Graph Neural Networks for Drug Screening

Abstract

Drug discovery is a lengthy and costly process that often involves small, noisy, and imbalanced datasets. In our study, we investigate the use of graph neural networks (GNNs) for predicting molecular homeostatic activity in neuronal cells through transfer learning. We evaluate Graph Convolutional Networks (GCNs) and Message Passing Neural Networks (MPNNs) with transfer learning, comparing their performance to Random Forest and non-transfer GNN baselines. To guide the selection of source datasets for pre-training, we implement a molecular latent representation similarity framework across nine MoleculeNet datasets. Additionally, we fine-tune a foundational molecular model on our target dataset. We evaluate the models using five-fold cross-validation, using the Area Under the Receiver Operating Characteristic curve (AUC-ROC) and the Area Under the Precision-Recall curve (AUC-PR) as metrics. Our results indicate that transferring knowledge from high-similarity source datasets outperforms the baseline models. Moreover, source-to-target transfer is more effective than fine-tuning the foundation model; however, the foundation model exhibits superior generalisation capabilities. Finally, we employ a selected set of models to rank an unlabelled molecular dataset. Our findings demonstrate that GNNs, combined with similarity-guided transfer learning, enhance performance in predicting bioactivity within low-data and imbalanced settings, highlighting the importance of carefully selecting source datasets to avoid negative transfer.