Deep learning-based prediction of drug bitterness using molecular structures

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Poor medication adherence, especially in children, is often linked to drug bitterness.
• Existing machine learning models for bitterness prediction have limitations in validation and scope.

Purpose of the Study:

• To develop a multitask graph neural network (GNN) for predicting drug bitterness directly from molecular structures.
• To improve the accuracy and reliability of bitterness prediction using sweetness as an auxiliary task.

Main Methods:

• Utilized SMILES-derived molecular graphs as inputs for a multitask GNN.
• Employed scaffold-based five-fold cross-validation and an independent external dataset of pediatric oral drugs for performance evaluation.
• Compared the multitask GNN against a single-task GNN.

Main Results:

• The multitask GNN consistently outperformed the single-task GNN across all validation folds, showing higher ROC-AUC and F1-scores.
• External validation on 185 pediatric drug compounds demonstrated high accuracy, correctly identifying 139 of 157 bitter compounds (sensitivity: 0.8854, precision: 0.9026, F1-score: 0.8939).
• Taste class overlap analysis indicated that bitterness is not solely determined by simple structural similarity.

Conclusions:

• Multitask molecular representation learning significantly enhances drug bitterness prediction.
• The developed GNN serves as a sensitive in silico screening tool for early-stage pharmaceutical development.
• Improved bitterness prediction can aid formulation design and enhance patient medication adherence.