BitterSweetForest: A Random Forest Based Binary Classifier to Predict Bitterness and Sweetness of Chemical Compounds

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Taste perception is crucial for nutrient intake and poison avoidance.
• Genetic and evolutionary factors significantly influence taste perception.
• Predicting molecular taste properties is vital for food science and toxicology.

Purpose of the Study:

• To develop and validate a machine learning model for discriminating sweet and bitter tastes of molecules.
• To create the first open-access model, BitterSweetForest, utilizing KNIME workflow for taste prediction.
• To apply the model to diverse chemical datasets, including natural products, drugs, and toxic compounds.

Main Methods:

• Development of a machine learning model based on molecular fingerprints.
• Utilizing a Random Forest classifier within a KNIME workflow.
• Validation through cross-validation and an independent test set.
• Application of Bayesian-based feature analysis for chemical feature discrimination.

Main Results:

• BitterSweetForest achieved 95% accuracy and 0.98 AUC in cross-validation.
• Independent testing showed 96% accuracy and 0.98 AUC for taste prediction.
• The model predicted a significant portion of natural products and approved drugs as bitter.
• Toxic compounds were predominantly predicted as bitter, supporting the bitter-as-poison hypothesis.

Conclusions:

• BitterSweetForest is a reliable tool for predicting molecular taste (sweet/bitter).
• The model demonstrates the link between bitter taste and toxicity.
• This platform offers valuable insights into taste perception and chemical safety.