Integrating Metabolomics Domain Knowledge with Explainable Machine Learning in Atherosclerotic Cardiovascular Disease...

Area of Science:

• Metabolomics
• Machine Learning
• Cardiovascular Disease Research

Background:

• Metabolomic data's high dimensionality and variability pose analytical challenges.
• Machine learning (ML) offers powerful tools for extracting insights from complex metabolomic datasets.
• Integrating metabolomics with explainable ML enhances model interpretability in atherosclerosis research.

Purpose of the Study:

• To identify key metabolites associated with atherosclerotic cardiovascular disease (ASCVD) using explainable ML and metabolomics domain knowledge.
• To refine predictive performance and interpretability of ML models for ASCVD detection.
• To evaluate the effectiveness of identified metabolites in distinguishing ASCVD cases from controls in external cohorts.

Main Methods:

• Utilized Partial Least Squares Discriminant Analysis (PLS-DA) for dimensionality reduction of metabolomic data.
• Employed eXtreme Gradient Boosting (XGBoost) for metabolite selection and ASCVD characterization.
• Integrated metabolomics domain knowledge (chemical categorization) with SHAP analyses for enhanced model interpretability.

Main Results:

• Identified 56 metabolites for ASCVD discrimination, with lipids, organic acids, and organic oxygen compounds being primary superclasses.
• XGBoost model with selected metabolites achieved an Area Under the Curve (AUC) of 0.75 in the training cohort.
• SHAP analysis highlighted cholesterol, 3-methylhistidine, and glucuronic acid as impactful features; external validation yielded an AUC of 0.93.

Conclusions:

• A combination of metabolites can effectively build classifiers for ASCVD.
• Integrating metabolite categorization with SHAP analysis improves model interpretability, revealing metabolite-specific contributions to ASCVD risk.
• This approach demonstrates the potential for improved diagnostic and prognostic tools in cardiovascular disease research.