1449-P: Machine Learning Reveals Metabolic Signatures in Patients with Type 1 Diabetes

Background: Metabolic dysbiosis has been linked to the development of type 1 diabetes. However, there are few studies reflecting the metabolic signatures in patients with type 1 diabetes based on machine learning. Therefore, we aim to investigate the serum metabolic alterations and signatures in patients with type 1 diabetes.

Methods: We recruited 29 type 1 diabetes patients and 29 healthy controls with matching age, sex and ethnicity. Serum metabolites were analyzed using liquid chromatograph-mass spectrometry (LC-MS). Four machine-learning approaches (Logistic Regression, Support Vector Machine, Gaussian Naive Bayes, and Random Forest) were used to screen potential T1D-related biomarkers.

Results: Between the T1D group and the control group, 150 different metabolites were identified. These metabolites were significantly enriched in three metabolic pathways (P<0.05), purine metabolism, ketone body synthesis and degradation, and methyl butyrate metabolism. We combined comprehensive analysis and four machine-learning algorithmsto screen for biomarkers associated with type 1 diabetes. The top 20 metabolites coincided with 150 metabolites identified by LC-MS. In addition, three metabolites were identified as T1D diagnostic biomarkers, including docosahexaenoic acid, estrane and inosine. The combination of these metabolites also accurately diagnosis type 1 diabetes.

Conclusions: In this study, purine metabolism, synthesis and degradation of ketone bodies, and impaired methyl butyrate metabolism were identified as metabolic disturbance associated with type 1 diabetes. These biomarkers by using machine-learning approaches may have important clinical significance in the diagnosis of type 1 diabetes and in the follow of responses to therapeutic interventions.