Date of Award

1-10-2024

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

attention;Drug side effects;dual-view;Heterogeneous graph neural networks;Link prediction;similarities

Supervisor

Alioune Ngom

Rights

info:eu-repo/semantics/embargoedAccess

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Abstract

Drug adverse side effects (ASEs) have substantial impacts on public health, healthcare costs, and drug discovery processes. Hospital admissions and emergency department visits are frequently attributed to adverse drug reactions (ADRs), incurring significant expenses. Identification of ASEs during the drug discovery process can slow down and prevent many candidate molecules from being selected as commercial drugs. As medication usage continues to rise, effective management of drug side effects becomes increasingly crucial. Previous works have relied on extracting and utilizing single-perspective drug features such as chemical structure, and topological information, or combining associated information between drugs and other bio-markers using Knowledge Graphs. More recent works learn in series to fuse drug representation from multiple perspectives – (microscopic) drug molecules feature and (macroscopic) over a heterogeneous network (created using a combination of various biological entity associations). In this study, we propose a novel Similarity-based Dual View Heterogeneous Graph Neural Network (SDV-HGNN) that simultaneously learns microscopic/intra-view drug substructures features using its molecular graph representation and macroscopic/inter-view drug and side-effect features over a connectivity-enhanced Drug-Adverse Side Effect Network (DSN). We introduced four additional edges between drugs and three between side effects using multi-context-specific defined similarity metrics. Our approach frames the problem as a binary classification task within the context of link prediction on a graph using a novel SDV-HGNN. We performed 10-fold cross-validation to show the superiority of our model and reported an AUROC of 0.92316±0.0014, AUPR 0.91458 ± 0.0025, and F1 0.85611 ± 0.0020.

Available for download on Tuesday, February 11, 2025

Share

COinS