AdptDilatedGCN: Protein-ligand binding affinity prediction based on multi-scale interaction fusion mechanism and...

Predicting protein-ligand binding affinity is crucial for drug discovery. However, existing prediction methods often make insufficient use of the features of proteins and ligands, lack interactions between different information, and have difficulty in capturing and integrating both fine- and coarse-grained information of proteins and ligands at the same time. Therefore, we propose a deep learning method called AdptDilatedGCN, which fully interacts with protein and ligand features through multi-scale interaction and fusion mechanism, mining potential fine-grained and coarsely grained information and fusing it, combining skip connections and multi-scale fusion strategy to improve feature utilization and enhance features. In the model, dilated Graph Convolutional Network(GCN) combines multi-head attention mechanism and dilated convolution to overcome the limitations of traditional GCN, while taking into account the interactions between amino acid nodes in the protein graph. In addition, an adaptive Gated Recurrent Unit(GRU) is designed to dynamically enhance the model's ability to integrate multi-source information in adaptive GCN. The Vina weights obtained from AutoDock Vina are used to supplement and further enhance the features learned from GCN. Finally, The model achieves Pearson correlation coefficients of 0.837 and 0.803 on the CASF-2016 and CASF-2013 public test sets, respectively.