Evaluation of deep learning architectures for predicting ligand interactions with neurologically relevant GPCRs.
TL;DR
It is revealed that while models achieve strong performance on structurally novel ligands, generalization to novel GPCRs remains challenging, with particularly poor prediction for out-of-family receptors.
Abstract
G protein-coupled receptors (GPCRs) are therapeutic targets for over 30% of approved drugs, yet specific GPCR subtypes act as molecular initiating events in several neurotoxic adverse outcome pathways. Therefore, knowledge of GPCR-ligand interactions is critical for drug discovery and computational toxicology. However, accurate predictions of GPCR-ligand binding can be challenging due to receptor conformational flexibility, complex membrane environment, and lack of selectivity among ligands. Drug-target interaction (DTI) models that jointly encode protein and ligand representations offer a promising approach to predict these interactions. In this study, we identified 119 neurologically relevant GPCRs and evaluated three deep learning architectures for creating a unified GPCR-ligand DTI model: dual-projection cosine similarity networks and transformer encoders (both using pre-computed embeddings) as well as bidirectional cross-attention networks (with frozen or fine-tuned encoders). Unlike prior studies that rely primarily on random splits, we evaluated model performance across random-split, cluster-split, and novel protein scenarios to provide realistic estimates of generalization. All the models performed well, achieving area under the receiver operating characteristic curve (AUROC) values of 0.91 on random-split validation, 0.81 on cluster-split testing (structurally distinct ligands), and 0.64 on novel GPCR generalization (proteins unseen during training). However, in the unseen-proteins test set, predictions were less accurate for GPCRs from protein families not represented in the training data or those with contrasting ligand interactions. We also assessed the models' ability to detect ligand selectivity, achieving an AUROC of 0.71 on ligands with at least five known GPCR interactions, although sensitivity remained low at 0.49. Our curated neurological GPCR dataset and rigorous evaluation framework provide realistic model assessment, reveal where current models succeed or fail, and provide practical guidance for deploying GPCR-ligand predictors in drug discovery and toxicity screening.Scientific contributionOur work connects computational toxicology and GPCR pharmacology by systematically evaluating transformers and similarity-based proteochemometric architectures for predicting CNS-relevant GPCR-ligand interactions under rigorous cold-start conditions. Our analysis reveals that while models achieve strong performance on structurally novel ligands, generalization to novel GPCRs remains challenging, with particularly poor prediction for out-of-family receptors. These results provide a rigorous evaluation framework for predicting ligand interactions against neurologically relevant GPCRs and highlight the need for improved protein representation learning in computational toxicology frameworks.