BitterPredict

BitterPredict predicts the bitterness of chemical compounds from their chemical structures to support toxicology, pharmaceutical development, and natural product research.

Key Features:

• Machine Learning Approach: Employs an Adaptive Boosting (AdaBoost) algorithm using decision trees to improve predictive accuracy by focusing on correcting previous errors.
• Descriptor Utilization: Represents molecules using physicochemical and ADME/Tox (Absorption, Distribution, Metabolism, Excretion, and Toxicity) descriptors.
• High Predictive Accuracy: Reports over 80% accuracy on hold-out test sets and 70–90% classification success across three independent external datasets and sensory test validations.
• Extensive Database Integration: Uses BitterDB for positive (bitter) samples and compiles non-bitter molecules from literature for the negative set.

Scientific Applications:

• Toxicology and Safety Assessment: Predicting bitterness can help infer potential toxicity because many bitter compounds function as natural deterrents.
• Pharmaceutical Development: Reports that approximately 66% of clinical and experimental drugs are bitter, informing drug formulation and patient compliance strategies.
• Natural Product Research: Predicts that approximately 77% of natural products are bitter, supporting exploration and utilization in health-related applications.

Methodology:

Data collection uses BitterDB for positive samples and literature sources for negative samples; compounds are characterized by physicochemical and ADME/Tox descriptors; an AdaBoost classifier with decision trees is trained on these features and validated against independent datasets and sensory tests.