Identification of biomarkers that distinguish chemical contaminants using a gradient feature selection method

In this study, we divided 105 compounds plus vehicle control into 14 compound classes. On the basis of gene expression profiles of in vitro primary cultured hepatocytes, we comprehensively compared various normalization, feature selection and classification algorithms for the classification of these 14 class compounds. We found that normalization had little effect on the averaged classification accuracy. Two support vector machine methods LibSVM and SMO had better classification performance. When feature sizes were smaller, LibSVM outperformed other classification methods. Simple logistic algorithm also performed well. At the training stage, usually the feature selection method SVM-RFE performed the best, and PCA was the poorest feature selection algorithm. But overall, SVM-RFE had the highest overfitting rate when an independent dataset used for a prediction in this case. Therefore, we developed a new feature selection algorithm called gradient method which had a pretty high training classification as well as prediction accuracy with the lowest over-fitting rate. Through the analysis of biomarkers that distinguished 14 class compounds, we found a goup of genes that mainly invovled in cell cylce were significanly downregulated by the metal and inflammatory compounds, but were induced by anti-microbial, cancer related drugs, pesticides, and PXR mediators.本研究将105种化合物与溶媒对照（vehicle control）划分为14个化合物类别。基于体外原代培养肝细胞的基因表达谱，本研究针对这14类化合物的分类任务，全面比较了多种归一化、特征选择与分类算法。研究发现，归一化操作对平均分类准确率的影响极小。两种支持向量机（Support Vector Machine, SVM）方法——LibSVM与SMO——展现出更优的分类性能。当特征维度较小时，LibSVM的表现优于其他分类算法。简单逻辑回归算法同样表现优异。在训练阶段，特征选择方法SVM-RFE通常表现最佳，而主成分分析（Principal Component Analysis, PCA）则是性能最差的特征选择算法。但总体而言，当使用独立数据集进行预测时，SVM-RFE的过拟合率在所有算法中最高。为此，本研究开发了一种名为梯度法的新型特征选择算法，该算法兼具较高的训练分类准确率与预测准确率，且过拟合率最低。通过对区分14类化合物的生物标志物进行分析，本研究发现：主要参与细胞周期（cell cycle）的一组基因，会被金属类与炎症类化合物显著下调，却会被抗微生物药物、癌症相关药物、农药以及孕烷X受体（Pregnane X Receptor, PXR）介导剂诱导上调。