【作者】 张红燕；

【导师】 袁哲明；

【作者基本信息】 湖南农业大学 ， 生物信息学， 2015， 博士

【摘要】 肿瘤是多基因与环境共同作用的结果,大规模基因表达谱技术的出现及其飞速发展为肿瘤研究提供了一种全新的技术平台。基于基因表达谱的数据挖掘对致病基因发现、肿瘤临床诊断、药物疗效判断和发病机理阐明等意义重大。肿瘤基因表达谱数据多具特征维数高、样本小或相对小、样本背景差异大、存在批次效应等非随机噪声、冗余度高、非线性、基因间存在互作效应等特点,传统的统计方法和模式识别方法应用受限。本文针对基因表达数据特点,围绕信息基因选择方法和分类器构建展开研究,主要结果如下：(1)基于支持向量机发展了高维特征选择新方法二元矩阵重排过滤器BMSF (Binary Matrix Shift Filter)。大多数信息基因选择方法只考虑单个基因或成对基因的作用,却未考虑多个基因之间的相互作用。本文提出的BMSF算法综合考虑了多基因间的互作关系,通过引入随机产生的一个中间(0,1)二元矩阵,将分类问题转化为回归问题,实现了核函数参数寻优前提下基于支持向量机的高维特征选择。在基因选择过程中,对保留在模型中的基因子集根据其在肿瘤分类中对其他基因的贡献情况进行递归优化并反复更新。对9个癌基因表达二分类数据集, BMSF均以较小的信息基因子集获得了远优于文献报道的留一法预测精度,所选信息基因子集能同时提高多个分类器的留一法预测精度。(2)基于卡方测验发展了鲁棒的高维特征选择与无需训练的直接分类新算法TSG(Top-scoring genes)。预测精度既与特征选择有关,又受分类器的影响；训练是多数分类器产生过拟合的主要原因。主流算法TSP (Top score pairs)家族既是特征选择方法又是分类器,本文克服TSP不能反映样本大小、所选信息基因恒为偶数个、多分类时算法繁琐等缺陷,提出TSG算法。TSG提出并实现了基于转导推理、无需训练的直接分类,其决策过程为：先假定某个待测样本属于正(+)类,合并待测样本与训练样本得卡方值Chi+；再假定待测样本属于负(-)类,合并待测样本与训练样本得卡方值Chi-；如Chi+> Chi-,则待测样本属于正类,反之,则判为负类。多分类类推。TSG的特征选择过程为：先选取出得分最高的基因对TS2作为初始信息基因子集,接着每次从剩余的基因中挑选一个与已入选基因联合效应最好的基因添加到信息基因子集中,并根据训练集的留一法精度自动确定最终的信息基因子集。TSG对9个二分类和10个多分类数据独立预测均获得了明显优于文献报道的结果,特别是其训练集留一法预测精度与独立测试集预测精度相当接近,在部分数据集上独立测试精度甚至优于训练集留一法预测精度,显示TSG独特的、无需训练的直接分类能有效控制过拟合。(3)基于互作与卡方测验发展了信息基因选择新方法χ~2-IRG-DC (Chi-square test-based Integrated Rank Gene and Direct Classifier).χ~2-IRG-DC特征选择过程为：先利用单基因卡方值和成对基因互作卡方值,计算基因的综合加权得分,得基因的重要性排序；再基于χ~2-DC分类器序贯引入排序基因,并依训练集的留一法精度为第一标准、卡方增益为第二标准去冗余,获得了更为鲁棒的信息基因子集；最后基于χ~2-DC和信息基因实施独立预测。χ~2-IRG-DC继承TSG优点的同时,进一步通过基因综合加权评分大幅降低了算法复杂度,通过引入第二标准卡方增益增强了特征选择的鲁棒性。对9个二分类和10个多分类肿瘤基因表达谱数据集的独立预测精度表明,χ~2-IRG-DC模型明显优于文献报道；作为特征选择方法,χ~2-IRG-DC明显优于mRMR、SVM-RFE、HC-K-TSP、TSG等四种参比特征选择方法；作为分类器,χ~2-DC明显优于NB、KNN等参比分类器,与SVM分类器性能可比。本文方法对于推进高维数据特征选择和肿瘤分类识别具有重要理论意义和实用价值。更多还原

