Year: 2023
Type: dissertation
Abstract: Accurate prediction of Drug-Protein Interactions (DPIs) is critical to understanding how drugs work and helping to accelerate drug discovery. Recent advances in Graph Neural Networks (GNN) have shown promise in effectively modeling complex relationships in biological networks. The Deep Multi-Graph Neural Networks for Drug-Protein Interactions (DMGNNs) framework, which integrates three specialized GNNs, is presented in this thesis: Drug-Drug Interaction GNN (DDI-GNN), Protein-Protein Interaction GNN (PPI-GNN), and Drug-Protein Interaction GNN (DPI-GNN). Compared to previous works, our novel approach distinctly divides the learning process into three separate tasks where drug-drug similarities, protein-protein similarities and drug-protein interactions are learned separately. For computing and learning similarities, DDI-GNN uses drug-drug graph features while PPI-GNN is based on protein-protein graph features. Once these two GNNs finish training, learned embeddings are passed to the DPI-GNN to predict drug-protein interactions. DMGNNs framework is a powerful approach for the accurate modeling of complex drug-protein relationships with potential applications in drug discovery and development. The results show that DMGNNs outperformed the state-of-the-art with an average improvement of about 28% using the F1-score. This confirms that modeling similarities of drug and protein graphs separately helps in extracting a more representative set of embeddings as compared to one combined graph learning task.
Source: StFX Scholar
Author: Zamil, Md Yosuf
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
Funder: St. Francis Xavier University
OpenAlex ID: W7208208488