Iterative Partial Least Squares for Discovering Biomarkers of Diabetic Nephropathy
|關鍵字:||代謝體學;糖腎病;生物標記;偏最小平方法;Metabolomics;Diabetic Nephropathy;Biomarker;Partial Least Squares|
結果：在DM和DN1的分類中選出44個代謝物做為分類模型，並達到AUC為0.8858(10-fold cv)；DM和DN2的模型選出124個代謝物，其AUC為0.9985；DN1和DN2的分類模型選出30個代謝物並達到AUC = 0.9552，另以三種糖腎病分期所建立的回歸模型選出140個代謝物，其方均根誤差為1.4836，所有選出的代謝物交集後，有15個代謝物在各個模型中均有挑選，其中有10個代謝物隨著分期階段遞增或遞減。
Background: Diabetic nephropathy (DN) is a kidney disease caused by long term diabetes mellitus (DM). In the developed countries, the number of patients with DN has increased annually, caused a huge financial burden to patients itself and countries. The progression of DN can be slowed down or well controlled once it is treated at early stage. However, there is no significant clinical symptoms in the early stage of DN, which thus makes it difficult for early detection of DN. Metabolomics can reveal the causes of the disease by analyzing the differences of metabolite profile between patients and healthy individuals. This study aims to find the potential biomarkers of diabetic nephropathy by analyzing different stage of DN patients’ metabolite profile. Methods: We collected 54 urine samples from different stage of DN patients (DM, DN1, DN2, 18 patients per group). After extracting metabolite profile by LC-ESI-TOF-MS, this study proposed an iterative PLS method to analyze the different metabolite between these groups. Then combined the results observed from different comparison to find metabolites that showed significant difference between three groups. Results: In DM vs. DN1 classification model, there were 44 metabolites used for differentiating DM and DN1 patients, with 10-fold CV AUC = 0.8858. The model chose 124 metabolites for the classification of DM and DN2 and yield AUC = 0.9985. As for the classification of DN1 and DN2, the model used 30 metabolites and achieved AUC = 0.9552. In the regression model, there were 140 metabolites used to predict patients UACR, with RMSE = 1.4836. The intersection of metabolites observing from different model showed 15 metabolites that has been chosen among all models. 10 of 15 metabolites showed increasing or decreasing trend as the severity of DN increase. Conclusions: Method proposed in this study can have better performance with less variables compared to regular VIP>1 method. The metabolites discovered in this study showed relations with severity of DN, but the relations between DN and these metabolites are still unclear. This method and the metabolites should be confirmed in further research with a larger cohort.
|Appears in Collections:||Thesis|