Building a model for predicting metabolic syndrome using artificial intelligence based on an investigation of whole-genome sequencing.

BACKGROUND

The circadian system is responsible for regulating various physiological activities and behaviors and has been gaining recognition. The circadian rhythm is adjusted in a 24-h cycle and has transcriptional-translational feedback loops. When the circadian rhythm is interrupted, affecting the expression of circadian genes, the phenotypes of diseases could amplify. For example, the importance of maintaining the internal temporal homeostasis conferred by the circadian system is revealed as mutations in genes coding for core components of the clock result in diseases. This study will investigate the association between circadian genes and metabolic syndromes in a Taiwanese population.

METHODS

We performed analysis using whole-genome sequencing, read vcf files and set target circadian genes to determine if there were variants on target genes. In this study, we have investigated genetic contribution of circadian-related diseases using population-based next generation whole genome sequencing. We also used significant SNPs to create a metabolic syndrome prediction model. Logistic regression, random forest, adaboost, and neural network were used to predict metabolic syndrome. In addition, we used random forest model variables importance matrix to select 40 more significant SNPs, which were subsequently incorporated to create new prediction models and to compare with previous models. The data was then utilized for training set and testing set using five-fold cross validation. Each model was evaluated with the following criteria: area under the receiver operating characteristics curve (AUC), precision, F1 score, and average precision (the area under the precision recall curve).

RESULTS

After searching significant variants, we used Chi-Square tests to find some variants. We found 186 significant SNPs, and four predicting models which used 186 SNPs (logistic regression, random forest, adaboost and neural network), AUC were 0.68, 0.8, 0.82, 0.81 respectively. The F1 scores were 0.412, 0.078, 0.295, 0.552, respectively. The other three models which used the 40 SNPs (logistic regression, adaboost and neural network), AUC were 0.82, 0.81, 0.81 respectively. The F1 scores were 0.584, 0.395, 0.574, respectively.

CONCLUSIONS

Circadian gene defect may also contribute to metabolic syndrome. Our study found several related genes and building a simple model to predict metabolic syndrome.