DBA-GNN: A Sparse and Dense Feature Extraction Bifrequency Graph Neural Network for Drug Property Prediction

Graph neural networks (GNNs) have significantly enhanced the analysis of drug molecules represented as graph-structured data. However, existing methods often neglect the sparse or dense characteristics of graphs and struggle with attribute prediction accuracy due to small sample sizes. To address these challenges, we introduce Drug Bifrequency Attention Graph Neural Network (DBA-GNN). DBA-GNN features a two-stage learning process: a warm-up stage where graph data undergo random masking to enhance foundational modeling capabilities, followed by a task-oriented stage using the Bifrequency Attention Mixer (BAM) module, which models sparse or dense graph characteristics to improve adaptability and predictive accuracy. Our experiments on a dataset of 188 compounds show that DBA-GNN achieves an accuracy of 92.4% and sensitivity of 93.6%, outperforming state-of-the-art models by up to 10%.