Affinity prediction of inhibitor-kinase based on mixture of experts enhanced by multimodal feature semantic analysis.

Accurate identification of inhibitor-kinase binding affinity is crucial for drug discovery. However, many deep learning models often overlook high-order feature information from biological networks and face challenges related to the cold-start problem when extracting biomolecular features. The EGFR family among human kinases serves as a primary drug target in cancer and other diseases. In this study, to improve prediction accuracy, we propose Mokin: Affinity Prediction of Inhibitor-Kinase Based on Mixture of Experts Enhanced by Multimodal Feature Semantic Analysis .This work is the first to utilize a Mixture of Experts (MoE) system as a bridge for extracting drug molecular features, while simultaneously integrating protein and protein-protein interaction (PPI) network features to enrich semantic representations.We innovatively design a global memory gated block router within the MoE to capture historical global features and explore latent content in the current state. By integrating protein and PPI network features, we obtain comprehensive biological characteristics, which are then combined with drug molecular features through a bilinear attention mechanism for downstream prediction. Our study achieves state-of-the-art performance in inhibitor-kinase binding mechanisms, especially excelling in cold-start experimental settings. Furthermore, the final stage kinase analysis further validates the model's generalization capability.