Bitter or Not? BitterPredict- Tool for Predicting Taste from Chemical Structure

Bitter taste is a significant factor in animal’s choice of food. Animals avoid eating bitter food components, many of which are toxic. Nevertheless, it is known today that bitterness is not always noxious and that some of the bitter compounds have beneficial effects on health. Interestingly bitter taste receptors are also expressed in many extraoral tissues and emerge as novel targets for therapeutic indications such as asthma and infection.

Bitter compounds (gathered in the BitterDB http://bitterdb.agri.huji.ac.il/dbbitter.php) dramatically vary in their structures. Therefore, identifying bitter molecules based on their chemical structures is a very challenging task.

Here we present a machine learning classifier, BitterPredict, which predicts whether a molecule is bitter or not, based solely on its chemical structure. To this end we used: BitterDB as the true positives set, non-bitter molecules that were gathered from literature and enriched by random molecules as true negative set, physicochemical and ADME/TOX descriptors for the molecules, and Adaboost (decision tree based) algorithm.

BitterPredict correctly classifies over 85% of the compounds in the hold-out test set, and between 70% to 90% of the compounds in three independent external sets. The fraction of sp³-hybridized (tetrahedral) carbon atoms out of total carbon count (Fsp³) and Hydrophobic component of the saturated carbon and attached hydrogen (FOSA) descriptors are the most important contributors to the classifier.

Interestingly, but not surprisingly, the classifier suggests that a small portion (10%) of compounds found in food, and a large portion (70%) of clinical and experimental drugs, are bitter.