Abstract
Determining molecular taste remains a significant challenge in food science. Here, we present FART (Flavor Analysis and Recognition Transformer), a chemical language model capable of predicting molecular taste from chemical structure. Trained on the largest public dataset (15,025 compounds) of molecular tastants to date, FART is the first model capable of parallel predictions across four taste categories: sweet, bitter, sour, and umami. FART achieves an accuracy above 91% for parallel taste prediction and outperforms previous state-of-the-art binary classifier models that specialize on predicting one taste class. Its transformer architecture allows for interpretability through gradient-based visualization of molecular features. The model identifies key structural elements driving taste properties and demonstrates utility in analyzing known tastants as well as novel compounds. By releasing both the model and dataset, we equip the food science community with tools for rapid taste prediction, accelerating flavor compound development and enabling systematic exploration of taste chemistry.