Integrative machine learning approaches for predicting disease risk using multi-omics data from the UK Biobank

Area of Science:

• Biomedical Informatics
• Genomics
• Metabolomics

Background:

• Current disease risk prediction often relies on individual data types like genomics or demographics.
• An integrated multi-omics approach is needed for comprehensive disease risk assessment and stratification.

Approach:

• Trained and compared machine learning models (Lasso, MLP, XGBoost, AdaBoost) using integrated multi-omics data.
• Incorporated ROC-AUC scores for evaluating prediction accuracy across various diseases and feature sets.
• Utilized Cox proportional hazard models for survival analysis.

Key Points:

• Integrated multi-omics data significantly enhanced disease risk prediction for 8 diseases.
• Metabolomic data's contribution was marginal compared to demographic, genetic, and biomarker features.
• Metabolomics can substitute for standard biomarker panels when they are unavailable.

Conclusions:

• Multi-omics data integration is a powerful tool for improving disease risk identification and stratification.
• While metabolomics has a smaller impact than other omics, it provides a valuable alternative biomarker source.