基于多尺度局部累积特征和神经网络的抗肿瘤药物反应预测

Prediction of Antitumor Drug Response Based on Multiscale Local Cumulative Features and Neural Networks

目前已有的研究结果表明现有抗肿瘤药物的有效性高度依赖于患者的基因组学特征.如何为每位肿瘤患者量身定制最佳的治疗方案是重要又富有挑战性的前沿课题.针对该课题,文中提出抗肿瘤药物反应预测方法,运用机器学习技术,对患者肿瘤基因测序数据进行处理、特征提取及建模,预测各种不同抗肿瘤药物的疗效反应.首先,提出基于多尺度关联规则的数据挖掘方法,对基因组学数据进行不同尺度的特征挑选.进而通过累积窗函数对挑选后的基因组学数据进行局部累积,进一步执行数据压缩,提取具有较强整体表达性的基因特征信息.然后,以多层全连接神经网络为模型、以提取的多尺度累积基因特征为输入样本,进行训练和建模.最后,分别采用特征融合和决策融合,实现某一肿瘤基因测序数据对于各种不同抗肿瘤药物反应结果的预测.在COSMIC、GDSC数据库上的仿真实验表明,文中方法在敏感性、特异性、准确率、特性曲线面积值等关键性能指标上均取得较优值.

Medical research results show that the effectiveness of an antitumor drug is highly dependent on the genomic characteristics of patients.How to customize an optimal medical treatment for each tumor patient is an extremely important and challenging research topic.Aiming at this subject,a set of methods to predict the efficacy response of various antitumor drugs is proposed by machine learning technology for data processing,feature extraction and modeling of the tumor gene sequences of patients.Firstly,a data mining algorithm based on multiscale association rules is proposed for feature selection at different scales of genomics data.Then,the selected genomics data are locally accumulated by the cumulative window function to further compress the data and extract the gene feature information with stronger overall expression.Based on the above,a fully connected multi-layer neural network is designed and the extracted multiscale cumulative gene features are treated as input samples to train the network.Finally,two fusion methods,including feature fusion and decision fusion,are utilized to predict the responses of a tumor gene sequence to different antitumor drugs,respectively.Results of simulation experiments show that the proposed approach is superior in key performance indexes,such as sensitivity,specificity,accuracy and the area under characteristic curve.