Monoclonal antibody inhibitors targeting PD-1/PD-L1 immune checkpoints are gradually entering the market and have achieved certain positive effects in the treatments of various types of tumors. However, with the expansion of application, the limitations of antibody drugs and problems such as excessive homogenization of research gradually appear, making small-molecule inhibitors the new focus of researchers. This study aims to use ligand-based and structure-based binding activity prediction methods to achieve virtual screening of small-molecule inhibitors targeting PD-L1, thereby helping to accelerate the development of small molecule drugs. A dataset of PD-L1 small-molecule inhibitory activity from relevant research literature and patents is collected and activity judgment classification models with intensity prediction regression models are constructed based on different molecular featurization methods and machine learning algorithms. The two types of models filter 68 candidate compounds with high PD-L1 inhibitory activity from a large drug-like small molecule screening pool (ZINC15). Ten of these compounds not only have good drug similarities and pharmacokinetics, but also exhibit comparable binding affinities and similar mechanisms of action with previous reported hotspot compounds in molecular docking. This phenomenon is further verified in subsequent molecular dynamics simulation and the estimation of binding free energy. In this study, a virtual screening workflow integrating ligand-based method and structure-based method is developed, and potential PD-L1 small-molecule inhibitors are effectively screened from large compound databases, which is expected to help accelerate the application and expansion of tumor immunotherapy.