基于人工神经网络挖掘急性心肌梗死患者脂质代谢枢纽基因及靶向中药预测

Mining lipid metabolism hub genes and targeted traditional Chinese medicine prediction in patients with acute myocardial infarction based on artificial neural networks

摘要:

目的]建立基于GEO数据库急性心肌梗死患者脂质代谢相关基因人工神经网络(ANN)模型并评价其效果,同时对枢纽基因进行靶向中药预测。 [方法]通过GEO数据库获取6份急性心肌梗死芯片数据,合并GSE48060、GSE60993和GSE66360作为训练数据集,GSE34198、GSE61144和GSE12288分别作为验证数据集,使用LASSO、随机森林和SVM-RFE三种机器学习算法筛选出与急性心肌梗死患者脂质代谢相关的枢纽基因,随后,基于上述筛选出的枢纽基因构建ANN模型。在独立验证集GSE34198、GSE61144和GSE12288中验证模型的效能,通过曲线下面积(AUC)分析评估ANN模型的临床应用价值；随后构建列线图以预测疾病发生概率,并采用RT-qPCR验证枢纽基因的表达水平。最后,基于枢纽基因预测靶向中药,并通过分子对接技术验证预测结果。 [结果]共筛选得到17个脂质代谢相关差异基因,其中上调基因14个,下调基因3个。进一步通过LASSO、随机森林及SVM-RFE三种机器学习算法筛选,得到ACSL1、DGAT2、MBOAT2和CH25H这4个枢纽基因,并据此构建了ANN诊断模型。使用该模型分别绘制训练组与验证组诊断的ROC曲线,结果显示:训练集的AUC值为0.936(P＜0.001)；验证集GSE34198、GSE61144和GSE12288的AUC值依次为0.786(P＜0.001)、0.821(P＜0.01)和0.900(P＜0.001)。基于4个脂质代谢相关枢纽基因构建了列线图,校准曲线显示,列线图模型的预测概率与理想模型几乎一致。RT-qPCR实验进一步验证了上述结果的可靠性。基于枢纽基因筛选出27味中药,其四气五味以温、平、寒为主,味多苦、辛、甘,归经多属肝、脾、肺及心经,功效以清热解毒、活血通络、逐水祛痰为主。分子对接结果显示,芸香、椰子和鱼脑石的结合能分别为－6.9 kcal/mol、－7.4 kcal/mol和－8.2 kcal/mol,提示其含有的小分子化合物与靶蛋白之间存在良好的相互作用。 [结论]本研究基于机器学习方法筛选获得急性心肌梗死患者4个脂质代谢相关枢纽基因(ACSL1、DGAT2、MBOAT2和CH25H),并据此构建了具有较高诊断效能的ANN模型。进一步结合靶向中药预测,为急性心肌梗死的早期识别及潜在干预策略提供了新的生物信息学依据。

Abstract:

Aim To establish an artificial neural network (ANN) model based on the GEO database for lipid metabolism-related genes in patients with acute myocardial infarction and evaluate its effectiveness, while predicting targeted traditional Chinese medicines for hub genes. Methods Six acute myocardial infarction chip data were obtained from the GEO database, and GSE48060, GSE60993 and GSE66360 were merged as training datasets, GSE34198, GSE61144 and GSE12288 were used as validation datasets, respectively. LASSO, random forest and SVM-RFE machine learning algorithms were used to screen hub genes related to lipid metabolism in patients with acute myocardial infarction, subsequently, an ANN model was constructed based on the selected hub genes. The effectiveness of the model was validated in independent validation sets GSE34198, GSE61144 and GSE12288, the clinical application value of ANN model was evaluated through area under the curve (AUC) analysis. Subsequently, a nomogram was constructed to predict the probability of disease occurrence, and RT-qPCR was used to validate the expression levels of hub genes. Finally, targeted traditional Chinese medicines were predicted based on hub genes, and the prediction results were validated through molecular docking technology. Results A total of 17 differentially expressed genes related to lipid metabolism were screened, including 14 upregulated genes and 3 downregulated genes. Four hub genes (ACSL1, DGAT2, MBOAT2 and CH25H) were further screened using three machine learning algorithms:LASSO, Random Forest and SVM-RFE, and an ANN diagnostic model was constructed. The ROC curves for the diagnosis of the training and validation groups were plotted using this model. The AUC value for the training set was 0.936 (P＜0.001), while the AUC values for the validation sets GSE34198, GSE61144 and GSE12288 were 0.786 (P＜0.001), 0.821 (P＜0.01), and 0.900 (P＜0.001), respectively. A nomogram was constructed based on four lipid metabolism related hub genes, and the calibration curve showed that the prediction probability of the nomogram model was almost consistent with the ideal model. RT-qPCR experiments further validated the reliability of the above results. Based on the hub genes, 27 traditional Chinese medicines were screened, with the four natures and five flavors mostly belonging to warm, calm, cold, bitter, spicy and sweet, and mostly returning to the liver, spleen, lungs and heart meridians. Their functions were mainly to clear heat and detoxify, promote blood circulation and unblock meridians, and drain dampness and eliminate phlegm. The molecular docking results showed that the binding energies of Ruta graveolens, Coconut and Ambergris Fish Bone were －6.9 kcal/mol, －7.4 kcal/mol and －8.2 kcal/mol, respectively, indicating a good interaction between the small molecule compounds they contain and the target protein. Conclusions This study employed machine learning methods to identify four lipid metabolism-related hub genes (ACSL1, DGAT2, MBOAT2 and CH25H) in acute myocardial infarction patients, and constructed an ANN model with high diagnostic efficacy. Furthermore, targeted traditional Chinese medicine prediction was performed based on these hub genes, providing new bioinformatics evidence for the early identification and potential intervention strategies of acute myocardial infarction.