Evaluation of deep learning architectures for predicting ligand interactions with neurologically relevant GPCRs

Area of Science:

• Computational toxicology
• Pharmacology
• Drug discovery

Background:

• G protein-coupled receptors (GPCRs) are crucial drug targets, but their role in neurotoxicity necessitates understanding GPCR-ligand interactions.
• Predicting these interactions is challenging due to receptor flexibility and environmental factors.
• Drug-target interaction (DTI) models offer a promising computational approach.

Purpose of the Study:

• To evaluate deep learning architectures for unified GPCR-ligand DTI modeling.
• To assess model generalization using realistic validation strategies (random, cluster, and novel protein splits).
• To provide practical guidance for deploying GPCR-ligand predictors in drug discovery and toxicity screening.

Main Methods:

• Identified 119 neurologically relevant GPCRs.
• Evaluated three deep learning architectures: dual-projection cosine similarity, transformer encoders, and bidirectional cross-attention networks.
• Assessed model performance using random-split, cluster-split, and novel protein generalization scenarios.

Main Results:

• Models achieved high AUROC (0.91) on random splits and good performance (0.81) on cluster splits (structurally distinct ligands).
• Generalization to novel GPCRs (unseen proteins) was less accurate (AUROC 0.64), especially for out-of-family receptors.
• Ligand selectivity prediction yielded an AUROC of 0.71, but sensitivity was low (0.49).

Conclusions:

• Current deep learning models show strong performance for novel ligands but face challenges generalizing to novel GPCRs.
• A rigorous evaluation framework is crucial for assessing GPCR-ligand predictors.
• Improved protein representation learning is needed for robust computational toxicology frameworks.