【Abstract】 Tumors are the consequences of interactions between multiple genes and the environment. The emergence and rapid development of large-scale gene-expression technology provide an entirely new platform for tumor investigation. Data mining based on gene expression profile plays an important role in the discovery of pathogenic genes, clinical diagnosis of tumor, judgment of therapeutic effect and mechanism of pathogenesis. Tumor gene expression profile data has the following features:high dimensionality, small or relatively small sample size, large differences in sample backgrounds, presence of nonrandom noise (e.g., batch effects), high redundancy, nonlinearity, and pairwise gene interactions. Traditional statistical methods and pattern recognition methods are limited. In this paper, According to the characteristics of gene expression profile data, the methods of information gene selection and the constructions of the classifier are studied. The main results are as follows:(1) Binary Matrix Shuffling Filter (BMSF), a new feature selection method for high-dimensional data based on support vector machine, is proposed. Most methods for gene selection in literature focus on screening individual or pairs of genes without considering the possible interactions among genes. In this paper, considering the interaction between multiple genes and introducing a random binary matrix, BMSF converts the classification problem to a regression problem. According to the optimization of the kernel function parameter, the high-dimensional feature selection is realized based on support vector machine regression. During the gene selection process, the set of genes to be kept in the model was recursively refined and repeatedly updated according to the effect of a given gene on the contributions of other genes in reference to their usefulness in cancer classification. The 9 binary classification datasets are expressed in the BMSF, and the accuracy of the prediction is far better than that reported by the literature. The small number of informative genes selected from each dataset leads to significantly improved leave-one-out (LOOCV) classification accuracy across all 9 datasets for multiple classifiers. Our method also exhibits broad generalization in the genes selected since multiple commonly used classifiers achieved either equivalent or much higher LOOCV accuracy than those reported in literature.(2) Top Scoring Genes (TSG), a new method based on chi-square test, is developed for high dimensional feature selection and direct classification. The prediction accuracy is related to not only the feature selection but also the classifier. Training is the major reason for model overfitting. The TSP family performs both feature-selection and classification. In this paper, TSG is developed. It overcomes the following problem of TSP family:it cannot reflect the size differences among samples, the numbers of the selected information genes are always even, and the algorithm of multi classification is complex.The direct classification with no demand for training is proposed and implemented for the first time, which is based on transduction inference. The classification process of TSG is as follows: Assume that a test sample belongs positive (+) class, and the chi-square value of the combined test sample and training samples is denoted as Chi+; Assume that the test sample belongs to the negative (-) class, and the chi-square value of the combined test sample and training samples is denoted as Chi-; If Chi+> Chi-, the test sample will be assigned to positive (+) class.Otherwise, it will be assigned to negative (+) class.The multiclass classification can be realized by analogy. The feature selection process of TSG is as follows: It starts with the top two genes and adds additional gene, which has the best combined effect with selected genes, into the candidate gene set to perform informative gene selection. The algorithm automatically reports the total number of informative genes selected with leave-one-out cross validation.The algorithm was applied to 9 binary and 10 multi-class gene expression datasets involving human cancers. The TSG classifier outperforms other classifiers in most of the 19 datasets. In particular, In particular, the prediction accuracy of the training dataset is quite close to that of the independent test dataset. The accuracy of the test dataset is even higher than that of the training dataset, and the results show TSG can effectively control the overfitting by the direct classification which need not train.(3) Chi-square test-based integrated rank gene and direct classifier (χ~2-IRG-DC) is developed for gene selection based on chi-square test and gene interactions. The feature selection process of χ~2-IRG-DC is as follows:First, we obtained the weighted integrated rank of gene importance from chi-square tests of single and pairwise gene interactions. Then, we sequentially introduced the ranked genes, and removed redundant genes according to leave-one-out cross-validation accuracy and the gain of chi-square value within the training set to informative genes. Finally, we determined the accuracy of independent test data by utilizing the genes obtained above with χ~2-DC. χ~2-IRG-DC method inherits the advantages of TSG, while greatly reduce the complexity of the feature selection by the weighted integrated rank of gene importance, and enhance the robustness of the feature selection by introducing chi-square gain as the second standard. The independent test accuracies of nine binary and ten multiclass tumor gene-expression datasets showed that χ~2-IRG-DC is obviously superior to the literature reported. As a feature selection method, χ~2-IRG-DC is better than mRMR, SVM-RFE, HC-K-TSP, TSG As a classifier, χ~2-DC is better than NB and KNN, and it is similar to the performance of SVM classifier.The proposed methods have important theoretical and practical value for promoting the feature selection of high dimensional data and classification of tumor.更多还原