基于序列的RNA甲基化修饰位点预测研究

发布时间：2018-07-27 15:35

【摘要】：RNA的转录后修饰在生命体中普遍存在,并且在许多生物过程中扮演了重要的角色,而RNA的甲基化修饰则是RNA转录后修饰中一个重要的分支。RNA甲基化是指在RNA的某些核苷酸分子上发生甲基化修饰的现象,常见的包括N6-甲基腺苷、N1-甲基腺苷等。近年来的研究表明,RNA甲基化修饰可以影响RNA的转录、代谢、剪接和稳定性;能与相关的蛋白质结合,从而调控基因的表达。并且,RNA甲基化修饰与肿瘤、肥胖症等疾病也有所关联。因此,从RNA序列中精确地识别出RNA甲基化修饰位点是一项重要的任务。传统的基于物理化学实验来识别RNA甲基化位点的方法成本高、耗时长且规模较小。近年来出现的高通量测序方法能进行高效的、大规模的RNA甲基化位点识别,但其本质仍然是基于生物化学实验的方法。因此,设计一种基于机器学习理论的RNA甲基化修饰位点预测方法,是十分必要的。本文对基于序列的RNA甲基化修饰位点预测问题进行了深入研究,主要工作如下:(1)研究了 RNA的基本性质,并提出了一种新的RNA序列的特征提取方法。位置特异性偏好思想在蛋白质修饰位点预测问题中得到了成功的应用。受此启发,本文将位置特异性偏好思想应用于RNA序列,提出了位置特异性核苷酸/二核苷酸偏好特征,用于RNA序列的特征提取。该思想利用统计方法,分别计算每种核苷酸在正负样本集合的序列中每个位置的出现频率,并且利用正负样本集合之间的差异进行特征编码。实验结果表明,该特征提取方法能进一步地提升N6-甲基腺苷位点预测的精度。(2)对基于序列的N6-甲基腺苷位点预测问题进行了研究,并提出了名为TargetM6A的预测方法。TargetM6A方法利用所提出的位置特异性核苷酸/二核苷酸偏好特征,并结合传统的核苷酸组成成分特征,来对RNA序列进行特征提取;对提取出的特征使用增量特征选择方法进行特征筛选,选出更有判别力的特征成分子集;最终使用支持向量机算法训练预测模型。实验结果表明,相对于现有的基于序列的N6-甲基腺苷位点预测方法,本文提出的TargetM6A方法在基准数据集上取得了更好的预测结果。(3)对基于序列的N1-甲基腺苷位点预测问题进行了研究,并提出了名为TargetM1A的预测方法。针对最近新发布的N1-甲基腺苷实验数据,进行了数据的处理和采样,构建了 3个基于物种和6个基于组织细胞的N1-甲基腺苷位点的数据集。TargetM1A方法提取了数种基于RNA序列的特征,并使用极限随机树算法作为分类器。TargetM1A方法在基于物种和基于组织细胞的预测模型的交叉验证实验中都取得了不错的性能;它对于现有的基于实验来研究N1-甲基腺苷位点的方法来说,是一个有益的补充工具。(4)对于所提出的TargetM6A和TargetM1A方法都提供了在线的预测服务,供其他研究人员免费地使用。
[Abstract]:Posttranscriptional modification of RNA is widespread in organisms and plays an important role in many biological processes. RNA methylation is an important branch of RNA posttranscriptional modification. RNA methylation refers to the methylation of RNA on some nucleotide molecules, including N6-methyladenosine N1-methyladenosine. Recent studies have shown that RNA methylation can affect the transcription, metabolism, splicing and stability of RNA, and can bind to related proteins to regulate gene expression. And RNA methylation is also associated with diseases such as cancer, obesity, and so on. Therefore, accurate identification of RNA methylation sites from RNA sequences is an important task. Traditional methods based on physicochemical experiments to identify RNA methylation sites are costly, time-consuming and small. In recent years, high-throughput sequencing methods can be used to identify RNA methylation sites on a large scale, but their essence is still based on biochemical experiments. Therefore, it is necessary to design a RNA methylation modification site prediction method based on machine learning theory. In this paper, the prediction of RNA methylation modification sites based on sequences is studied. The main work is as follows: (1) the basic properties of RNA are studied, and a new feature extraction method for RNA sequences is proposed. The idea of position specific preference has been successfully applied to the prediction of protein modified sites. Inspired by this, the idea of location-specific preference is applied to RNA sequences, and a location-specific nucleotide / dinucleotide preference feature is proposed for feature extraction of RNA sequences. Using statistical method, this idea calculates the occurrence frequency of each nucleotide in each position in the sequence of positive and negative sample sets, and encodes the features by using the difference between positive and negative samples. The experimental results show that this feature extraction method can further improve the accuracy of N6-methyladenosine site prediction. (2) Sequence-based prediction of N6-methyladenosine sites is studied. A prediction method called TargetM6A. TargetM6A is proposed to extract the RNA sequence by using the location-specific nucleotide / dinucleotide preference feature and combining with the traditional nucleotide component characteristics. Incremental feature selection method is used to screen the extracted features to select the more discriminant subset of feature components. Finally support vector machine algorithm is used to train the prediction model. The experimental results show that compared with the existing sequence-based prediction methods for N6-methyladenosine sites, The TargetM6A method proposed in this paper obtains better prediction results on the datum data set. (3) the prediction problem of N1-methyladenosine sites based on sequence is studied, and a prediction method called TargetM1A is proposed. According to the newly released experimental data of N1-methyladenosine, the data were processed and sampled. Three data sets based on N1-methyladenosine locus and six histiocyte-based data sets were constructed. TargetM1A method was used to extract several RNA sequence-based features. Using the LRT algorithm as the classifier. TargetM1A method has achieved good performance in the cross-validation experiments based on the prediction model based on species and tissue cells. It is a useful supplementary tool for the existing experiment-based methods to study N1-methyladenosine sites. (4) both the proposed TargetM6A and TargetM1A methods provide online predictive services for free use by other researchers.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：Q811.4

【相似文献】