Feature Selection Pipelines with Classification for Non-targeted Metabolomics Combining the Neural Network and...

Area of Science:

• Metabolomics
• Computational Biology
• Biomedical Data Analysis

Background:

• Non-targeted metabolomics generates vast datasets with numerous features but limited samples.
• Analyzing such high-dimensional data requires advanced methods beyond classical statistics.

Purpose of the Study:

• To develop a novel computational pipeline for analyzing high-resolution mass spectrometry metabolomics data.
• To identify a metabolic fingerprint associated with prediabetes using machine learning.

Main Methods:

• A pipeline combining convolutional neural networks (CNNs) with statistical approaches and a genetic algorithm was developed.
• The method was applied to metabolomics data from a lifestyle intervention cohort undergoing oral glucose tolerance tests.
• Feature selection was performed using CNN classification and a genetic algorithm for relevance ranking.

Main Results:

• The pipeline achieved a precision-recall score exceeding 0.9 on the test set for classification.
• Approximately 200 features with high predictive scores were identified, characterizing metabolic changes in prediabetic individuals.
• The approach successfully identified a metabolic fingerprint for the prediabetic class.

Conclusions:

• The developed framework offers a novel approach for complex modeling of high-resolution mass spectrometry data using CNNs.
• This method enhances the analysis of non-targeted metabolomics data, aiding in the identification of disease-related metabolic signatures.
• The pipeline provides a powerful tool for understanding metabolic alterations in conditions like prediabetes.