Machine learning models to predict sweetness of molecules.

Area of Science:

• Computational chemistry
• Machine learning
• Food science

Background:

• Sweetness is a crucial taste influencing human preference.
• The rise of type-2 diabetes necessitates the identification of low-calorie sweeteners.
• Predictive models for molecular sweetness are essential for developing healthier alternatives.

Purpose of the Study:

• To develop accurate computational models for predicting the sweetness of small molecules.
• To identify potential low-calorie sweeteners using machine learning.
• To create a comprehensive dataset and accessible web server for sweetness prediction.

Main Methods:

• Manual curation of an extensive dataset (SweetpredDB) comprising 671 sweet molecules with experimental sweetness values.
• Application of regression-based machine learning algorithms, including Gradient Boost and Random Forest Regressors.
• Development of a user-friendly web server (Sweetpred) for predicting sweetness.

Main Results:

• Gradient Boost and Random Forest Regressors demonstrated superior performance with correlation coefficients of 0.94 and 0.92, respectively.
• The developed models exhibit state-of-the-art accuracy compared to previous studies.
• The SweetpredDB dataset and Sweetpred web server are made publicly available.

Conclusions:

• Machine learning models can accurately predict molecular sweetness.
• The developed tools and dataset facilitate the discovery of novel, low-calorie sweeteners.
• This research contributes to managing diabetes and promoting healthier dietary choices.