Graph neural networks

• Main Author: Lian, Ran
• Other Authors: Luu Anh Tuan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/169974
• Institution: Nanyang Technological University

Molecular property predictions are crucial for drug discovery and development. Predicting the properties of molecular compounds could essentially speed up the research process in areas such as drug designing and chemical substance discovery. In recent years, Graph Neural Networks (GNNs) have become increasingly attractive methods for molecular property prediction due to their abilities to analyze graph structural data with chemical structures being easily displayed as graphs. In this paper, I will perform a comparative study on some of the state-of-the-arts architectures used today, including Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), Attentive FP and Path-Augmented Graph Transformer Networks (GATNs) for molecular property prediction on 5 benchmark datasets (HIV, Tox21, BBBP, ClinTox and BACE). With fixed hyperparameters choices on different deep learning architectures, the experimental results showed that the PAGTN model outperformed other GNN architectures on several datasets. Finally, to simplify the drug discovery process for pharmaceutical scientists, I proposed one possible application using the best model which could be adopted by them in the development and testing of new drugs.