Drug-target affinity prediction based on bilinear networks and protein large language model

Assessing the binding affinity of pharmaceutical agents to their respective protein targets is a crucial phase in the formulation of new drugs. Current research on prediction algorithms based on deep learning is extensive but primarily focuses on improvements in the representation of proteins and drug molecules, neglecting the representation of interactions between proteins and small molecule drugs, which leads to a lack of biological interpretability. Moreover, the rapid development of protein large language models has introduced new methods for protein representation. Consequently, this study introduces a novel approach for forecasting drug-target affinity, employing bilinear networks coupled with protein large language models, designated as BplmDTA. Initially, BplmDTA employs a protein language model to obtain feature vectors for the target; it also uses extended connectivity fingerprints (ECFP) to represent drug molecules, thereby obtaining drug feature vectors from the constructed embedding layer. Subsequently, the model calculates the interaction between the two feature vectors using a bilinear network. Finally, the affinity between the drug and the target is predicted through a multi-layer feedforward neural network. The findings from the conducted experiments demonstrate that the introduced model surpasses current techniques in forecasting drug-target interactions, offering enhanced efficacy and interpretability.