基于多组学数据的乳腺癌预后预测模型构建

Construction of Breast Cancer Prognosis Prediction Model Based on Multi-Omics Data

摘要: 本文主要从UCSC Xena数据库中已经整理好的关于TCGA数据库的乳腺癌数据中，挑选了拷贝数变异、RNA基因表达量、RNA外显子表达量三个组学方面的数据。首先，基于三个组学数据的维度远大于样本量的特征，分别对三个组学的数据进行方差阈值过滤，初步筛选过滤掉变化幅度不大的变量，再使用mRMR进行滤波式的变量选择方法，即最大化特征与分类变量之间的相关性，最小化特征之间的相关性，各自筛选得到50个变量。对于离散型的天数表型数据，采用阈值方法将其转化为0-1分类变量，最终将因变量与自变量进行合并，并划分测试集、训练集，使用svm、XGBoost、Logistic、RandomForest四种方法对结果变量进行预后预测，并采用特定的指标对这四种算法进行比较，运用在训练集上，最终得到XGBoost、Logistic两种算法的预测效果要优于svm、RandomForest。

Abstract: In this paper, we mainly selected the three omics data of copy number variation, RNA gene expres-sion, and RNA exon expression from the breast cancer data on the TCGA database that have been collated in the UCSC Xena database. Firstly, based on the characteristics of the three omics data whose dimensions are much greater than the sample size, the variance threshold filter is performed on the three omics data, the variables with little change are filtered out initially, and then the vari-able selection method is filtered by using mRMR, that is, to maximize the correlation between the features and the categorical variables, minimize the correlation between the features, and filter 50 variables each. For the discrete number of days phenotypic data, the threshold method is used to convert it into a 0-1 categorical variable, and finally the dependent variable is merged with the in-dependent variable, and the test set and the training set are divided, and the outcome variable is predicted by svm, XGBoost, Logistic, randomForest, and the four algorithms are compared with spe-cific indicators, and the training set is applied to the training set, and finally XGBoost. The predic-tion effect of logistic algorithms is better than that of svm and RandomForest.