Machine-Learning Approach Reveals Microbiome, Metabolome, Lipidome, and Their Interaction in Type 1 Diabetes Mellitus

Background: An increasing amount of research shows that the gut microbiota and metabolites play a role in type 1 diabetes mellitus (T1D). We aimed to use machine learning to explore gut microbiota, serum metabolites, and lipids signatures in T1D individuals. Methods: We evaluated 137 individuals in a cross-sectional cohort that included 38 T1D patients, 38 healthy controls, and 61 T1D patients for validation. After clinical examination and biospecimen collection, we characterized the gut microbiome profile with 16S rRNA gene amplicon sequencing and analyzed serum metabolites and lipids with liquid chromatography-mass spectrometry. All molecular data were analyzed using a combination of univariate, multivariate, and machine-learning approaches (Logistic Regression, Support Vector Machine, Gaussian Naive Bayes, and Random Forest). Results: Machine-learning approaches using microbiota composition did not accurately predict T1D status (model accuracy=0.7555), while the accuracy of model using metabolite composition was 0.9333. Based on bacterial species-level composition, Ruminococcus torques, Anaerostipes, Veillonella, Erysipelotrichaceae UCG-003, Blautia, and Coprococcus were coincident microbes which all increased in T1D. Increased 3-hydroxybutyric acid and 9-oxo-ode (AUC=0.70 and 0.67) were meaningful coincident metabolites in T1D. PC(36:4e)(rep) was the most significant lipid (coefficient index=3.11e-9, increased in T1D). We confirmed the biological relevance of the microbiome, metabolome, and lipidome features in the validation group. Ruminococcus torques was positively associated with 3-hydroxybutyric acid (p<0.01). Conclusions: By using machine-learning algorithms and multi-omics, we demonstrated that T1D patients are associated with altered microbiota, metabolites, and lipidomic signatures or functions. Machine-learning approaches have potential clinical applications in T1D diagnostics and treatment. Disclosure H.Tan: None. J.Yan: None. S.Luo: None. J.Weng: None. X.Zheng: None. Y.Shi: None. T.Yue: None. D.Zheng: None. C.Wang: None. Z.Liu: None. D.Yang: None. Y.Ding: None. W.Xu: None. Funding National Natural Science Foundation of China (82100822); Anhui Provincial Natural Science Foundation (2008085MH248, 2008085MH278); Guangdong Basic and Applied Basic Research Foundation (2019A1515010979); National Key R&D Program (2017YFC1309600)