局部晚期直肠癌新辅助化放疗病理学缓解的预测基因分析

Predictive genes for pathological response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer

目的局部晚期直肠癌仅10%~30%的患者显示出病理完全缓解。文中利用基因表达谱分析预测局部晚期直肠癌新辅助化放疗(NCRT)病理学缓解的基因标签。方法从GEO数据库中筛选具有新辅助化放疗后病理学缓解分级的局部晚期直肠癌基因表达谱数据,并获得4个数据集GSE35452、GSE46862、GSE68204和GSE53781。合并前3个数据集并对批次效应校正后分为训练集(n=121)和内部验证集(n=53),最后1个作为外部验证集(n=26)。训练集中单因素Logistic和t检验筛选与病理学缓解相关基因(未多重校正P<0.05),并按P值排秩。取P<0.05所有基因纳入LASSO算法,取前50个基因纳入SVM算法构建预测模型,并在对应的验证集中验证。反复随机取样500次以判断标签基因和模型的稳定性。取LASSO算法纳入模型次数最多的21个基因作为构建预测模型的候选基因,分别以在174例合并数据集或外部独立验证集中的Logistic回归系数与表达值乘积和作为放化疗敏感指标进行验证。在 174例合并数据集中分析 NCRT缓解组与未缓解组间差异表达基因和调控网络。结果 GSE35452、GSE46862和 GSE68204数据集共 12 803个基因纳入分析。LASSO 算法在内部验证集中诊断病理学缓解的准确性、特异性和灵敏度分别为 0.523(95% CI: 0.396~0.642)、0.578(95% CI: 0.373~0.762)、0.464(95% CI: 0.258~0.700)。SVM 法在内部验证集中的诊断准确性、特异性和灵敏度分别为 0.504(95% CI: 0.377~0.623)、0.596(95% CI: 0.393~0.830)、0.405(95% CI: 0.182~0.650)。21 基因放化疗敏感指标对 174 例合并数据集和外部独立验证集的病理学缓解诊断的 AUC 分别为0.863 (95% CI: 0.811~0.912)和 0.925 (95% CI: 0.817~1.000)。结论由于个体间肿瘤异质性的影响,基于特定人群基因表达构建的新辅助放化疗缓解预测模型在不同人群中效能较低。调控网络分析表明参与直肠癌浸润、转移以及干性转化的机制可能介导了直肠癌放化疗抵抗。

Objective Only 10%-30% of locally advanced rectal cancer(LARC)respond pathologically to neoadjuvant chemo-radiotherapy(NCRT). This study was to search for a feasible gene signature predicting pathological response to NCRT in LARC. Methods Four datasets GSE35452,GSE46862,GSE68204 and GSE53781 relating to the mRNA expression matrix and tumor regression grading of LARC after NCRT were obtained from the Gene Expression Omnibus. The first three datasets were merged into one and divided into training sets(n = 121)and internal validation sets(n = 53)after batch effect removal,and the last dataset was used as external validation sets(n = 26). Pathological response-related genes in the training sets were identified by univariate logistic regression and t-test(crude P < 0.05)and ranked by the P-value. All the genes with P < 0.05 were subjected to the least absolute shrinkage and selection operator(LASSO)and the first 50 to the support vector machine algorithm(SVM)for the establishment of predicting models,followed by verification in the corresponding validation set. Random sampling was repeated 500 times to determine the stability of the selected gene signatures and models. With the 21 most important genes revealed by LASSO as the candidates for model construction,the sensitivity index for NCRT was calculated as the total sum of coefficients in logistic regression and expression values in the merged datasets and external validation sets. The differentially expressed genes were identified between the response and non-response groups in the 174 merged datasets and subjected to regulatory network analysis. Results A total of 12 803 genes from the GSE35452,GSE46862 and GSE68204 datasets were included in the analysis. The accuracy,specificity and sensitivity of LASSO for predicting the pathological response in the internal validation sets were 0.523(95% CI:0.396-0.642,0.578(95% CI:0.373-0.762) and 0.464(95% CI:0.258-0.700),while those of SVM were 0.504(95% CI:0.377-0.623),0.596(95% CI:0.393-0.830)and 0.405(95% CI:0.182-0.650),respectively. The area under the ROC curve(AUC)for pathological response prediction was 0.863 (95% CI:0.811-0.912)in the 174 merged datasets and 0.925(95% CI:0.817-1.000)in the external validation sets. Conclusion The model for predicting response to NCRT established using the expression of candidate genes identified from a specific set of patients has a frustratingly low capacity in an independent set,probably because of high tumor heterogeneity among different individuals. Regu? latory network analysis indicates that radiotherapy-resistance in rectal cancer may be mediated by the mechanisms underlying the inva? sion,metastasis and transformation of the malignancy.