Predicting Molecular Taste: Multi-Label and Multi-Class Classification

Predicting the taste of chemical compounds is a complex task and has been a challenge for decades. This study explores the application of machine learning to predict taste profiles of chemical compounds using the ChemTastesDB dataset, comprising 2,944 tastants categorized into 44 taste labels and 9 taste classes. Addressing the challenges of label imbalance and correlation, the dataset was preprocessed using iterative stratified sampling and feature representations such as Mordred descriptors, Morgan fingerprints, and Daylight fingerprints. Baseline random forest models, along with binary relevance and classifier chains, were employed for multi-label classification, with evaluation metrics including micro-averaged F1 scores, precision, and recall. Results demonstrated that binary relevance models, particularly with Morgen fingerprints, achieved superior F1 scores, outperforming classifier chains likely due to random label ordering. Label correlation analysis via co-occurrence matrices and community detection revealed significant associations between taste labels, providing deeper insights into molecular taste interactions. Feature importance analysis highlighted structural elements influencing taste prediction. This work underscores the potential of computational models in advancing flavor science and paves the way for future exploration with deep learning and optimized label dependencies.