Decision Tree-Based Ensemble Method for Bitter Taste Prediction

In the field of chemistry, the problem of determining the bitter taste of a certain compound is considered a difficult task which dramatically affects the development in the fields of drug discovery and beverage science. Confidently, with the innovation in artificial intelligence (AI) numerous research endeavours have surfaced, aiming to tackle this intricate problem by using machine learning-based approaches to detect whether a molecule exhibits bitterness or not. In more detail, these research articles showed their excellent model performances in cross-validation and their test sets. Nevertheless, in terms of three external test sets (Phyto-Dictionary, UNIMI, Bitter new) which were established by the research of BitterPredict [1], however, these external evaluations reveal a stumbling block─namely, the intricate chemical diversity inherent in each test set. Consequently, existing models find themselves constrained, and they can only be used in a specific test set or a certain chemical domain. Against this backdrop, our contribution is to propose a novel method: a decision tree-based ensemble model, which uses the predicted outputs of our EC-ANN, M2V-ANN [23] and Chemopy-RF [2] models to classify a molecule as bitterant or not, the process of data augmentation is included and specified for a distinct task. The practical experiment has supposed that our ensemble model successfully increases the overall performance compared to SOTA models [1]–[4] in three test sets, which obtained 0.954, 0.846 and 0.8 in F1-score from the combination of our models and the SOTA models [2].