A deep hybrid GNN based on edge-conditioned and graph isomorphism network convolutions for PC-3 anticancer screening
Adrian S. Remigio, Jeffrey A. Aborot, Jonel Hong · 2023
Computational methods are employed for efficient drug discovery for cancer treatment. Recent efforts on virtual anticancer screening have utilized various machine and learning approaches, which are advantageous for resolving relationships in large and complex data. In this study, a Deep hybrid GNN architecture based on graph isomorphism network (GIN) and edge-conditioned convolution (ECC) layers was applied for the anticancer screening of small molecules for treatment of PC-3 tumors. After obtaining the graph representation of small molecules, the corresponding graphstructured data were fed to the Deep hybrid GNN for model training via adaptive moment estimation and evaluation. An exhaustive grid hyper-parameter tuning of various network architecture configurations and hyper-parameters were also included in the study to determine their influence on model performance. Our proposed Deep hybrid GNN model outperformed baseline GNN models for the PC-3 anticancer screening task. However, a small gap in predictive performance existed between the Deep hybrid GNN and baseline GNN based on ECC, suggesting the effective embedding capabilities of ECC for such task. The hyper-parameter tuning experiment indicate sufficiently high accuracies were obtained for various network architecture configurations. In most instances, an accuracy and area under the receiver operating curve (AUC) greater than 0.85 and 0.80, respectively, were obtained. Although the proposed GNN model can be utilized to facilitate anticancer screening, further optimization and validation are required prior to clinical translation.