The Use of Machine Learning Algorithms and the Mass Spectrometry Lipidomic Profile of Serum for the Evaluation of Tacrolimus Exposure and Toxicity in Kidney Transplant Recipients

Abstract

Tacrolimus has a narrow therapeutic window; a whole-blood trough target concentration of between 5 and 8 ng/mL is considered a safe level for stable kidney transplant recipients. Tacrolimus serum levels must be closely monitored to obtain a balance between maximizing efficacy and minimizing dose-related toxic effects. Currently, there is no specific tacrolimus toxicity biomarker except a graft biopsy. Our study aimed to identify specific serum metabolites correlated with tacrolinemia levels using serum high-precision liquid chromatography–mass spectrometry and standard laboratory evaluation. Three machine learning algorithms were used (Naïve Bayes, logistic regression, and Random Forest) in 19 patients with high tacrolinemia (8 ng/mL) and 23 patients with low tacrolinemia (5 ng/mL). Using a selected panel of five lipid metabolites (phosphatidylserine, phosphatidylglycerol, phosphatidylethanolamine, arachidyl palmitoleate, and ceramide), Mg2+, and uric acid, all three machine learning algorithms yielded excellent classification accuracies between the two groups. The highest classification accuracy was obtained by Naïve Bayes, with an area under the curve of 0.799 and a classification accuracy of 0.756. Our results show that using our identified five lipid metabolites combined with Mg2+ and uric acid serum levels may provide a novel tool for diagnosing tacrolimus toxicity in kidney transplant recipients. Further validation with targeted MS and biopsy-proven TAC toxicity is needed.

Keywords

• tacrolimus
• kidney transplant
• metabolomic biomarkers
• nephrotoxicity
• machine learning
• kidney graft function