Explainable Prediction of Protein-Ligand Interaction in Bitter Taste using Classical Machine Learning and Graph Neural Networks

Bitter taste is one of the five basic tastes that humans and other animals can perceive and is often linked to toxins and harmful compounds that can cause poisoning or disease. It involves the detection of diverse chemical compounds, such as alkaloids, phenols, and peptides, by a family of G protein-coupled receptors, known as taste receptor type 2 (TAS2R) or bitter taste receptors. Interestingly, TAS2Rs participate in the regulation of glucose homeostasis, modulation of immune and inflammatory responses, and may have implications for various diseases, such as obesity, diabetes, asthma, and cancer due to their expression in various extra-oral tissues. Therefore, they might represent promising therapeutic targets for several pathologies. However, the available data on receptor-ligand associations is limited and incomplete, requiring a costly and time consuming in vitro screening of possible TAS2R ligands. For this reason, in silico methods to predict bitterant-TAS2R interactions could represent powerful tools to help in the selection of ligands and targets for experimental studies. These methods can also pave the way to the discovery of novel bitter compounds, suggest the design of new bitter blockers or enhancers, and elucidate the molecular mechanisms of bitter taste. Among these computational methods, Machine Learning (ML) is a branch of artificial intelligence that applies algorithms to large datasets to learn from patterns and make predictions, while Deep Learning (DL), is a branch of ML that operates with Neural Networks. In particular, Graph Neural Networks are well suited to represent molecules and their interactions, hence these models could be used to clarify the underlying relationships between chemical features and compound behavior. In this context, the aim of this work is to develop an explainable model to predict the interactions between any given bitter molecule and TAS2Rs both via traditional ML and DL. Together with the classification task, these models were built to be either self-explainable or explainable through custom methods. The results demonstrate that the classical ML model outperforms both state-of-the-art models and the DL model, although the latter offers great potential in terms of explainability.