【作者】 孙超；

【导师】 孟军；

【作者基本信息】 大连理工大学 ， 计算机应用技术， 2014， 硕士

【摘要】 MicroRNA (miRNA)是一种非编码的RNA序列,这些长度约为21个碱基的序列在动物、植物、病毒基因的后转录过程中发挥着巨大的作用。MiRNA可以靶向到它们对应的靶基因,抑制某些基因的转录和翻译,进而阻止相应性状的表达。大量科学实验表明,miRNA参与到很多的生物反应中,这些生物反应可以由生化作用产生也可以由非生化作用产生。为了更好地了解生物的基因调控过程和生长机理,必须对这些miRNA和它们对应的靶基因有更深入的了解。近些年来,以计算机强大的存储和计算能力为基础,通过机器学习方法建立模型的方式被广泛的应用到了生物miRNA分类预测、miRNA靶基因研究和miRNA相互作用分析等研究中。现有的miRNA分类模型单纯的把哺乳动物做为研究对象,或者仅仅可以识别植物miRNA前体而无法给出miRNA的位置。因此,本文在支持向量机(Support Vector Machine, S VM)算法的基础上,构建了一个可以分类植物miRNA前体,并给出前体中成熟miRNA所在位置的集成分类模型miPlantPreMat。从第一个miRNA被发现开始,科学界提出了许多miRNA相关的特征,通过总结以往的这些特征和对植物miRNA序列的分析,本文共提取了152个miRNA序列结构相关特征。从每个特征的信息增益值和在SVM分类过程中的贡献率综合分析出发,提出了一种有放回的支持向量机递归特征消除(Support Vector Machine-Recursive Feature Elimination,SVM-RFE)特征子集选择方法,同时结合了合成少数样本过采样(Synthetic Minority Over-sampling Technique, SMOTE)数据预处理方法和网格优化参数方法,分类效果有所提升。在拟南芥,大豆,番茄等9种植物的数据集上的测试结果表明,该方法对植物的miRNA前体和成熟体分类有着高分类精度和普适性特点。通过miPlantPreMat分类模型,从番茄基因组中预测到了522条潜在的miRNA序列,进一步通过psRNATarget预测得到了3214个miRNA与番茄靶基因相互作用关系,并通过这些关系构建了番茄miRNA通过靶基因的协同作用网络,为生物实验提供了参考。总之,本文成功地利用基于SVM和植物miRNA序列结构特征的分类模型及相关算法解决了miRNA前体和成熟体分类问题。通过对具有代表性的植物和模拟数据的实验分析,无论是在特征的质量还是模型分类能力方面,效果都得到了一定的提高。更多还原

【Abstract】 MicroRNAs (miRNAs) are a family of-21nt non-coding RNAs that play important roles at the post-transcriptional level in animals, plants and viruses. These molecules can identify target genes in order to prevent them from transcription or translation. A large number of researches have shown that miRNAs are involved in the biological response to a variety of biological and non-biological effects. Identification of these molecules and their targets can aid understanding of regulatory processes.Recently, calculating prediction method based on machine learning has been widely used for predicting biological miRNAs, researching miRNA target genes, analysising miRNA-miRNA interaction and many other purposes. However, most of these methods are only for mammals or can not predict miRNAs within pre-miRNAs. Therefore, miPlantPreMat, an integrated classification model based on support vector machine (SVM) was trained specifically for distinguishing real/pseudo plant pre-miRNAs together with their miRNAs.Since the first miRNA had been found, a lot of features related to miRNA were put forward in scientific community. By summing these features and a conventional plant miRNA sequence analysis, an initial set of152novel features related to sequential structure were picked. From a comprehensive analysis of the information gain value of each property and the contribution rate in the SVM classification process, we proposed an improved SVM-RFE method, attribute selection made by this method together with SMOTE method and grid search parameters, the classification results improved. Test results on arabidopsis, soybean, tomato plants and other6kinds of data sets show that the method of miRNA precursors and mature plants classification has high classification accuracy and universal characteristics. A total of522potential tomato miRNA were predicted from the tomato genome by miPlantPreMat.3214miRNA target genes interacting were derived by psRNATarget. Finally, a miRNA target gene interaction networks were constructed through these relationships which provides a reference for experimental biology.In summary, this paper successfully using SVM-based and structural features of plant classification models and algorithms to solve the classification problems of pre-miRNAs and mature miRNAs. Experimental analysis on plants and simulated data show that, either from feature quality or classifiers capacity, the effect has been improved.更多还原