态模型,提高自动提取DNA序列数据局部特征和远程特征的能力,也改进了动态分配特征权重和排除特征噪音的能力。在四个数据集上进行交叉验证,从预测准确度和精度等方面进行全面的评估和分析。实验结果表明,集成模型在核小体定位预测上显著优于独立模型及传统机器学习方法;与权威核小体定位预测模型相比,预测准确度、精度等指标有所提升。这表明整合DNA数据的多表征以及整合多种模型对核小体定位预测是合理和有效的。更多还原

【Abstract】 Nucleosome localization refers to the position of DNA double helix structure relative to histone,and plays an important role in regulating the life activities of cells.Predicting nucleosome localization is not only helpful to understand the life process of organisms,but also to prevent and treat a variety of diseases.Measuring nucleosome localization through biological experiments will consume a lot of time and resources.With the development of high-throughput technology,DNA sequence data is gradually enriched.Therefore,using DNA sequences to predict nucleosome localization has become an important research direction.Because the single dimension representation of DNA limits the advantages of deep learning model in extracting DNA features and prediction,this thesis mainly improves from the following two aspects:(1)Based on the classical sequence feature representation methods,several characterization methods of nucleosome DNA sequences were designed.These methods characterize DNA sequences from different dimensions,highlight the weight characteristics of K-nucleotide groups in DNA sequences while retaining the original DNA sequence information,and consider the context and global relationship between Knucleotide groups;On this basis,combined with convolution neural network,bidirectional cyclic neural network,gating mechanism and self attention mechanism,four modes of nucleosome localization prediction models are designed;In addition,the effects of different length K-nucleotide groups,different number and types of neural network units on nucleosome localization prediction model were comprehensively analyzed through multiple groups of control experiments.(2)The nucleosome localization prediction model is integrated in the optimal way by using the weight distribution and stacking integration method to obtain an end-to-end multimodal model,which improves the ability to automatically extract the local and remote features of DNA sequence data,and also improves the ability to dynamically allocate feature weights and eliminate feature noise.Cross validation is carried out on four data sets to comprehensively evaluate and analyze the prediction accuracy.The experimental results show that the integrated model is significantly better than the independent model and traditional machine learning methods in nucleosome location prediction;Compared with the authoritative nucleosome positioning prediction model,the prediction accuracy and accuracy are improved.This shows that integrating multiple representations of DNA data and integrating multiple models is reasonable and effective for nucleosome localization prediction.更多还原