Machine Learning Approach for Analysis of Metabolomics Predicts Persistent Albuminuria in Patients with Sickle Cell Anemia

Background: Sickle cell anemia (SCA) is associated with an increased risk of chronic kidney disease (CKD), which presents in its early stages as persistent albuminuria (PA). Understanding metabolic alterations underlying PA may aid diagnosis, prognosis, and management. This study aimed to identify disease-specific metabolites and associated biological pathways linked to PA in SCA using plasma metabolomics.

Methods: Untargeted metabolomics analysis of 280 SCA patients was performed using liquid chromatography coupled to high-resolution mass spectrometry (LC-HRMS). PA was defined based on urine albumin-creatinine ratio (UACR) ≥30 mg/g at least twice over 6 months or UACR ≥100 mg/g, if only one value was available. Data were preprocessed by quantile normalization, log10 transformation and auto-scaling. Differentially expressed metabolites (DEMs) between PA and those without PA were extracted using the linear model framework of R′s limma package with adjusted p≤0.05 for multiple tests. Pathway analysis was conducted to identify biological pathways enriched in DEMs. To assess predictive power of DEMs and their ability to predict PA, we performed machine learning with an Xtreme Gradient Boosting (XGBoost) model, which builds sequential decision trees, where each tree aims to correct errors of previous trees. We applied hyperparameter tuning to optimize model performance. A grid search-based approach was utilized to identify top metabolites contributing to model predictions, and 5-fold cross-validation for model evaluation. Model performance was assessed with area under the receiver operating characteristics (AUROC), sensitivity and specificity and their standard deviations.

Results: Of 280 patients, 161 (57.5%) were female; 100 (35.7%) had PA. Metabolomic analysis revealed 52 differentially expressed metabolites in patients with PA vs. those without PA. Significant alterations were observed in several metabolic pathways, including tryptophan, nicotinate and nicotinamide, histidine, glycine and serine metabolism and bile acid biosynthesis. Among specific metabolites, hydroxypropionylcarnitine, C4-Carnitine, O-[(2E)-hexenedioyl] carnitine and Tetranor-12R-HETE, kynurenic acid, quinolinic acid, methylhistamine and phosphohydroxypyruvic acid differentiated SCA patients with PA from those without PA.

Cross-validation of the XGBoost model with the top 18 metabolites yielded an AUROC of 0.75 (±0.07), sensitivity of 0.81 (±0.05), and specificity of 0.55 (±0.14).

Conclusions: Tryptophan metabolism, particularly via the kynurenine pathway, is linked to inflammation and oxidative stress, key factors in CKD pathophysiology. Impaired NAD⁺ metabolism exacerbates oxidative stress and inflammation, impacting kidney function. Altered levels of methylhistamine reflect heightened immune activity and inflammation, contributing to kidney damage. Imbalances in glycine and serine metabolism are part of one-carbon metabolism with altered phosphohydroxypyruvic acid levels, an intermediate in serine biosynthesis, and suggest disruptions in serine and glycine metabolism in CKD. Bile acids exert inflammatory and fibrotic effects on the kidneys. Elevated carnitine derivatives indicate disruptions in fatty acid metabolism and mitochondrial oxidation, contributing to oxidative stress and kidney injury. This metabolic dysregulation can lead to reduced energy production and increased oxidative stress in renal cells, contributing to cellular injury and dysfunction. Collectively, these metabolites highlight the complex metabolic disruptions associated with PA in SCA. Alterations in these metabolic pathways may contribute to development and progression of CKD through inflammation, oxidative stress and disruptions in cellular metabolism. Understanding these relationships and mechanistic roles of these metabolites will enhance our ability to predict, monitor, and treat kidney disease in SCA.

XGBoost model achieved an optimal AUROC utilizing 18 metabolites, indicating good discriminatory power between patients with PA and those without PA. The model showed a high sensitivity, but a moderate specificity, suggesting that it effectively identifies patients with PA but may produce false positives. Future work should focus on model optimization to improve specificity while maintaining sensitivity and external validation in an independent dataset.

Derebail:UpToDate: Honoraria; Novartis: Consultancy; iCell Gene: Consultancy; Amgen: Consultancy. Desai:NMDP: Other: Study Monitor ; Chiesi: Honoraria; Novo Nordisk: Research Funding; Pfizer: Consultancy, Research Funding; Novartis: Research Funding. Ataga:GSK: Consultancy, Honoraria; Hillhurst Biopharmaceuticals: Consultancy, Honoraria; Novartis: Honoraria; Novo Nordisk: Research Funding; Sanofi: Consultancy; Vertex: Membership on an entity's Board of Directors or advisory committees; Vitalant: Membership on an entity's Board of Directors or advisory committees; Forma Therapeutics: Honoraria, Membership on an entity's Board of Directors or advisory committees, Research Funding; Agios Pharmaceuticals: Honoraria, Membership on an entity's Board of Directors or advisory committees.