An Ensemble Spectral Prediction (ESP) model for metabolite annotation.

Area of Science:

• Metabolomics
• Computational Chemistry
• Bioinformatics

Background:

• Metabolomics research faces a significant challenge in annotating measured spectra with accurate chemical identities, with only a small fraction of spectra currently identifiable.
• Existing computational approaches for metabolite annotation involve mapping candidate molecules to spectra or query spectra to molecular candidates, with the best matching spectrum determining the target molecule.
• Limited research has focused on incorporating rank learning tasks to improve the accuracy of determining the target molecule in spectral annotation.

Purpose of the Study:

• To develop a novel machine learning model, Ensemble Spectral Prediction (ESP), to enhance metabolite annotation accuracy in metabolomics.
• To improve the process of assigning chemical identities to measured spectra from biological samples.
• To leverage and combine existing neural network-based annotation models for superior performance.

Main Methods:

• Proposed a novel machine learning model, Ensemble Spectral Prediction (ESP), integrating multilayer perceptron (MLP) and Graph Neural Network (GNN) models.
• ESP learns optimal weightings for MLP and GNN spectral predictor outputs to generate a refined spectral prediction for query molecules.
• Enhanced baseline MLP and GNN models by incorporating peak dependencies via label mixing and multi-tasking on spectral topic distributions, with training data stratified by molecular formula.

Main Results:

• The ESP model demonstrated significant performance gains, improving average rank by 23.7% over MLP baselines and 37.2% over GNN baselines when evaluated on the NIST 2020 dataset and PubChem candidate sets.
• ESP achieved performance improvements over state-of-the-art neural network approaches for metabolite annotation.
• Analysis indicated that annotation performance is influenced by the training dataset, the size of the candidate set, and the similarity between candidate molecules and the target molecule.

Conclusions:

• The Ensemble Spectral Prediction (ESP) model offers a substantial advancement in metabolite annotation accuracy within metabolomics.
• The integration of MLP and GNN models, coupled with rank learning, provides a more robust approach to spectral identification.
• The developed ESP tool, including code and a trained model, is publicly available to facilitate broader adoption and further research in the field.