摘要目的:基于非负矩阵分解（NMF）聚类算法鉴定胶质母细胞瘤（GBM）肿瘤微环境相关分子亚型，建立风险评分模型预测GBM预后。方法:提取癌症基因图谱库和GBM患者转录组数据（ n=161），基于NMF算法鉴定最佳分子聚类亚型，探索不同聚类亚型在预后方面是否存在显著差异。 Cox模型回归筛选差异表达基因并构建风险评分将GBM患者分为高风险组和低风险组。绘制受试者工作特征（ROC）曲线评估风险评分在不同队列中预测GBM患者1、3年总生存率的准确性，在基因表达综合队列中进行验证。 结果:3年总生存率GBM患者聚类C2亚型好于聚类C1亚型（0.12∶0.25， P=0.015）。两种聚类亚型在6种不同免疫免疫细胞等浸润方面均不同（均 P<0.05）。训练组（ n=97）和外部验证组（ n=64）GBM患者高风险组3年总生存率差于低风险组（0.18∶0.31， P=0.002和0.06∶0.16， P=0.036）。预后模型在癌症基因图谱库和外部验证组预测3年准确性的ROC曲线下面积分别为0.74和0.82。 结论:基于NMF算法成功聚类两种GBM分子亚型，大多数表现为低免疫细胞等浸润型。进一步构建了肿瘤微环境相关基因的预测模型，为GBM患者治疗提供理论基础。

abstractsObjective:To identify tumor microenvironment (TME) related molecular subtypes of glioblastoma (GBM) using non-negative matrix factorization (NMF) clustering algorithm, and to establish a TME-related risk score model for predicting GBM prognosis.Methods:The transcriptome data of GBM patients were extracted from the Cancer Genome Atlas Program (TCGA) public databases ( n=161). NMF algorithm was used to identify the best molecular clusters, and the potential significant differences in prognosis between different clusters were explored. The differentially expressed genes associated with GBM prognosis were screened by Cox regression. The risk score (RS) was established to divide GBM patients into the high-risk and low-risk groups. The accuracy of RS in predicting 1-and 3-year overall survival (OS) rates in different groups was evaluated by receiver operating characteristic (ROC) curve, and the external applicability of risk score was validated in the gene expression omnibus cohort. Results:The 3-year OS rate showed that Cluster2 subtype had significantly longer than Cluster1 subtype (0.12 vs 0.25, P=0.015). The infiltration of six different immune cells showed significant differences between the two molecular subtypes (all P<0.05). The 3-year OS rate of the high risk group was worse than that of the low risk group among GBM patients in the training group ( n=97,0.18 vs 0.31, P=0.002) and the external validation group ( n=64,0.06 vs 0.16, P=0.036). The areas under the ROC curve for the TCGA and the external validation group were 0.74 and 0.82, respectively. Conclusions:This study successfully clustered two molecular subtypes of GBM based on NMF algorithm, most of GBM tends to Cluster2 subtype. Furthermore, a prediction model based on TME-related genes was established, providing potential theoretical basis for the treatment of GBM patients.