构建6-基因直肠癌新辅助放化疗后区域淋巴结完全退缩的人工神经网络预测模型

A six-gene model using an artificial neural network to predict regional lymph node metastasis after neo- adjuvant chemoradiotherapy for locally advanced rectal cancer

目的筛选直肠癌新辅助放化疗（CRT）后区域淋巴结完全退缩（ypN-）的分子标志物，构建基于人丁神经网络（ANN）的CRT后ypN-的预测模型。方法从公共基因芯片数据库下载一组直肠癌新辅助放化疗患者的芯片数据（GSE46862），包括69例直肠癌样本．共64例样本纳入研究。21例直肠癌CRT后ypN-设为ypN-组，43例直肠癌CRT后区域淋巴结阳性残留（ypN＋）设为ypN＋组。通过GCBI在线平台分析数据。比较直肠癌CRT后ypN-组与ypN＋组的基因表达谱，寻找差异表达基因以筛选分子标志物。通过绘制受试者工作特征（ROC）曲线筛选排名前6名的差异表达基因进行模型构建。ANN预测模型构建采用SPSSModeler完成。按照7：3将总样本随机裂解成训练样本和测试样本。运用训练样本进行ANN预测模型构建，用测试样本进行独立回代验证。观察指标：（1）差异表达基因筛选结果。（2）ANN预测模型分析结果。绘制ROC，计算曲线下面积（AUC）评估各分子标志物和ANN预测模型的预测能力。结果（1）差异表达基因筛选结果：共筛选出50个差异表达基因。前6名的基因分别为IL6、AKRlBl、AREG、SELE、ROB01、CD274。（2）ANN预测模型分析结果：选择上述6个基因，构建ANN预测模型。该ANN为3层7-5-2结构。对ANN贡献最大的是IL6，之后分别为ROB01、AKRlBl、AREG、CD274、SELE。该模型的AUC为0．929。其灵敏度为96．7％，特异度为85．7％，预测训练样本的准确率为93．2％。当进行独立回代时，其预测灵敏度为92．3％，特异度为85．7％，预测独立测试样本的准确率为90．0％。结论基于多分子标志物（IL6、ROB01、AKRlBl、AREG、CD274、SELE）成功构建了CRT后ypN-的精准预测模型，稳定性好．可为直肠癌CRT后保直肠手术的临床决策提供参考。

Objective To screen out the potential gene to predict regional lymph node metastasis after neoadjuvant chemoradiotherapy （CRT） for locally advanced rectal cancer （LARC） and develop a 6-gene model using an artificial neural network （ANN）. Methods The gene expression profiles （GSE46862） of locally advanced rectal cancer undergoing preoperative chemoradiotherapy from 64 specimens （ 21 with ypN- and 43 with ypN＋） were downloaded from the gene expression omnibus （GEO） database. The differentially expressed genes were identified to screen out the potential biomarkers through the Gene-Cloud of Biotechnology Information （GCBI） platform. The top 6 genes were screened out for building model. An ANN model was trained and validated using the SPSS Modeler software. The study samples were allocated randomly into the training sample group and testing sample group with a 7：3 ratio. The training samples and testing samples were respectively used for building an ANN model and independent back-substitution test. Observation indicators ： （ 1 ） screening results of differentially expressed genes ; （2） analysis results of ANN model. The receiver operating characteristic （ ROC ） curve was drawn and the area under the curve （AUC） was calculated to evaluate the predictive abilities of ANN and each biomarker. Results （ 1 ） Screening results of differentially expressed genes ： A total of 50 genes were screened. Six top genes included IL6, AKR1B1, AREG, SELE, ROBO1 and CD274. （2） Analysis results of ANN model： Six top genes were selected to construct a three-layer ANN model with a 7-5-2 structure. The IL6 made the greatest effect on the ANN model, followed by ROBO1, AKR1B1, AREG, CD274 and SELE. The AUC was 0. 929. The sensitivity and specificity of ANN model were 96. 7% and 85.7%, and accuracy of training samples was 93.2%. In the independent back-substitution test, sensitivity and specificity were 92. 3% and 85.7%, and accuracy of testing samples was 90. 0%. Conclusion The prediction ANN model based on multiple molecular markers （ IL6, ROBO1, AKR1B1, AREG, CD274 and SELE） for regional lymph node metastases in LARC patients after CRT would be beneficial in selecting potential candidates for rectum-preserving surgery following CRT for LARC.