Artificial Intelligence in the Identification of Key Metabolites Associated with Diabetic Kidney Disease

초록

Background: Despite the increasing prevalence of DKD, reliable biomarkers for its early detection remain scarce. This study aimed to identify urine and serum metabolites that differentiate between DKD stages through bioinformatics analysis. Methods: We analyzed 92 participants, categorizing them by eGFR and ACR combinations to identify criteria that best distinguish metabolites. Using these criteria, we trained and validated separate AI models. The criterion with the highest accuracy was selected for further investigation of metabolites that varied under this criterion. To understand the relationships and functions of these metabolites, we conducted a metabolic network analysis. Results: The classification based on ACR achieved the highest prediction accuracy. We identified nine urine and thirteen serum metabolites common among the top 20 from each group. Four metabolites—Adenosine and 5-MTA in urine, and m2,2G and cis-Aconitic acid in serum—demonstrated significant differences across ACR groups (Fig 1). Metabolic network analysis revealed hub proteins and networks linking these metabolites, with significant mRNA expression differences in hub proteins between healthy controls and those with DKD in both urine and serum networks (Fig 2). Notably, IL4I1 mRNA was identified in both urine and serum. Conclusions: The ACR-based classification demonstrated the highest accuracy. Urinary adenosine and 5-MTA, and serum m2,2G and cis-aconitic acid showed distinct patterns across ACR groups, highlighting their potential as biomarkers. Metabolic network analysis revealed hub proteins, including IL4I1, and networks connecting these differentially expressed metabolites. Further investigation into IL4I1’s involvement in DKD is recommended.