A Five-Plasma Protein–Based Algorithm for Predicting Incident CKD in Type 2 Diabetes

KEY POINTS: Proteomics analyses consistently identified nine independent protein predictors for CKD in both Chinese and European participants with type 2 diabetes. A novel prediction model, "eGFR+five-protein panel predicting CKD," reliably predicted CKD with good performance (area under the receiver operating characteristic curve ≥0.80). The new "eGFR+five-protein panel predicting CKD" model exhibited superior performance over traditional clinical models. BACKGROUND: Currently, there is a lack of reliable biomarkers for noninvasive prediction of CKD in patients with diabetes. This study aimed to identify novel protein predictors and construct a prediction model for incident CKD in participants with type 2 diabetes applicable across different populations. METHODS: A targeted Olink plasma proteomics profiling analysis, involving 368 proteins, was conducted in a nested case-control study comprising 132 incident CKD cases and 132 non-CKD controls, matched for age, sex, duration of diabetes, and eGFR, recruited from a long-term prospective cohort of Chinese type 2 diabetes participants (median approximately 9-year follow-up). False discovery rate was applied for multiple testing corrections. A q -value <0.2 was considered statistically significant. Three machine-learning approaches (Boruta, support vector machine, and eXtreme gradient boosting) were used for feature selection. Independent associations of proteins with incident CKD with adjustments for conventional clinical risk factors were examined in the training set ( n =1580, including 173 incident CKD cases and 1407 non-CKD controls; 70%) of the UK Biobank Pharma Proteomics Project (UKB-PPP) (median approximately 13.5-year follow-up). The least absolute shrinkage and selection operator-based Cox-regression analysis was used to develop the prediction model, which was subsequently validated in the testing set ( n =677, including 85 incident CKD cases and 592 non-CKD controls; 30%) of UKB-PPP. RESULTS: Among the 18 identified protein features predictive of incident CKD, 12 showed significant associations with consistent direction of effects in both cohorts. A prediction model ("eGFR+five-protein panel predicting CKD [eGFR+FPPC]") combining eGFR and five proteins (α-1-microglobulin/bikunin precursor, matrix metallopeptidase 7, placental growth factor, TNF-related apoptosis-inducing ligand receptor 2, and kidney injury molecule-1) was constructed and validated in the testing and training sets of the UKB-PPP, respectively. The "eGFR+FPPC" model achieved superior predictive performance in both training (area under the receiver operating characteristic curve [95% confidence interval]: 0.82 [0.79 to 0.85]) and testing (area under the receiver operating characteristic curve [95% confidence interval]: 0.80 [0.74 to 0.86]) sets of UKB-PPP and yielded fewer indeterminate results compared with conventional clinical models. CONCLUSIONS: The "eGFR+FPPC" model performed better than conventional clinical models in predicting incident CKD in type 2 diabetes participants across different populations.