Deep Learning-Based Molecular Fingerprint Prediction for Metabolite Annotation.

Area of Science:

• Computational chemistry
• Bioinformatics
• Analytical chemistry

Background:

• Metabolomics studies heavily rely on Liquid Chromatography-Mass Spectrometry (LC-MS).
• Metabolite annotation is a significant challenge due to limited public tandem mass spectrometry (MS/MS) spectral libraries.
• Deep learning offers a novel approach to overcome these limitations by modeling complex relationships between molecular data and spectral measurements.

Purpose of the Study:

• To investigate deep learning methods for molecular fingerprint-based metabolite identification using MS/MS spectra.
• To rank putative metabolite identifications by comparing known and predicted molecular fingerprints.
• To evaluate the performance of deep learning models against established methods on benchmark datasets.

Main Methods:

• Trained three deep learning models to correlate molecular fingerprints with MS/MS spectra.
• Processed MS/MS spectral data from NIST, MoNA, and HMDB databases, including scaling, binning, and filtering.
• Applied feature selection to identify relevant m/z bins and molecular fingerprints.
• Evaluated model performance on CASMI 2016, 2017, and 2022 benchmark datasets for metabolite identification.

Main Results:

• Feature selection effectively reduced data redundancy, improving model training.
• Deep learning models achieved performance comparable to CSI:FingerID in ranking metabolite candidates.
• The trained models demonstrated efficacy in predicting and ranking metabolites based on molecular fingerprints and MS/MS spectra.

Conclusions:

• Deep learning presents a viable and promising approach for enhancing metabolite annotation in metabolomics.
• The developed methods offer a powerful alternative to traditional spectral matching techniques.
• This study highlights the potential of AI in addressing critical bottlenecks in biological data analysis.