MolAttnNet: A Predictive Model for Organic Drug Solubility Based on Graph Convolutional Networks and Transformer-Attention

The solubility of a drug in water is a crucial indicator of its therapeutic efficacy and safety. However, traditional prediction methods often overlook the fine local structural dependencies within molecules, limiting the accuracy of solubility predictions. To address this issue, we propose MolAttnNet, an innovative deep learning framework that combines Graph Neural Networks, Transformer encoders, and Long Short-Term Memory networks to enhance solubility prediction accuracy. The framework comprises three specialized modules: a Graph Convolutional Network for extracting local molecular structural features, a multi-granularity attention mechanism for capturing both local and global molecular dependencies, and an adaptive LSTM with chemically-informed forget gates for selective feature retention and noise attenuation. On an anticancer compound test set, MolAttnNet achieved an R2 of 0.5921, an RMSE of 0.5681, and an MSE of 0.3227, representing an 11.15% reduction in MSE compared to the SolPredictor baseline model. Ablation studies systematically validate the contribution of each module, demonstrating that their synergistic integration is essential for optimal performance. Furthermore, interpretability analysis confirms the model’s reliability in feature extraction and decision-making processes. The proposed framework provides a practical approach for drug solubility prediction and contributes to computational drug discovery methodologies.