Application of data augmentation techniques towards metabolomics.

Area of Science:

• Biochemistry
• Genetics
• Computational Biology

Background:

• Niemann-Pick Class 1 (NPC1) disease is a rare, neurodegenerative lysosomal storage disorder.
• Metabolomics datasets for NPC1 patients are often small and unbalanced, hindering analysis.

Purpose of the Study:

• To enhance predictive capabilities and identify novel biomarkers for NPC1 disease using urinary metabolomics data.
• To evaluate data augmentation (DA) techniques for improving analysis of limited patient datasets.

Main Methods:

• Employed computational intelligence-based DA techniques: noise addition, oversampling, and conditional generative adversarial networks.
• Utilized machine learning classification models and partial least squares for prediction.
• Analyzed urine samples from 13 untreated NPC1 patients and 47 controls.

Main Results:

• DA techniques successfully generated high-quality synthetic data, significantly improving model performance.
• Achieved increases in sensitivity (20%-50%), F1 score (6%-30%), and predictive capacity (0.3).
• Identified urinary branched-chain amino acids (valine, 3-aminoisobutyrate, quinolinate) as potential diagnostic and prognostic biomarkers.

Conclusions:

• DA is effective for improving biomarker discovery in small, unbalanced metabolomics datasets for NPC1 disease.
• Urinary metabolite profiles, particularly branched-chain amino acids, show promise for NPC1 diagnosis and monitoring.