EVA

EVA applies a convolutional neural network (CNN) to evaluate metabolic feature fidelity in LC-MS data by classifying extracted ion chromatogram (EIC) peak shapes to reduce false positives in feature extraction.

Key Features:

• Deep learning integration: A CNN trained on 25,095 EIC plots derived from 22 diverse LC-MS-based metabolomics projects.
• Manual quality assessment for training: Training labels were obtained by manual inspection classifying each EIC plot as good or poor quality.
• High classification accuracy: Demonstrated over 90% accuracy in classifying EIC shapes.
• False-positive reduction: Classifies EIC peak shapes to identify low-fidelity features arising from limitations of traditional peak-picking algorithms.

Scientific Applications:

• Metabolomics feature curation: Improves fidelity of feature extraction for LC-MS metabolomics datasets.
• Biomarker discovery: Enhances reliability of candidate biomarkers by reducing false positives in feature lists.
• Metabolic pathway analysis: Provides higher-quality feature sets for metabolic pathway elucidation.
• Systems biology studies: Supports downstream systems biology analyses with more reproducible LC-MS feature data.

Methodology:

Training a convolutional neural network on 25,095 manually annotated EIC plots to learn patterns of good versus poor EIC peak shapes indicative of feature fidelity and potential false positives.