Retip

Retip: Machine Learning-Based Retention Time Prediction for Untargeted Metabolomics

Retip predicts retention time (RT) of small molecules in high-performance liquid chromatography (HPLC) coupled with mass spectrometry (MS) to improve compound annotation in untargeted metabolomics by integrating RT prediction with MS/MS matching.

Key Features:

• Machine Learning Integration: Implements random forest, Bayesian-regularized neural network, XGBoost, light gradient-boosting machine (LightGBM), and Keras for RT prediction.
• Algorithm Performance: Keras minimizes overfitting with mean absolute error of 0.78 minutes for hydrophilic interaction liquid chromatography (HILIC) and 0.57 minutes for reversed-phase liquid chromatography (RPLC).
• Training Datasets: Trained and validated on the Fiehn HILIC dataset (981 primary metabolites and biogenic amines) and the RIKEN PlaSMA database (852 secondary metabolites).
• Software Integration: Integrates with MS-DIAL and MS-FINDER to refine candidate structure annotation.

Scientific Applications:

• Compound Annotation: Reduces candidate isomer structures by 68% in mouse blood plasma analysis using MS-FINDER, improving identification rates in liquid chromatography–mass spectrometry workflows.
• Untargeted Metabolomics: Enhances identification of unknown peaks by combining RT prediction with MS/MS spectral matching.

Methodology:

Machine learning models were trained on experimentally derived RT data from HILIC and RPLC platforms using publicly available metabolite datasets. Model performance was evaluated using training, test, and validation sets, with mean absolute error as the primary metric to assess prediction accuracy and overfitting.