Multi-modal protein-ligand binding affinity prediction method based on cross-channel fusion

The invention discloses a multi-modal protein-ligand binding affinity prediction method based on cross-channel fusion, and aims to improve the prediction performance of protein-ligand binding affinity. The method specifically comprises the following steps: constructing a protein-ligand compound structure and an affinity tag; preprocessing data, extracting a residue sequence, and constructing a protein pocket residue map, a ligand molecular map and a pocket residue-ligand atom interaction map; corresponding input features are generated, and feature extraction is carried out; fusing the output of all modules to generate vector representation of a complex; binding affinity is predicted through a fully connected layer. According to the method, protein pockets and ligand molecules are represented through a combined framework of GNN and Transformer, so that the problem of local limitation of GNN message passing is solved, and a global view is provided for nodes. In addition, through global representation of residue