Advancing the Prediction of MS/MS Spectra Using Machine Learning.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Biochemistry

Background:

• Tandem mass spectrometry (MS/MS) is crucial for identifying small molecules and metabolites.
• Current identification relies on matching experimental spectra to reference libraries, which have limited coverage.
• In silico spectral prediction methods are being developed to expand spectral libraries.

Purpose of the Study:

• To investigate the challenges in achieving fast and accurate in silico MS/MS spectral predictions for small molecules.
• To address limitations of generic machine learning benchmarking in evaluating spectral prediction algorithms.
• To propose strategies for enhancing the reliability of computational spectral prediction.

Main Methods:

• Review and analysis of machine learning and deep learning approaches for MS/MS spectral prediction.
• Evaluation of common benchmarking practices and their impact on reported accuracy.
• Identification of key factors influencing prediction performance.

Main Results:

• Generic benchmarking tactics can lead to misleadingly high accuracy scores for spectral prediction models.
• Prediction accuracy is significantly influenced by dataset curation and the selection of appropriate collision energies.
• Current algorithms face amplified challenges when predicting spectra for a wide range of small molecules.

Conclusions:

• Improving in silico MS/MS spectral prediction requires careful data curation and consideration of experimental parameters like collision energy.
• Closer collaboration between computational scientists and experimental mass spectrometrists is essential.
• Refined methodologies are needed to overcome current limitations and achieve truly accurate spectral predictions for broader applications.