Bitter or not? BitterPredict, a tool for predicting taste from chemical structure

Area of Science:

• * Computational chemistry and cheminformatics.
• * Sensory science and taste receptor research.
• * Pharmacology and toxicology.

Background:

• * Bitter taste perception evolved as a defense mechanism against toxic substances.
• * While often aversive, some bitter compounds possess beneficial health properties.
• * The vast chemical diversity of bitter molecules presents challenges for prediction and discovery.

Purpose of the Study:

• * To develop a machine learning classifier for predicting bitterness based on chemical structure.
• * To create a reliable tool for identifying bitter compounds in large datasets.
• * To explore the prevalence of bitterness in various chemical classes.

Main Methods:

• * Development of BitterPredict, a machine learning classifier using Adaptive Boosting (AdaBoost) and decision trees.
• * Training data derived from BitterDB (bitter compounds) and literature (non-bitter compounds).
• * Molecule representation using physicochemical and ADME/Tox descriptors.

Main Results:

• * BitterPredict achieved over 80% accuracy on a held-out test set.
• * Validation on independent external sets and sensory tests yielded 70-90% accuracy.
• * The model predicts a significant proportion of drugs and natural products are bitter.

Conclusions:

• * BitterPredict offers a rapid and dependable method for classifying compounds as bitter or non-bitter.
• * The tool aids in identifying novel bitter compounds with potential applications.
• * Findings suggest a high prevalence of bitterness in pharmaceuticals and natural products.