【作者】 王楠；

【导师】 马志强；

【作者基本信息】 东北师范大学 ， 应用技术， 2019， 硕士

【摘要】 RNA修饰是发生在生物体中将初级转录RNA转化为成熟RNA的加工过程,是基因调控领域的重要组成部分。同时该修饰作为转录后水平的调控方式,参与多种重要的生物学过程。迄今为止,已有超过150种RNA的转录后修饰在生命领域被发现。RNA修饰可以影响RNA剪接、RNA降解、蛋白质翻译和人体免疫调节等生物活动,在生物体的调节中起着不可或缺的作用。因此,对于已测得RNA基因修饰数据研究其修饰类型,有助于为科学家解释和发现RNA表观修饰调控提供新思路,更好地理解其分子机制和功能。虽然物理化学实验和基于高通量测序技术的RNA修饰位点识别方法一定程度上促进了RNA修饰研究的进展,并对RNA修饰的鉴定和生物功能的认识起到了积极的推动作用,但是完全通过生物实验以确定RNA修饰位点和类型不仅耗时且成本较高。因此,随着高分辨率实验数据的积累和大量计算模型的出现,使用生物信息学的方法对RNA序列信息进行分析预测和确定RNA修饰位点和类型是非常迫切和必要的,并且很快变成表观遗传领域的研究热点和前沿问题之一。近年来,已经有多个RNA修饰位点预测工具被发布。但是这些预测工具大部分都依靠于研究人员的先验知识,利用RNA修饰的单一或者多个特征构建预测模型,如何有效的选择特征对预测精度的提高带来了极大的挑战。此外一些预测方法只关注单一物种的单一类型或多种类型的RNA修饰位点,并且对小样本RNA修饰类型无法进行预测,导致相关研究人员缺乏真实有效的工具使用,因此这一研究领域仍然值得关注和深入探索,进行创新和改进。目前,深度学习算法在许多领域取得重大性的进展和突破,代表性的模型有长短期记忆网络。与标准的前馈神经网络不同,该方法使用人工循环神经网络架构,处理整个数据序列,挖掘出原始数据的潜在信息,并自行对相关特征进行选择和组合,这极大地避免了人工选择特征的主观性,因此依托深度学习算法改善和提高RNA修饰位点和类型的实验性能具有良好的应用和研究前景。鉴于此,本文利用深度学习框架和RNA序列信息构建RNA修饰位点和类型预测模型,是一种新式的多物种多类别RNA修饰位点预测器。该工具基于双向门控循环单元(BGRU)和迁移学习,充分利用多种RNA修饰位点数据及特征的相互关联,为不同类型的RNA修饰位点构建预测模型。为了有效训练和评估模型性能,本文的方法进行十折交叉验证并使用独立测试集,我们还与多种深度学习模型进行了对比,同时还与之前发布的3种可使用的RNA修饰位点预测工具进行对比。实验结果显示,本文的模型不仅比之前提出的预测工具相比具有更好的性能,而且能较为准确地对多物种多类别的RNA修饰位点进行预测。此外该模型的预测结果和性能也证明了在该研究领域具有发展潜力。更多还原

【Abstract】 RNA modification is a biological process in all living organisms that turns primary transcribed RNA into mature RNA,and is an important part of field of transcriptional regulation.Meanwhile,this modification is involved in many vital biological processes as a regulation mode of post-transcriptional levels.To date,more than 150 post-transcriptional modifications of RNA have been found in the field of life.RNA modification can affect biological activities such as RNA splicing,RNA degradation,protein translation and human immune regulation,playing an indispensable role in the regulation of organisms.Therefore,researching the modification types of RNA gene modification data can provide scientists a new insight to explain and discover the regulation of RNA epigenetic modification and better understand its molecular mechanism and function.Although physical chemistry experiments and high-throughput sequencing-based RNA modification sites identification methods have promoted the progress of RNA modification research to a certain extent,and have stimulated the recognition of RNA modifications and understanding of biological functions,RNA modification site and types determined entirely by biological experiments is consuming.Thus,with the accumulation of high-resolution experimental data and the emergence of a large number of computational models,it is very urgent and necessary to analyze RNA sequence information to predict and determine RNA modification sites and types using bioinformatics approaches.Moreover,it is becoming one of the research hotpots and frontier issues in the field of epigenetics.Several RNA modification site predictors have been developed in recent years.However,most of these prediction tools depend on the knowledge of researchers,using single or multiple features of RNA modification to build prediction models,and how to effectively select features brings great challenges to the improvement of prediction accuracy.In addition,some prediction methods only focus on the single type or multiple types of RNA modification sites in a single species,and cannot predict the types of RNA modification in small samples,resulting in the lack of real and effective tools for related researchers.Consequently,this research field is worthy of attention,exploration,innovation and amelioration.At present,deep-learning algorithms have achieved significant progress and breakthroughs in many fields and representative models including long short-term memory networks.Unlike standard feedforward neural networks,this model is an artificial recurrent neural network architecture to process data sequence and mine the potential information of original data.It can select and combine relevant features by itself,which greatly avoids manually selecting features.Hence,relying on deep learning algorithms to improve and enhance the experimental performance of RNA modification sites and types has a good prospect of application and research.In consideration of above,we use deep-learning architecture and RNA sequence information to construct an RNA modification site and type prediction model,which is a new type of multi-species,multi-class RNA modification site predictor.Based on bidirectional Gated Recurrent Unit(BGRU)and transfer learning,this tool makes full use of data correlation and characteristics of multiple RNA modification sites to build predictive models for different types of RNA modification.In order to train and evaluate model performance,we compared the experimental results with various deep-learning models under 10-fold cross-validation and independent test,and also compared with 3 previously available RNA modification site prediction tools.Experimental results showed that our model not only has better performance than previous predictive tools,but also can accurately predict multi-species and multi-class RNA modification site.Additionally,the prediction results and performance of the model also prove that it has development potential in this research area.更多还原