SGCL-DPI: structure-guided curriculum learning for drug-protein interaction prediction

Predicting drug–protein interactions (DPIs) is a critical challenge in bioinformatics and drug discovery, as each computational approach provides only a partial view of these complex molecular relationships. Deep learning techniques such as graph neural networks (GNNs) capture local structural patterns from molecular graphs, whereas classical algorithms like Random Forest (RF) leverage global molecular descriptors. We introduce Structure-Guided Curriculum Learning forDrug-Protein Interaction Prediction (SGCL-DPI), a structure-guided curriculum learning framework that initially leverages the global molecular insights captured by a RF model to guide and progressively refine the structural pattern learning of a GNN, enhancing drug–protein interaction prediction accuracy. SGCL-DPI employs a curriculum learning strategy in which an RF teacher model provides initial high-level predictive guidance to a GNN student model, and the focus of training gradually shifts from the RF to the GNN. The training objective integrates three components: binary cross-entropy for correct interaction classification, knowledge distillation to align the GNN’s outputs with the RF’s predictions, and a structural consistency term to maintain similarity-based relational patterns in the learned representations. We evaluated SGCL-DPI on two benchmark datasets: BindingDB and a challenging STITCH-derived dataset. On the BindingDB dataset, a standalone RF baseline using classical molecular descriptors achieved an area under the receiver operating characteristic curve (AUC-ROC) of 99.18% and an area under the precision–recall curve (AUPR) of 99.14%, outperforming many deep learning models. This result highlights the strong predictive power of traditional descriptors on this dataset. On the more difficult STITCH-derived hard split, SGCL-DPI attained a balanced performance, with an F1-score of 67.06%, an AUC-ROC of 82.33%, and an AUPR of 71.69%. Notably, the model outperformed both a purely GNN-based deep model and the traditional RF baseline in terms of F1-score, demonstrating superior ability to predict interactions for entirely unseen drug–protein pairs. These findings demonstrate that SGCL-DPI effectively bridges classical machine learning and deep learning approaches for DPI prediction. By integrating global descriptor-based knowledge with graph-based structural learning, the proposed framework significantly improves predictive accuracy and generalization on challenging interaction prediction tasks, highlighting a promising direction for future DPI prediction research. The full implementation of the proposed framework is publicly available at https://github.com/soufanom/SGCL-DPI to ensure transparency and reproducibility.