Drug-target affinity prediction method and device based on lightweight cross-attention bridge

Abstract

The application discloses a drug-target affinity prediction method and device based on a lightweight cross-attention bridge, comprising the following steps: constructing a drug-target affinity dataset; extracting global semantic features of drugs and proteins by using a frozen pre-training large model, and synchronously extracting local structure features by using an expert encoder; constructing a lightweight cross-attention bridge module to realize deep fusion of global knowledge and local features; modeling the cross-modal correlation between drugs and proteins by using a bidirectional cross-attention module; designing a multi-objective loss function based on mean square error and consistency index to optimize parameters; and finally, performing affinity prediction by using the optimized model. By using the architecture of "frozen large model + trainable expert + cross-attention bridge", the application reduces the calculation cost, fully excavates the prior knowledge of the pre-training model and the task-specific local structure information, and improves the accuracy and generalization ability of the affinity prediction.

Images