BitterMasS: Predicting Bitterness from Mass Spectra

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Metabolomics

Background:

• Bitter compounds are prevalent in natural products and pharmaceuticals.
• Current machine learning models predict bitterness based on chemical structure, but most metabolites lack assigned structures.
• A significant portion of the metabolome remains uncharacterized, limiting bitterness prediction.

Purpose of the Study:

• To develop a machine learning model, BitterMasS, for predicting bitterness directly from experimental mass spectra.
• To assess the performance of a spectrum-based bitterness prediction strategy compared to structure-based methods.
• To enable the identification of bitter compounds in complex mixtures without prior structural elucidation.

Main Methods:

• A Random Forest classifier (BitterMasS) was trained on 5414 experimental mass spectra of bitter and nonbitter compounds.
• The model's performance was evaluated using internal and external test sets, including newly acquired spectral data.
• Comparison of spectrum-bitterness prediction versus spectrum-structure-bitterness prediction strategies.

Main Results:

• BitterMasS achieved high precision (0.83) and recall (0.90) on an internal test set.
• External validation demonstrated good performance: 67% precision and 93% recall for literature spectra, and 58% accuracy and 99% recall for newly measured spectra.
• The spectrum-bitterness approach outperformed structure-based methods in effectiveness and compound coverage.

Conclusions:

• BitterMasS effectively predicts bitterness from mass spectra, overcoming limitations of structure-based prediction.
• This method allows for the identification of bitter compounds in the "dark" metabolome.
• Applications include metabolomics, comparative bitterness analysis, and monitoring bitterness over time.