Virtual drug screening that provides possible drug candidates facilitates early-stage drug discovery. It works by large scale predicting native-like protein-ligand complexes (PLC) from an abundance of docking decoys. Many affinity predicting models currently in use fail to provide reliable prediction because of a lack of non-binding data during model training, lost critical physical-chemical features, and difficulties in learning abstract information with limited neural layers. In this paper, we developed a deep learning model, DeepBindBC for classifying putative ligands as binding or non-binding. Our model incorporates information of non-binding interactions, making it more suitable for real applications. ResNet model architecture and more detailed atom type representation guarantee implicit features can be learned more accurately. DeepBindBC identified a novel human pancreatic $\alpha$-amylase binder validated by a fluorescence spectral experiment (Ka $=1.0\times 10^{5}\mathrm{M}$). Furthermore, we proposed a virtual screening pipeline by incorporating multiple complementary methods, such as DFCNN, Autodock vina docking, DeepBindBC, and pocket molecular dynamics simulation. Three potential inhibitors of pancreatic $\alpha$-amylase were identified by the proposed pipeline, and interestingly most of them contain glycan groups. Additionally, an online webserver based on the model is available at http://cbblab.siat.ac.cn/DeepBindBC/index.php for the convenience of the users.