Targeted metabolomics identified novel metabolites, predominantly phosphatidylcholines and docosahexaenoic acid-containing lipids, predictive of incident chronic kidney disease in middle-to-elderly-aged Chinese adults.

Background: Evidence is limited regarding the association of circulating metabolites with decline of kidney function, letting alone their value in prediction of development of chronic kidney disease (CKD).

Methods: This study included 3802 participants aged 64.1 ± 7.4 years from the Dongfeng-Tongji cohort, among whom 3327 were CKD-free at baseline (estimated glomerular filtration rate [eGFR] > 60 ml/min per 1.73 m). We measured baseline levels of 211 metabolites with liquid chromatography coupled with mass spectrometry, including 25 amino acids, 12 acyl-carnitines, 161 lipids, and 13 other metabolites.

Results: The mean (SD) absolute annual change in eGFR was -0.14 ± 4.11 ml/min per 1.73 m per year, and a total of 472 participants who were free of CKD at baseline developed incident CKD during follow-up of 4.6 ± 0.2 years (14.2 %). We identified a total of 22 metabolites associated with annual eGFR change and survived Bonferroni correction for multiple testing, including seven metabolites associated with eGFR increase (six being docosahexaenoic acid [DHA]-containing lipids) and 15 associated with eGFR decline (nine being phosphatidylcholines [PCs]). Among them, eight metabolites obtained non-zero coefficients in least absolute shrinkage and selection operator (LASSO) regression on incident CKD, indicating predictive potential, including one amino acid (arginine), one acyl-carnitine (C2), one lysophosphatidylcholine (LPC 22:6), two PCs (32:1 and 34:3), one triacylglycerol (TAG 56:8 [22:6]) and two other metabolites (inosine, niacinamide), and the composite score of these eight metabolites showed an odds ratio (OR) of 8.79 (95 % confidence interval [CI]: 7.49, 10.32; P < 0.001) per SD increase in association with incident CKD. The addition of the metabolite score increased the c-statistic of the reference model of traditional risk factors (including baseline eGFR) by 0.065 (95 % CI: 0.046 to 0.084; P = 3.39 × 10) to 0.765 (0.742 to 0.788) in 1000 repetitions of 10-fold cross-validation, while the application of two advanced machine learning algorithms, random forest (RF), and extreme gradient boosting (XGBoost) models produced similar c-statistics, to 0.753 (0.729 to 0.777) and 0.778 (0.733 to 0.824) with increases of 0.074 (0.055 to 0.093; P = 4.11 × 10) and 0.073 (0.032 to 0.114; P = 4.00 × 10), respectively.

Conclusions: In this study, we identified 22 metabolites associated with longitudinal eGFR change, nine of which were PCs and six were DHA-containing lipids. We screened out a panel of eight metabolites which improved prediction for the development of CKD by 9 % beyond traditional risk factors including baseline eGFR. Our findings highlighted involvement of lipid metabolism in kidney function impairment, and provided novel predictors for CKD risk.