Using Machine Learning to Predict Obesity Based on Genome-Wide and Epigenome-Wide Gene–Gene and Gene–Diet Interactions

Abstract

Obesity is associated with many chronic diseases that impair healthy aging and is governed by genetic, epigenetic, and environmental factors and their complex interactions. This study aimed to develop a model that predicts an individual’s risk of obesity by better characterizing these complex relations and interactions focusing on dietary factors. For this purpose, we conducted a combined genome-wide and epigenome-wide scan for body mass index (BMI) and up to three-way interactions among 402,793 single nucleotide polymorphisms (SNPs), 415,202 DNA methylation sites (DMSs), and 397 dietary and lifestyle factors using the generalized multifactor dimensionality reduction (GMDR) method. The training set consisted of 1,573 participants in exam 8 of the Framingham Offspring Study (FOS) cohort. After identifying genetic, epigenetic, and dietary factors that passed statistical significance, we applied machine learning (ML) algorithms to predict participants’ obesity status in the test set, taken as a subset of independent samples (n = 394) from the same cohort. The quality and accuracy of prediction models were evaluated using the area under the receiver operating characteristic curve (ROC-AUC). GMDR identified 213 SNPs, 530 DMSs, and 49 dietary and lifestyle factors as significant predictors of obesity. Comparing several ML algorithms, we found that the stochastic gradient boosting model provided the best prediction accuracy for obesity with an overall accuracy of 70%, with ROC-AUC of 0.72 in test set samples. Top predictors of the best-fit model were 21 SNPs, 230 DMSs in genes such as CPT1A, ABCG1, SLC7A11, RNF145, and SREBF1, and 26 dietary factors, including processed meat, diet soda, French fries, high-fat dairy, artificial sweeteners, alcohol intake, and specific nutrients and food components, such as calcium and flavonols. In conclusion, we developed an integrated approach with ML to predict obesity using omics and dietary data. This extends our knowledge of the drivers of obesity, which can inform precision nutrition strategies for the prevention and treatment of obesity.Clinical Trial Registration: [www.ClinicalTrials.gov], the Framingham Heart Study (FHS), [NCT00005121].

Holistic Integration of Omics Tools for Precision Nutrition in Health and Disease

“…The integration of genomics with other functional omics analyses may significantly improve the accuracy of host genetic data for explaining health outcomes [ 6 ]. For example, using multiple machine learning algorithms, best predictors of obesity status were identified, and included single-nucleotide polymorphisms (mapped to genes such as STXBP6 , BBX , PLXDC2 , PCDH15 , TPH2 , PCDH15 , CALN1 , FGF14 , LRRN1 , ACTBP2 , RBMXP1 , and ZNF32 ) together with differentially methylated sites (in proximity to CPT1A , ABCG1 , SLC7A11 , RNF145 , and SREBF1 genes) and interactions with dietary factors encompassing specific foods, micronutrients, and bioactive compounds [ 7 ]. Additionally, the consortium of single-nucleotide polymorphisms in genes related to obesity and cardiometabolic diseases, low adherence to the Mediterranean diet, and harboring specific urolithin metabotypes (as biomarkers of the gut microbiota), was able to predict obesity in childhood and adolescence [ 8 ].…”

The potential of DNA methylation as a biomarker for obesity and smoking

“…This is not surprising, because obesity is known to have a genetic background; therefore, incorporation of the polygenic score naturally adds to the prediction accuracy. The most recent and currently the most comprehensive study aimed at developing a predictor for obesity risk combined BMI‐associated SNPs and methylation CpG sites with dietary and lifestyle factors in machine‐learning algorithms [105]. The best‐performing model in the study had an overall accuracy of 70% in predicting current obesity by using 21 SNPs, 230 CpG sites, and 26 dietary factors—such as processed meat, high‐fat dairy, French fries, artificial sweeteners, and alcohol intake.…”

Disease classification for whole blood DNA methylation: meta-analysis, missing values imputation, and XAI

“…Classifying different human conditions - phenotypes or pathologies - using DNA methylation data from a single tissue is more difficult. Phenotype classification can question smoking or obesity status, although existing results suggest that such conditions may not be clearly reflected in DNA methylation [21, 31, 32]. Classification of cases and controls for certain diseases is also performed using DNA methylation data.…”