GraphDrug: A Multimodal Graph Learning for Predicting Bioactivity and Pharmacokinetics of Drug Candidates

Abstract

The accurate prediction of drug bioactivity and pharmacokinetic (PK) properties is a cornerstone of early-stage drug discovery. Traditional computational models rely heavily on molecular descriptors and handcrafted features, limiting their generalizability and performance. In this study, we introduce GraphDrug, a graph neural network (GNN)-based platform that learns molecular representations directly from molecular graphs to predict bioactivity and ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties. Our model incorporates advanced graph convolutional networks with attention mechanisms to capture complex molecular interactions and hierarchical structural dependencies. Benchmarked against several public datasets including MoleculeNet and TDC, GraphDrug consistently outperforms traditional machine learning baselines and SMILES-based deep learning approaches. The platform offers an interpretable, scalable, and end-to-end pipeline for virtual screening and lead optimization in drug development.