Evaluating Methods for High-Dimensional Mediation in Metabolomics Data

Area of Science:

• Metabolomics
• Statistical Genetics
• Bioinformatics

Background:

• High-dimensional data in metabolomics presents challenges for mediation analysis.
• Existing methods like HIMA and HDMA require evaluation for accuracy and scalability.
• Understanding indirect effects is crucial for elucidating biological pathways.

Purpose of the Study:

• To compare the performance of high-dimensional mediation analysis methods (HIMA, HDMA) and the Meet-in-the-Middle (MITM) approach.
• To assess the accuracy in estimating total indirect effects (TIE) and component indirect effects (CIEs).
• To evaluate sensitivity and specificity across various simulation scenarios.

Main Methods:

• Simulated metabolomics data with varying sample sizes, mediator set sizes, and correlation structures.
• Evaluation of HIMA (Zheng et al.) and HDMA (Gao et al.) for estimating TIE and CIEs.
• Assessment of the Meet-in-the-Middle (MITM) approach and its performance metrics.

Main Results:

• HDMA provided the most accurate TIE estimates in independent metabolite scenarios; HIMA also showed reliable CIE estimation.
• MITM generally underestimated TIE; HIMA's TIE estimates improved with larger mediation effect sizes.
• In correlated settings, CIE estimation was infeasible, and all methods underestimated TIE; sensitivity decreased with smaller effects and sample sizes.

Conclusions:

• HIMA offers accurate mediation results but may reduce dimensionality, potentially excluding features.
• Applying parallel mediation approaches (MITM, HIMA) and focusing on overlapping results is recommended.
• There is a need for robust, scalable mediation methods specifically for untargeted metabolomics data.