基因表达谱数据的特征选择方法应用策略研究

发布时间：2018-04-15 19:06

本文选题：基因表达谱数据 + 特征基因选择　；参考：《第三军医大学》2016年硕士论文

【摘要】：研究背景:随着分子生物学技术和基因微阵列的不断发展,通过定量测量我们可以很容易获取大量的基因表达谱数据,而且得到的数据质量也越来越高。基因表达谱数据具有维数高,样本小,且存在大量噪声的特点。有研究学者针对基因表达谱数据的特点,利用统计学习和模式识别的方法来对数据进行有效的挖掘。目前最为常用的方法是对表达谱数据进行降维,选择出具有代表性的信息基因,在保证具有较高分类精度的前提下同时提高学习算法的性能和效率。特征选择通过剔除不相关和冗余的特征,而达到降低特征个数,提高模型精确性,减少运行时间的目的。目前虽已有了大量的特征选择算法,但如何针对基因表达谱数据结构自身特点来选择最优的算法还存在着一定困难。因此本研究结合生物学和模式识别的相关知识,通过对一些常用的特征选择算法进行比较,最终提出一种可供参考的选择标准。研究方法:本研究中通过模拟不同特征数、样本量、分类情况以及噪声大小的数据集,选用了8个特征选择算法,并在3种分类器中进行测试,以分类精确度和计算复杂度作为衡量指标来对每一种方法的优劣性进行评价。最后将其应用到真实数据集中,并对每种方法的结果进行分析和比较,选择出最佳的特征选择方法。研究结果:对于不同特点的表达谱数据集分别使用三种类别的特征选择方法来分析都能有效的降低特征维度,通过比较分析后我们发现:SVM-RFE算法在特征数较少和样本量较小的时候有很好的分类结果;Wrapper SVM算法在较少样本量和较多PCR特征基因情况下的分类效果较好;Wrapper k-NN算法对具有较少特征数和较多的FCR特征基因数据集有较好的实用性;Relief F算法能很快的从高维数据集中获得特征子集,该算法在特征数较多和样本量较大的情况下效果好;mRMR算法同样适合特征数较多的情况,并在信噪比较高的情况也有也有很好的效果。研究结论:对肿瘤的类别进行划分并从中挖掘有生物学含义的知识时,信息基因的筛选是非常重要的步骤。对分类无关基因的排除可以进一步降低数据量或减少数据的维数,从而提高分类器的性能。本文从已有的大量特征选择算法中,选取了一些常用的方法并针对表达谱数据的不同特点来进行对比研究,最终建立方法应用策略,为基因表达谱数据的分析提供了方法学上的指导。
[Abstract]:Background: with the development of molecular biology technology and gene microarray, it is easy to obtain a large amount of gene expression profile data by quantitative measurement, and the quality of the obtained data is getting higher and higher.The data of gene expression profile have the characteristics of high dimension, small sample and large amount of noise.According to the characteristics of gene expression profile data, some researchers use the methods of statistical learning and pattern recognition to mine the data effectively.At present, the most commonly used method is to reduce the dimension of the expression spectrum data, select the representative information gene, and improve the performance and efficiency of the learning algorithm under the premise of high classification accuracy.By eliminating irrelevant and redundant features, feature selection can reduce the number of features, improve the accuracy of the model and reduce the running time.Although there are a lot of feature selection algorithms, it is still difficult to select the optimal algorithm according to the characteristics of gene expression profile data structure.So this study combined with the knowledge of biology and pattern recognition, through the comparison of some commonly used feature selection algorithms, finally put forward a reference selection criteria.Research methods: in this study, eight feature selection algorithms were selected and tested in three classifiers by simulating data sets of different feature numbers, sample sizes, classification conditions and noise size.The classification accuracy and computational complexity are used to evaluate the advantages and disadvantages of each method.Finally, it is applied to the real data set, and the results of each method are analyzed and compared, and the best feature selection method is selected.The results showed that for the expression data sets with different characteristics, using three kinds of feature selection methods can effectively reduce the feature dimension.Through comparison and analysis, we find that the small number of features and the small sample size of the small number of features of the small number of features and small sample size of the small number of classification results of the wrapper SVM algorithm in the case of less sample size and more PCR feature genes, the classification effect is better than that of the wrapper k-NN algorithm.With fewer feature numbers and more FCR feature gene datasets, Relief F algorithm can quickly obtain feature subsets from high dimensional datasets.The algorithm has good performance in the case of more features and larger sample size. It is also suitable for the case of more features, and also has a good effect in the case of higher signal-to-noise ratio (SNR).Conclusion: the screening of information genes is a very important step in the classification of tumors and the mining of biological knowledge.The elimination of classification independent genes can further reduce the amount of data or reduce the dimension of data, thus improving the performance of the classifier.In this paper, we select some common methods from the existing feature selection algorithms, and compare and study the different characteristics of the expression spectrum data. Finally, we establish the application strategy of the method.It provides methodological guidance for the analysis of gene expression profile data.
【学位授予单位】：第三军医大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：Q811.4

【相似文献】