Ionic liquid toxicity is a major cause of morbidity and mortality worldwide. However, determining the toxicity of ionic liquids is still a challenge for the industry. In this study, we propose a graph attention network (GAT) based framework specifically adapted for predicting ionic liquid toxicity using graph-based molecular representations. In order to improve the model’s accuracy, we added physicochemical descriptors and hyperparameter optimization using the Optuna framework. In addition, Grad-CAM (gradient-weighted class activation mapping) visualizations and permutation feature importance analysis were employed to visualize the significance of the nodes and edges in the graphs. Experimental results demonstrate that the EnhancedGAT achieves lower prediction errors and stronger correlations compared to traditional baselines, indicating high predictive reliability.