Benchmarking MS/MS Featurization Strategies for Machine Learning-Driven Metabolite Structure Annotation

Area of Science:

• Metabolomics
• Computational Chemistry
• Bioinformatics

Background:

• Untargeted metabolomics experiments generate vast amounts of MS/MS spectra, often unannotated due to incomplete spectral libraries.
• Machine learning offers a path to annotate metabolites beyond direct library matches, but requires effective numerical representations of MS/MS spectra.
• Existing spectral featurization methods lack systematic comparison, hindering optimal metabolite annotation.

Purpose of the Study:

• To systematically benchmark various spectral featurization methods for MS/MS data.
• To evaluate the impact of vector dimensionality on performance.
• To assess the accuracy of different featurization strategies in retrieving correct metabolite structures from large databases.

Main Methods:

• Over 71,000 unique compounds with merged-energy MS/MS spectra were used.
• Benchmarked methods included fixed/adaptive binning, frequent-peaks, spectrum hashing, and learned embeddings (Spec2Vec, MS2DeepScore, DreaMS, SpecEmbedding).
• 105 neural network models were trained to predict Mol2Vec embeddings and retrieve structures, with retrieval assessed at 0.1, 3, and 10 ppm mass tolerances.

Main Results:

• Adaptive binning, frequent-peaks, and DreaMS demonstrated the most accurate embedding predictions.
• Top-1 retrieval accuracy reached 46% at 0.1 ppm, 44% at 3 ppm, and 38% at 10 ppm on the test dataset.
• Top-5 accuracies reached up to 77%, with performance varying significantly based on mass tolerance and featurization strategy.

Conclusions:

• Clear performance differences exist among spectral featurization strategies.
• Metabolite structure retrieval accuracy is strongly dependent on mass precision.
• There is a need for evaluation metrics that align with structure-level annotation tasks in metabolomics.