基因表达对膀胱癌生存结局的预测:稀疏与混合Cox模型的实证比较研究

Prognostic Prediction of Bladder Cancer with Gene Expressions:An Empirical Comparison Study of Sparse and Mixed Cox Models

目的研究稀疏Cox(coxlasso)与混合Cox模型(coxlmm)在全基因表达数据中对膀胱癌预后的预测表现。方法通过计算一致性指数(C-index)评价两种模型在膀胱癌全基因表达数据中(TCGA,GSE31684和GSE13507)的预测精度,同时在混合Cox模型中将膀胱癌的生存方差划分为临床(PCE)和转录组(PGE)两部分。结果当TCGA数据集为训练集时,coxlmm预测能力(C-index=0.676)高于coxlasso(C-index=0.655),两者在外部验证集的C-index分别为0.527和0.534。当三个合并数据集为训练集时coxlmm(C-index=0.671)比coxlasso(C-index=0.650)的预测精度提高2.1%。当GSE31684为训练集时,coxlmm(C-index=0.553)比coxlasso(C-index=0.550)的预测精度提高0.3%,两个模型在外部验证集上的C-index分别为0.632和0.633。生存方差划分表明膀胱癌的临床贡献高于转录组的贡献(PCE=14.95%,PGE=10.88%)。结论成功构建了一种用于膀胱癌的预后预测的coxlmm模型,揭示了整合全转录组信息可在一定程度上提高膀胱癌预后预测能力。

Objective To compare the predictive performance of sparse Cox model(coxlasso)and Cox linear mixed model(coxlmm)for prognosis of bladder cancer with whole gene expression profile.Methods The prediction accuracy of the two models was evaluated with bladder cancer datasets(TCGA,GSE31684 and GSE13507)using the concordance index(C-index);the survival variance of bladder cancer was partitioned into the relative contribution of clinical and transcriptional components within the framework of coxlmm.Results By integrating whole genes and clinical covariates in the TCGA training set,we confirmed that the power for prognostic prediction of coxlmm model(C-index=0.676)was higher than coxlasso model(C-index=0.655).The C-index of two models was 0.527 or 0.534 in the external validation set.When the three data sets were combined to be the training set,the prediction accuracy of coxlmm(C-index=0.671)is higher than that of coxlasso(C-index=0.650).When the GSE31684 was the training set,the prediction accuracy of coxlmm(C-index=0.553)is also slightly higher than that of coxlasso(C-index=0.550).The C-index of two models was 0.632 or 0.633 in the external validation set.The survival variance partition demonstrated the clinical contribution was higher than transcriptomic contribution for bladder cancer.Conclusion This study successfully built and validated a coxlmm model for the prognostic prediction of bladder cancer.This study indicated the aggregation of genome-wide transcriptomic information can improve prognostic prediction power of bladder cancer.