A Comparative Study of Graph Neural Network Models for Drug-Target Interaction Prediction
Jaqueline Bitencourt, Anderson Rocha Tavares · 2025
Accurately predicting drug-target interactions (DTI) is crucial for computational drug discovery, yet there’s a research gap in evaluating existing graph neural network (GNN) models rather than developing novel architectures. This study provides a comparative analysis of three state-of-the-art GNN architectures – GraphSAGE, Graph Attention Network (GAT), and Graph Isomorphism Network (GIN) – for predicting interactions between chemical compounds and five protein targets. Using a dataset of 73,938 samples representing interactions between compounds and five protein targets derived from PubChem, we implement a robust evaluation framework with hyperparameter optimization and cross-validation. Our results show GraphSAGE achieves the highest accuracy (93%) and precision (79%), while GIN exhibits superior recall (72%). This work contributes to the field by: (1) providing a comprehensive evaluation framework for GNN models in DTI prediction; (2) offering empirical evidence of architecture-specific advantages for different application contexts; and (3) introducing a new benchmark dataset that facilitates reproducibility and further research in computational drug discovery.