Abstract: It is important to identify the bitter substances in natural compounds and determine their bitterness threshold for finding out the bitter molecules that affect the flavor of food and developing some foods with unique flavors. Identifying bitter molecules and predicting the threshold of bitter molecules based on the quantitative structure-activity relationship is a low-cost and rapid method. This research used Molecular Operating Environment (MOE), Chemopy and Mordred to generate 2D molecular descriptor to establish bitterness molecular recognition models with Support Vector Machine (SVM) and Random Forests (RF) algorithms. This study used above descriptors to establish bitterness threshold prediction models with Partial Least Squares Regression (PLSR), Random Forests Regression (RFR), k-Nearest Neighbor Regression (kNNR), and Principle Component Regression (PCR) algorithms. The results showed that the MOE-RF model had the highest accuracy of 0.982, the ChemoPy-PLSR model had the best bitterness prediction effect with a coefficient of determination of 0.85 and a root mean square error of 0.43. The two models would be combined to predict whether the molecule has bitterness and the threshold of bitterness or not.