BitterSweet

BitterSweet predicts bitter and sweet taste profiles of small molecules using machine learning to identify molecular correlates that define the gradient between bitter and sweet tastes.

Key Features:

• Bitter–Sweet classification: Predicts bitter, sweet, or tasteless labels for small molecules.
• Machine learning models: Employs state-of-the-art machine learning models for taste-profile prediction.
• Dataset compilation: Compiles datasets of small molecules for model development and evaluation.
• Molecular descriptors: Utilizes a diverse array of molecular descriptors to capture chemical characteristics influencing taste perception.
• Descriptor evaluation: Evaluates multiple sets of molecular descriptors and compares their predictive performance.
• Feature-block analysis: Identifies key features and feature blocks that contribute to accurate taste prediction.
• Application to specialized chemical databases: Applies models to large chemical datasets including FlavorDB, FooDB, SuperSweet, Super Natural II, DSSTox, and DrugBank.

Scientific Applications:

• Flavor science: Predicts taste properties of flavor compounds to inform flavor formulation and analysis.
• Food chemistry: Screens natural and synthetic food-derived molecules for bitter or sweet taste profiles.
• Drug development: Assesses taste-related properties of drug candidates relevant to palatability and off-target taste effects.
• Chemosensory research: Investigates molecular correlates of taste perception and the bitter–sweet sensory axis.
• Large-scale compound screening: Enables screening of extensive chemical libraries for desired taste attributes.

Methodology:

Compile datasets of small molecules; compute diverse molecular descriptors; train state-of-the-art machine learning models; evaluate multiple descriptor sets and compare predictive performance to identify important features and feature blocks; apply models to datasets including FlavorDB, FooDB, SuperSweet, Super Natural II, DSSTox, and DrugBank.