Assessing individual genetic susceptibility to metabolic syndrome: interpretable machine learning method

Area of Science:

• Genetics
• Computational Biology
• Metabolic Diseases

Background:

• Genome-wide association studies (GWAS) have identified genetic factors for metabolic syndrome (MetS).
• Current methods lack machine learning (ML)-based models for assessing individual genetic susceptibility to MetS.
• This gap limits personalized risk prediction and prevention strategies for MetS.

Purpose of the Study:

• To develop and validate ML-based genetic risk score (GRS) models for predicting MetS occurrence.
• To assess individual genetic susceptibility to MetS using single-nucleotide polymorphisms (SNPs).
• To integrate GRS with conventional risk factors for improved MetS prediction.

Main Methods:

• Feature selection using Least Absolute Shrinkage and Selection Operator (LASSO).
• Construction of GRS models using six ML algorithms, including extreme gradient boosting (XGBoost).
• Internal validation via fivefold cross-validation and model performance evaluation using ROC curves; SHapley Additive exPlanations (SHAP) for model interpretation.

Main Results:

• The XGBoost model achieved superior discriminative performance (AUC = 0.837) for MetS prediction.
• SHAP analysis provided insights into SNP effects and interactions influencing MetS risk.
• A combined model integrating GRS and conventional risk factors demonstrated excellent performance (AUC = 0.962).

Conclusions:

• A reliable XGBoost-based GRS model and prediction platform were established for assessing individual MetS genetic susceptibility.
• The model offers high interpretability, supporting personalized MetS primary prevention.
• Integrating GRS with traditional risk factors enhances MetS prediction accuracy.