e-Bitter: Bitterant Prediction by the Consensus Voting From the Machine-Learning Methods.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Experimental screening of bitter compounds is costly and time-consuming.
• Existing in-silico prediction models often rely on hypothetical data.
• A robust dataset of experimentally verified bitterants and non-bitterants is needed.

Purpose of the Study:

• To develop accurate computational models for predicting bitter compounds.
• To create a user-friendly software tool for bitterant identification.
• To establish a reliable benchmark dataset for in-silico bitterant prediction.

Main Methods:

• Compiled an experimental dataset of 707 bitterants and 592 non-bitterants.
• Employed consensus machine learning models combined with molecular fingerprints.
• Validated models using five-fold cross-validation, Y-randomization, and applicability domain analysis.

Main Results:

• Achieved high performance metrics including 0.929 accuracy, 0.954 sensitivity, and 0.856 MCC.
• Developed a standalone graphical software named "e-Bitter" for automated prediction.
• Demonstrated the efficacy of consensus modeling for bitterant classification.

Conclusions:

• The developed consensus model provides a reliable in-silico method for bitterant prediction.
• "e-Bitter" software offers a valuable, free resource for food scientists.
• This work sets a new standard for bitterant prediction using experimental data.