TUPocket: 3D Convolutional Neural Network Combined with Transformer for Ligand Binding Site Detection

In the field of drug development, it is crucial to accurately predict the binding sites of protein interactions, which directly affects the design efficiency and development efficiency of targeted drugs. A targeted drug can only play its due role when it binds specifically to a defined site within the protein structure. If the binding site prediction is inaccurate, it may result in unexpected side effects. Not only might the drug fail to treat the disease effectively, but it could also potentially harm the health of the individual taking the medication. Despite the classical Unet-3D model having shown initial potential in 3D image segmentation and its capability to handle most scenarios, its performance constraints increasingly manifest when dealing with intricate and complex biomolecular binding sites. As a result, this study uniquely incorporates the attention mechanism derived from Transformer, alongside a variety of complementary attention modules and technologies, thereby significantly enhancing both the predictive accuracy and generalization capabilities of the model. The experimental results demonstrate that the model exhibits superior ability of prediction and generalization. The enhanced model not only attains upper precision on experimental datasets, but also has excellent performance when dealing with uncharted data. The present study reveals the enhancement in the predictive capacity for protein binding sites achieved by the optimized Unet-3D architecture integrated with Transformer’s attention framework.