Improved One-Class Modeling of High-Dimensional Metabolomics Data via Eigenvalue-Shrinkage

Area of Science:

• Metabolomics
• Chemometrics
• Statistical modeling

Background:

• One-class modeling detects abnormal metabolite profiles using reference data.
• Current methods like Mahalanobis distance (MD) and principal component analysis (PCA) have limitations in untargeted metabolomics.
• Reliable detection of outlying metabolite profiles is crucial for applications in food safety and disease diagnosis.

Purpose of the Study:

• To propose novel distance measures for one-class modeling in untargeted metabolomics.
• To evaluate the performance of these new measures against standard methods.
• To provide a robust approach for identifying abnormal metabolite profiles.

Main Methods:

• Developed five new distance measures combining Mahalanobis distance (MD) with eigenvalue-shrinkage estimators for covariance matrix estimation.
• Utilized a cross-validation procedure for setting critical limits in outlier detection.
• Conducted simulation studies to assess performance based on type I error and outlier detection power.

Main Results:

• The proposed distance measures demonstrated superior performance compared to standard principal component-based one-class models.
• The new method achieved a better type I error rate (reduced false positives).
• Improved power for detecting abnormal metabolite profiles was observed.

Conclusions:

• The novel distance measures offer a more reliable and powerful approach for one-class modeling in untargeted metabolomics.
• The method is effective for analyzing data from techniques like LC-MS and NMR.
• This approach enhances applications in food safety and rare disease diagnosis.