McGPCR: A Multimodal Learning Model with Improved Applicability Domain Characterization for Predicting G Protein-Coupled Receptor Affinity of Plastic Chemicals

A variety of chemicals in plastics may pose risks to human health, while only a limited number have been extensively studied for their toxicity. Binding to G protein-coupled receptors (GPCRs) serves as a crucial molecular initiating event in identifying chemicals that induce toxic effects in humans. Given the diversity of GPCRs and chemicals, the binding affinity remains largely elusive, necessitating high-throughput models with the functionality of integrating chemical and receptor features to enable predictions across multiple receptors. Herein, a human GPCR affinity data set was constructed, containing 96,776 records between 59,599 compounds and 109 GPCRs. A multimodal learning model, McGPCR, was built to predict the GPCR binding affinity of chemicals by integrating multimodal features of molecular graphs and receptor binding sites. The McGPCR outperformed models with chemical structures as the only predictor variables. Applicability domain (AD) characterization based on feature-activity landscape analysis was proposed, which ensures the reliability of predictions. The McGPCR, along with the AD, was employed to predict affinities of over 9000 plastic chemicals. By integration of the affinity, persistence, bioaccumulation, and production volume, 30 plastic chemicals with potentially high environmental risks were identified. The McGPCR with AD characterization can serve as a powerful tool for identifying toxic chemicals harmful to human health.