SteroidXtract

SteroidXtract extracts steroid molecules from untargeted metabolomics tandem mass spectrometry (MS²) data to identify and quantify known and unknown steroids.

Key Features:

• Deep-Learning Framework: Employs a convolutional neural network (CNN) model trained on MS² spectra from MassBank of North America (MoNA) and an in-house steroid library to recognize steroid molecules.
• MS² Pattern Recognition: Identifies steroid-specific tandem mass spectrometry (MS²) spectral patterns for detection and quantification of known and unknown steroids.
• Data Augmentation Strategies: Uses intensity thresholding and Gaussian noise addition to augment limited steroid MS² spectra and reduce overfitting.
• Comparative Performance Analysis: Evaluates the CNN against random forest and XGBoost using nested cross-validation, reporting improvements in sensitivity, specificity, and robustness.
• Metabolomics Prioritization: Prioritizes biologically significant steroid molecules in untargeted metabolomics datasets with high throughput and sensitivity.

Scientific Applications:

• Steroid detection and quantification: Detection and quantification of known and previously unreported steroid molecules from untargeted MS² metabolomics data.
• Metabolomics feature prioritization: Prioritization of biologically relevant steroid features for follow-up in metabolomics studies.
• Steroid-focused biological research: Support for investigations of steroids in physiological systems and disease contexts.

Methodology:

Training a convolutional neural network (CNN) on MS² spectra from MassBank of North America (MoNA) and an in-house steroid library, applying intensity thresholding and Gaussian noise addition for data augmentation, and performing nested cross-validation comparisons with random forest and XGBoost to evaluate sensitivity, specificity, and robustness while recognizing steroid-specific MS² spectral patterns.