Diabetic kidney disease (DKD) is a leading cause of end-stage renal disease (ESRD) worldwide. Early detection is critical for the risk stratification and early intervention of progressive DKD. Serum creatinine (sCr) and urine output are used to assess kidney function, but these markers are limited by their delayed changes following kidney pathology, and lacking of both sensitivity and accuracy. Hence, it is essential to illustrate potential diagnostic indicators to enhance the precise prediction of early DKD. A total of 194 Chinese individuals include 30 healthy participants (Stage 0) and 164 incidents with type 2 diabetes (T2D) spanning from DKD's Stage 1a to 4 were recruited and their serums were subjected for untargeted metabolomic analysis. Random forest (RF), a machine learning approach, together with univariate linear regression (ULR) and multivariate linear regression (MvLR) analysis were applied to characterize the features of untargeted metabolites of DKD patients and to identify candidate DKD biomarkers. Our results indicate that 2-(α-D-mannopyranosyl)-L-tryptophan (ADT), succinyladenosine (SAdo), pseudouridine and N,N,N-trimethyl-L-alanyl-L-proline betaine (L-L-TMAP) were associated with the development of DKD, in particular, the latter three that were significantly elevated in Stage 2-4 T2D incidents. Each of the four metabolites in combination with sCr achieves better performance than sCr alone with area under the receiver operating characteristic curve (AUC) of 0.81-0.91 in predicting DKD stages. An average of 3.9 years follow-up study of another cohort including 106 Stage 2-3 patients suggested that "urinary albumin-to-creatinine ratio (UACR) + ADT + SAdo" can be utilized for better prognosis evaluation of early DKD (average AUC = 0.9502) than UACR without sexual difference.
Keywords: biomarkers; diabetic kidney disease; random forest; untargeted metabolomics.