PREDICTING DRUG TOXICITY ON TOX21 WITH A MULTI-TASK GNN-TRANSFORMER MODEL

As research into medicinal chemical synthesis deepens, an increasing number of novel drug molecules with promising clinical therapeutic potential are emerging. During drug development, many compounds are discarded due to excessive toxic side effects, while traditional toxicity testing faces challenges of high cost and lengthy timelines. To enable rapid toxicity assessment by researchers, this paper proposes a drug toxicity prediction model (TGT) based on GNN-Transformer. The model was constructed and validated using the Tox21 dataset, with caffeine selected for practical generalisation testing. The Tox21 dataset encompasses toxicity test results for diverse compounds, providing high-quality data. The model architecture leverages graph neural networks (GNN) to process molecular graph-structured data, effectively capturing complex topological relationships and chemical information within molecules. This transforms molecular graph structures into meaningful node feature representations. The Transformer component, with its exceptional sequence modelling capabilities, further learns from GNN-extracted features, capturing long-range dependencies between molecular structures. Through training and optimisation, the model demonstrated commendable performance in toxicity prediction tasks, achieving an average AUC of 0.7488 on the validation set. Its high accuracy in predicting drug toxicity was further validated through practical application on caffeine molecules, establishing it as an efficient and precise predictive tool for early-stage drug safety assessment.The GNN-Transformer drug toxicity prediction model proposed in this study enhances prediction reliability by integrating multi-task learning with interpretability techniques. It serves as an auxiliary pre-screening tool for drug development, thereby helping to shorten the research and development cycle and reduce costs.