基于GEO数据库对子痫前期关键基因筛选及 诊断模型构建的研究

Research on Key Gene Screening and Diagnostic Model Construction for Pre-Eclampsia Based on the GEO Database

摘要: 目的：旨在挖掘子痫前期(Preeclampsia, PE)患者相较正常孕妇胎盘组织中的关键差异表达基因及其临床诊断价值，并探索性构建候选诊断模型。方法：从Gene Expression Omnibus (GEO)公共数据库下载GSE74341、GSE10588和GSE25906三个PE胎盘组织基因表达数据集，筛选差异基因，采用生物信息学技术进行基因本体(Gene Ontology, GO)过表征分析和基因集富集分析(Gene Set Enrichment Analysis, GSEA)分析差异基因功能。以LASSO回归构建主诊断模型，随机森林和支持向量机作为敏感性分析模型；通过嵌套五折交叉验证、独立外部验证和留一队列交叉验证评估模型性能。结果：共筛选出686个显著差异基因，其中上调347个、下调339个。GO富集分析显示差异基因主要参与氧水平响应(response to oxygen levels)、缺氧反应(response to hypoxia)、氧含量降低响应(response to decreased oxygen levels)和骨化(ossification)等生物过程，GSEA分析提示脂肪因子信号通路、细胞周期和细胞黏附分子相互作用等通路呈趋势性富集。LASSO算法构建了由YWHAB、TIMM22、TMEM45A、KARS、QPCT、SDHC、ARHGEF16、CCNH、CLN6、CNOT2、COPS5、EGFR和EPC2组成的13基因诊断面板，其中YWHAB、TIMM22和TMEM45A为交叉验证中稳定入选的基因(选择频率 ≥ 80%)。模型在发现队列内嵌套交叉验证AUC = 0.937 (95%CI 0.870~0.989)，敏感性为0.875，特异性为0.944，准确率为0.912，Brier评分为0.139；模型在GSE25906独立外部队列上AUC = 0.659 (95%CI 0.510~0.803)，敏感性为0.696，特异性为0.703，准确率为0.700，Brier评分为0.259；三队列LOCO-CV平均AUC = 0.654 (范围0.513~0.846)。结论：本研究在PE胎盘组织中筛选出以YWHAB、TIMM22和TMEM45A为核心的候选差异表达基因，并构建了以13个基因为核心的LASSO候选诊断面板。该模型在发现队列内部具有良好判别能力，但外部队列和留一队列验证提示跨队列泛化能力有限，仍需在更大规模、表型一致且平台标准化的独立队列中进一步验证。

Abstract: Objective: To identify key differentially expressed genes (DEGs) in placental tissues from patients with preeclampsia (PE) compared with normal pregnant women, evaluate their clinical diagnostic value, and exploratorily construct candidate diagnostic models. Methods: Three PE placental tissue gene expression datasets, GSE74341, GSE10588, and GSE25906, were downloaded from the Gene Expression Omnibus (GEO) public database. DEGs were screened, and bioinformatics approaches were used to perform Gene Ontology (GO) over-representation analysis and Gene Set Enrichment Analysis (GSEA) to investigate their biological functions. A primary diagnostic model was constructed using LASSO regression, with random forest and support vector machine models used for sensitivity analysis. Model performance was evaluated using nested five-fold cross-validation, independent external validation, and leave-one-cohort-out cross-validation. Results: A total of 686 significant DEGs were identified, including 347 upregulated and 339 downregulated genes. GO enrichment analysis showed that these DEGs were mainly involved in biological processes such as response to oxygen levels, response to hypoxia, response to decreased oxygen levels, and ossification. GSEA suggested trend-level enrichment of pathways including adipocytokine signaling, cell cycle, and cell adhesion molecule interactions. A 13-gene diagnostic panel was constructed using the LASSO algorithm, consisting of YWHAB, TIMM22, TMEM45A, KARS, QPCT, SDHC, ARHGEF16, CCNH, CLN6, CNOT2, COPS5, EGFR, and EPC2. Among these, YWHAB, TIMM22, and TMEM45A were stable genes across folds, with selection frequencies of ≥ 80%. In nested cross-validation within the discovery cohort, the model achieved an AUC of 0.937, 95% CI: 0.870~0.989, with sensitivity of 0.875, specificity of 0.944, accuracy of 0.912, and a Brier score of 0.139. In the independent external validation cohort GSE25906, the model achieved an AUC of 0.659, 95% CI: 0.510~0.803, with sensitivity of 0.696, specificity of 0.703, accuracy of 0.700, and a Brier score of 0.259. The mean AUC in three-cohort leave-one-cohort-out cross-validation was 0.654, with a range of 0.513~0.846. Conclusion: This study identified candidate DEGs centered on YWHAB, TIMM22, and TMEM45A in PE placental tissues and constructed a 13-gene LASSO-based candidate diagnostic panel. The model showed good discriminative performance within the discovery cohort; however, external validation and leave-one-cohort-out validation indicated limited cross-cohort generalizability. Further validation in larger independent cohorts with consistent phenotypes and standardized platforms is warranted.