Deep kernel learning improves molecular fingerprint prediction from tandem mass spectra

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Untargeted metabolomics requires comprehensive spectral libraries for accurate structure annotation, which are currently incomplete.
• In silico methods, particularly machine learning models predicting molecular fingerprints from tandem mass spectra, are crucial for overcoming library limitations.
• Kernel support vector machines are effective for fingerprint prediction but struggle with large datasets due to cubic scaling in training time.

Purpose of the Study:

• To develop and evaluate a novel deep kernel learning method for molecular fingerprint prediction.
• To address the scalability limitations of traditional kernel support vector machines in handling large spectral datasets.
• To improve the accuracy and efficiency of in silico structure annotation in metabolomics.

Main Methods:

• Utilized the Nyström approximation to transform kernel methods into a linear feature map.
• Implemented and evaluated a linear support vector machine and a deep neural network (DNN) using this feature map.
• Assessed performance on a large cross-validated dataset (156,017 compounds) and three independent datasets (1734 compounds).

Main Results:

• The deep kernel learning approach, combining kernel methods with DNNs, demonstrated superior performance compared to standard kernel support vector machines.
• The proposed method outperformed a standalone DNN model directly applied to tandem mass spectra across all evaluation datasets.
• Achieved state-of-the-art results in molecular fingerprint prediction, enhancing its utility for metabolomics.

Conclusions:

• Deep kernel learning offers a scalable and highly accurate solution for molecular fingerprint prediction from tandem mass spectra.
• This advancement significantly improves the capabilities of in silico methods in metabolomics, aiding structure annotation and compound identification.
• The developed method is integrated into the SIRIUS software, making it accessible for broader research applications.