Machine Learning to Identify Molecular Markers for Metabolic Disease Development Using Mouse Models

BackgroundAging, Western diet (WD) intake, and bile acid (BA) receptor farnesoid X receptor (FXR) inactivation are risk factors for metabolic disease development including nonalcoholic fatty liver disease (NAFLD) and chronic inflammation-related health issues such as dementia. The progression of the metabolic disease can be escalated when those risks are combined. Inactivation of FXR is cancer prone in both humans and mice. The current study used omics data generated within the gut-liver axis to classify those risks using bioinformatics and machine learning approaches. MethodsDifferent ages (5, 10, and 15 months) of wild-type (WT) and FXR knockout (KO) male mice were fed with either a healthy control diet (CD) or a WD since weaning. Hepatic transcripts, liver, serum, and urine metabolites, hepatic bile acids (BAs), as well as gut microbiota were used for risk prediction. A linear support vector machine with K-fold cross-validation was used for classification and feature selection. ResultsIncreased urine sucrose alone achieved 91% accuracy in predicting WD intake. Hepatic lithocholic acid (LCA) and serum pyruvate had 100% and 95% accuracy, respectively to classify age. Association analyses showed hepatic LCA was positively associated with serum concentrations of acetone, a ketone body, and 1,3-dihydroxyacetone (DHA), but negatively correlated with serum pyruvate. Urine metabolites (decreased creatinine and taurine as well as increased succinate) or gut microbiota (increased Dorea, Dehalobacterium, and Oscillospira) could predict FXR functional status with greater than 90% accuracy. Integrated pathway analyses revealed that the predictors for diet and FXR expression were implicated in the central carbon metabolism in cancer. To assess the translational relevance, mouse hepatic transcripts were crosschecked with human NAFLD and hepatocellular carcinoma (HCC) datasets. WD-affected hepatic Cyp39a1 and Gramd1b expression were associated with human HCC and NAFLD, respectively. The metabolites and diseases interaction analyses uncovered that the identified features are implicated in human metabolic diseases, mental disorders, and cancer. ConclusionThe risk prediction using mouse models contributes to the identification of noninvasive biomarkers for early diagnosis of metabolic disease development.