Antibody therapeutics has been extensively studied in drug discovery and development within the past decades. One increasingly popular focus in the antibody discovery pipeline is the optimization step for therapeutic leads. Both traditional methods and in silico approaches aim to generate candidates with high binding affinity against specific target antigens. Traditional in vitro approaches use hybridoma or phage display for candidate selection, and surface plasmon resonance (SPR) for evaluation, while in silico computational approaches aim to reduce the high cost and improve efficiency by incorporating mathematical algorithms and computational processing power in the design process. In the present study, we investigated different graph-based designs for depicting antibody-antigen interactions in terms of antibody affinity prediction using deep learning techniques. While other in silico computations require experimentally determined crystal structures, our study took interest in the capability of sequence-based models for in silico antibody maturation. Our preliminary studies achieved satisfying prediction accuracy on binding affinities comparing to conventional approaches and other deep learning approaches. To further study the antibody-antigen binding specificity, and to simulate the optimization process in real-world scenario, we introduced pairwise prediction strategy. We performed analysis based on both baseline and pairwise prediction results. The resulting prediction and efficiency prove the feasibility and computational efficiency of sequence-based method to be adapted as a scalable industry practice.