Benchmarking Outlier Detection Methods for Detecting IEM Patients in Untargeted Metabolomics Data.

Area of Science:

• Biochemistry
• Computational Biology
• Clinical Diagnostics

Background:

• Untargeted metabolomics (UM) is a powerful tool for identifying metabolic disorders.
• Screening for inborn errors of metabolism (IEM) is crucial for early diagnosis and treatment.
• Existing outlier detection methods need evaluation for their efficacy in IEM patient profiling.

Purpose of the Study:

• To assess the performance of various outlier detection methods for identifying IEM patient profiles in UM data.
• To compare the effectiveness of 30 different outlier detection algorithms across multiple UM datasets.
• To determine the potential of these methods to aid in routine IEM screening.

Main Methods:

• Benchmarking 30 outlier detection methods on three untargeted metabolomics datasets.
• Evaluating method performance based on IEM patient detection rates and false positive counts.
• Investigating the impact of Principal Component Analysis (PCA) transformation prior to outlier detection.

Main Results:

• Significant variability in IEM detection performance was observed among the tested methods.
• DeepSVDD and R-graph demonstrated consistent performance across datasets.
• PCA transformation generally improved the performance of several outlier detection methods.
• Clinically relevant performance (90% detection, 0% false positives) was achieved by some methods on one dataset.

Conclusions:

• Outlier detection methods show potential to assist in routine IEM screening using UM data.
• Method selection and data preprocessing (e.g., PCA) are critical for optimal performance.
• Further advancements are necessary to achieve consistently clinically satisfying results for IEM detection.