Deep Learning-Based Prediction of Drug Bitterness Using Molecular Structures.

Bitterness is a major cause of poor medication adherence, particularly in pediatric patients. Although several machine learning models have been developed for bitterness prediction, most rely on single-task learning and have undergone limited validation using pharmaceutically relevant compounds. We developed a multitask graph neural network (GNN) that predicts drug bitterness directly from molecular structures using sweetness as an auxiliary task. SMILES-derived molecular graphs were used as model inputs. Predictive performance was evaluated using scaffold-based five-fold cross-validation and an independent external dataset of pediatric oral drugs after removing compounds overlapping with the model development dataset. Compared with a single-task GNN, the multitask model consistently improved predictive performance across all folds, yielding higher ROC-AUC values and F1-scores. UMAP and silhouette analyses revealed substantial overlap between taste classes, indicating that bitterness cannot be explained by simple structural similarity. For external validation, an independent dataset comprising 185 pediatric drug compounds was used. The model correctly identified 139 of 157 previously reported bitter compounds, achieving a sensitivity of 0.8854, a precision of 0.9026, an F1-score of 0.8939, and an average precision of 0.9339, exceeding the prevalence baseline (0.849). These findings demonstrate that multitask molecular representation learning improves bitterness prediction and suggest that the proposed model may serve as a high-sensitivity in silico screening tool for prioritizing potentially bitter compounds during early-stage pharmaceutical development, thereby supporting formulation design and improving patient adherence.