使用LASSO多變數線性回歸尋找糖腎病尿液代謝物生物標記

Abstract

背景：糖腎病是糖尿病最主要的併發症之一，糖腎病有不同的階段，從早期的微量白蛋白尿逐漸發展白蛋白尿，甚至進展到終末期腎臟病，本實驗的目的是從尿液代謝物中挑出一組代謝物組合能達到現今的診斷標準尿液白蛋白跟肌酸酐比值（UACR），並分析這些尿液代謝物的的代謝途徑。 方法：研究樣本包括五十四位二型糖尿病患者，經由統計分析及多重檢驗校正挑出顯著差異的代謝物，再進一步使用最小絕對值收斂和選擇算子（LASSO）邏輯式回歸分析執行類別分類，使用LASSO線性回歸分析預測UACR的對數數值。 結果：經過LASSO邏輯式回歸分析，從非白蛋白患者尿與輕微白蛋白尿患者的顯著差異代謝物中挑出三個代謝物，分類結果AUC為0.938，從輕微白蛋白尿患者與白蛋白尿患者的顯著差異代謝物中挑出十一個代謝物，分類結果AUC為0.966，從輕微白蛋白尿患者與白蛋白尿患者的顯著差異代謝物中挑出六個代謝物，分類結果AUC為0.963，經過LASSO線性回歸分析，UACR的對數與預測數值的相關係數為0.672。 結論：研究結果表示統計分析後的篩選會提升LASSO的分類能力，從中找到三組區分不同階段的糖腎病生物標記，其生物標記可能與抑制SIRT1蛋白質或與服用藥物有關。

Background: Diabetic nephropathy is the most important complication of diabetes mellitus, diabetic nephropathy has different stages from early stage microalbuminuria to macroalbuminuria, even lead to end-stage renal disease. The purpose of this study is found a urine metabolite combination that could fit diabetic nephropathy's criteria of clinical diagnosis, urine albumin creatinine ratio (UACR), furthermore, analyzes those metabolite's pathways. METHODS: The study sample included 54 type 2 diabetes mellitus patients. Statistical analysis and multiple hypotheses testing significant were used to remain significant metabolites, and then, least absolute shrinkage and selection operator (LASSO) logistic regression was performed to do the classification and LASSO linear regression was performed to do the log(UACR) value prediction. Results: In LASSO logistic regression, 3 metabolites were selected as the classification of normoalbuminuria and microalbuminuria (AUC=0.938), 11 metabolites were selected as the classification of normoalbuminuria and macroalbuminuria (AUC=0.966), 6 metabolites were selected as the classification of microalbuminuria and macroalbuminuria (AUC=0.963). In LASSO linear regression, the correlation coefficient of logarithmic UACR and predict value was 0.673. Conclusion: This study suggested that statistical analysis and multiple hypotheses testing result will enhance the performance of LASSO logistic regression, compared to result without statistical analysis. We revealed 3 metabolite combinations between different stages of diabetic nephropathy, those biomarkers may be related to inhibition of SIRT1 protein or could be medicine.