Variable importance analysis based on rank aggregation with applications in metabolomics for biomarker discovery

Area of Science:

• Metabolomics
• Bioinformatics
• Computational Biology

Background:

• Biomarker discovery in metabolomics is crucial for classification tasks.
• Existing variable importance methods yield inconsistent ranking results.
• The discrepancy between different ranking methods is often overlooked.

Purpose of the Study:

• To address the inconsistency in biomarker ranking from different methods.
• To introduce rank aggregation as a strategy for merging individual rankings.
• To evaluate the performance of rank aggregation for biomarker discovery.

Main Methods:

• Employed rank aggregation to merge variable rankings from multiple methods.
• Utilized nine distinct methods: three univariate filtering and six multivariate.
• Applied the strategy to two real-world metabolic datasets.

Main Results:

• Rank aggregation significantly improved predictive classification accuracy compared to using all variables.
• The aggregated rankings outperformed the penalized Least Absolute Shrinkage and Selection Operator (LASSO) method.
• Achieved higher prediction accuracy or a reduced, more interpretable set of selected variables.

Conclusions:

• Rank aggregation is an effective strategy for robust biomarker discovery in metabolomics.
• This approach enhances the reliability and interpretability of identified biomarkers.
• The method offers a superior alternative to single-method rankings and standard penalized methods.