SF-Matching

SF-Matching predicts metabolite structures from tandem mass spectrometry (MS/MS) spectra by matching substructure-associated fragmentation patterns using machine-learning random forest models.

Key Features:

• SubFragment-Matching Hypothesis: Assumes molecules sharing substructures produce analogous fragmentation patterns that can be exploited for identification from MS/MS spectra.
• Integration of Fragmentation Patterns and Substructures: Analyzes fragmentation patterns linked to shared molecular substructures to estimate the likelihood that a candidate structure generates observed fragments.
• Machine Learning Model - Random Forests: Employs random forest models to score candidate molecules based on their predicted fragmentation behavior.
• Pre-computation of Scores: Pre-computes scores across common biological molecular structure databases to enable rapid comparison of experimental spectra to candidate structures.
• Benchmarking and Performance: Benchmarked on benchmark datasets showing performance comparable to CSI: FingerID and complementary accuracy gains when combined with CSI: FingerID.
• Scalability and Data-Driven Improvement: Rarefaction analysis of the training dataset indicates performance improves as additional experimental MS/MS data are incorporated.

Scientific Applications:

• Untargeted metabolomics: Supports identification workflows for untargeted MS/MS data by predicting compound identities from complex spectra.
• Metabolic pathway elucidation: Aids elucidation of metabolic pathways via structural annotation of detected metabolites.
• Biomarker discovery: Facilitates discovery of novel biomarkers through structure-level annotation of MS/MS features.
• Biochemical process investigation: Enables studies of biochemical processes by providing compound-level identifications from MS/MS data.

Methodology:

Matches substructure-associated fragmentation patterns, analyzes fragmentation patterns and substructures, uses random forest models to assess and score candidate molecules, pre-computes scores across common biological molecular structure databases, and performs rarefaction analysis on the training dataset.