摘要
目的利用机器学习构建皮肤恶性黑色素瘤转移特征预测模型。方法基于TCGA数据库中470例皮肤恶性黑色素瘤样本数据(其中原发性肿瘤TP组103例,转移肿瘤TM组367例),分3种方法筛选转移特征基因:1.使用edge R与DESeq2R包筛选TP组与TM组表达差异基因;2.使用加权基因共表达网络分析(weighted correlation network analysis,WGCNA),鉴定肿瘤转移特征性状相关基因模块,并用GO(gene ontology)和GSEA(gene set enrichment analysis)富集分析转移特征相关基因的生物学功能;3.通过受试者工作特征曲线(receiver operating characteristic curve,ROC)筛选转移特征性基因,三者取交集后,利用机器学习XGBoost算法构建转移相关基因特征分类模型,并用来自GEO数据库的外部测试集GSE190113进行验证。结果共筛选TP组与TM组差异表达基因1804个,WGCNA分析发现黑色模块与转移性状密切相关(r=-0.46),该模块包含529个基因,ROC分析基于AUC(Area under curve)值>0.7标准共筛选806个基因,三者取交集后获得142个基因,利用XGBoost算法成功构建皮肤恶性黑色素瘤转移特征判定模型,内部测试集预测AUC值为0.88,外部测试集预测AUC值为0.85。结论本机器学习模型能较为准确预测皮肤恶性黑色素瘤的转移特征,对帮助判断患者的临床分型及分期、优化治疗方案具有一定指导意义。
Objective To construct a predictive model of metastatic characteristics of cutaneous malignant melanoma based on machine learning.Methods The sample data of 470 cases of cutaneous malignant melanoma were obtained from the TCGA database,and included 103 cases in the primary tumor TP group and 367 cases in the metastatic tumor TM group.The metastatic characteristic genes were screened by three methods:1.edgeR and DESeq2R packages were used to screen the differentially expressed genes between TP group and TM group;2.Weighted correlation network analysis(WGCNA)was used to identify the gene modules related to the characteristics of tumor metastasis,and GO(gene ontology)and GSEA(gene set enrichment analysis)were used to analyze the biological functions of genes related to metastasis characteristics;3.The transfer characteristic genes were screened by receiver operating characteristic curve(ROC).After the intersection of the three methods,the feature classification model of transfer related genes was constructed by machine learning XGBoost algorithm,and verified by the external test set GSE190113 from GEO database.Results A total of 1804 differentially expressed genes in TP group and TM group were screened.WGCNA analysis found that the black module was closely related to the metastasis character(r=-0.46),and this module contained 529 genes.ROC analysis screened806 genes based on the AUC(area under curve)value>0.7.After the intersection of the three methods,142 genes were obtained.The XGBoost algorithm was used to successfully construct the metastasis characteristic judgment model of skin malignant melanoma.The AUC value predicted of the internal test set was 0.88,and the predicted AUC value of the external test set was 0.85.Conclusion This machine learning model can accurately predict the metastatic characteristics of cutaneous malignant melanoma,and has certain guiding significance for judging the clinical classification and staging of patients and optimizing the treatment plan.
作者
王英伦
何孜灏
黎思颖
郑子豪
董博文
邓展程
陈秋锐
沈晗
WANG Yinglun;HE Zihao;LI Siying;ZHENG Zihao;DONG Bowen;DENG Zhancheng;CHEN Qiurui;SHEN Han(School of Life Sciencesand Biopharmaceutics,Guangdong Pharmaceutical University,Guangzhou 510006,China;Guangdong Province Key Laboratory for Biotechnology Drug Candidates,Guangzhou 510006,China)
出处
《广东药科大学学报》
CAS
2022年第3期85-92,共8页
Journal of Guangdong Pharmaceutical University
基金
广东省自然科学基金项目(2018A030313114)
广东省普通高校特色创新类项目(2021KTSCX053)
广东省大学生创新创业训练计划项目(S201910573039)。
关键词
皮肤恶性黑色素瘤
转移
机器学习
分类模型
cutaneous malignant melanoma
metastasis
machine learning
classification model