Applications of Graph Neural Networks in Drug Discovery and Drug Design: Trends, Challenges

Graph neural networks (GNNs) have emerged as a genuinely transformative class of deep learning architectures for computational drug discovery, making it possible to learn molecular representations directly from graph-structured chemical data, end-to-end, without relying on hand-engineered features. Where conventional descriptor-based or fingerprint-based approaches fall short, GNNs preserve the relational topology of molecular structures-enabling more accurate predictions of biological activity, physicochemical properties, and ADMET profiles (absorption, distribution, metabolism, excretion, and toxicity). This review critically examines the mechanistic foundations of the major GNN architectures in use today-message passing neural networks (MPNNs), graph convolutional networks (GCNs), graph attention networks (GATS), and equivariant geometric GNNs-and evaluates how they have been deployed across five core areas of drug discovery: molecular property prediction, virtual screening, de novo molecular design, protein-ligand interaction modeling, and ADMET profiling. We provide comparative analysis of architectures, benchmark datasets, and reported performance metrics, while taking a critical look at the persistent limitations that continue to challenge the field: out-of-distribution (OOD) generalization failure, interpretability gaps, conformational insensitivity, and data quality constraints. A dedicated section addresses the often-overlooked problems with benchmarking-dataset bias, inconsistent evaluation protocols, and reproducibility failures that make cross-study comparisons genuinely difficult. Finally, we identify the most promising emerging directions: GNN-transformer hybrid architectures, physics-informed learning, uncertainty quantification, and closed- loop autonomous discovery systems.

کلیدواژه ها:

graph neural networks ، drug discovery ، molecular property prediction ، ADMET profiling