Protein pocket-based multi-modal drug target affinity prediction method

The invention discloses a multi-modal drug target affinity prediction method based on a protein pocket. The method comprises the following steps: 1, constructing existing drug and protein sequence data into structural data through molecular force field optimization and a protein generation model; 2, predicting a drug binding site based on a protein structure, and clustering amino acids with high scores to obtain a potential binding pocket; 3, respectively using a pre-trained sequence encoder and a pre-trained structure encoder for medicine and protein pockets to obtain sequence features representing global information and structure features representing local information; and 4, applying DNN to carry out regression prediction of drug target affinity. According to the method, the problem of scarcity of spatial information in affinity prediction is greatly relieved by introducing a protein structure generated by a large model, the noise influence of a non-interaction area is effectively reduced through multi-mo