This paper proposes a new deep learning model, SAGERank, which predicts antibody-antigen docking outcomes using Graph Sample and Aggregate Networks. As recognition of foreign antigens by antibiotics is fundamental to immune response, this framework has the ability to accelerate the neoantigen discovery and improve adoptive immunotherapy efficacy for cancer treatment. By configuring Ab-Ag in a natural graph network, with atoms as nodes and interactions as edges, the model predicts docking quality and binding affinity. The framework was trained and evaluated on antibody-antigen docking datasets, as well as 66 cancer targets and their corresponding antibodies taken from the immune epitope database. Performance of the framework was compared with the success rate and scoring functions of other models. Results of the study include the SAGERank docking model outperforming major existing scoring functions and successfully predicted majority epitopes in a cancer target dataset. Additionally, SAGERank had outperformed another prediction algorithm, Alphafold3, in nanobody–antigen structure prediction. The study demonstrates the ability of inductive deep learning, to accurately capture the physicochemical features of amino acid interactions when working with limited datasets. This work suggests potential of SAGERank to accelerate the discovery and improvement of therapeutic antibody design in targeting cancer antigens, representing a promising tool in advancing immunotherapy research.