FocusLG: Focusing on local and global molecular representation for kinase inhibitor binding affinity prediction

Kinase inhibitor binding affinity prediction is a fundamental and important task in drug discovery. Existing approaches for kinase inhibitor affinity research suffer from limitations including insufficient feature representation, limited labeled data, and relatively low prediction accuracy of the model. We propose FOCUSLG, which is a kinase inhibitor affinity prediction method focusing on both local and global features of the inhibitor molecules. Specifically, the model employs ESM-2 as the protein feature encoder to effectively extract kinase features, utilizes ChemBERTa as the molecular feature encoder to capture global features of inhibitors, and incorporates a graph convolutional network (GCN) to focus on the local molecular features of inhibitors. In the task of kinase-inhibitor binding affinity prediction in the MAPK pathway (including RAF, MEK, and ERK), the model achieves prediction accuracy of 88.7% on RAF and 81.0% on ERK. Furthermore, on a dataset encompassing over 200 kinases, the method achieves accuracy and AUC values of 68.4% and 76.3%, respectively, and also shows good performance on the PDBbind dataset, indicating the strong generalization capability.