Performance evaluation of artificial intelligence methods in human CD47 antibody design

Abstract

The traditional approach to antibody design requires a large number of experiments and time, which is inefficient. This may delay the therapeutic window, consume a lot of human resources and increase the economic cost of research and development. With the rapid development of artificial intelligence technology in recent years, it has been widely used in the biomedical field, which greatly improves the efficiency of drug development, especially in antibody design and function optimization shows great potential. Therefore, many researchers now choose to use artificial intelligence methods to design humanized CD47 antibodies and compare them with those obtained from experimental screening to measure the feasibility of AI methods for antibody design. In this experiment, the target antibody was designed by remotely logging into the server using Finalshell, opening the DiffAb and AlphaPanda software environments using conda, and designing the target antibody to contain three heavy chains and three light chains of the antibody. The t-test was performed to compare the Root Mean Square Deviation, Sequence Identity and ddG of the DiffAb and AlphaPanda designs, and Pymol graphs were used to show a more intuitive design result. In this way, the feasibility of the AI method for antibody design was evaluated, the performance of the design software was tested, and its strengths and weaknesses were analyzed. By analyzing the data, it can be seen that the performance of DiffAb in designing human CD47 antibody is better than AlphaPanda; in the design of L_CDR2, the antibody designed by AlphaPanda is better than DiffAb in thermodynamic stability. It was found that the AI method is able to rapidly screen a large amount of data in terms of efficiency, which greatly improves the speed of antibody discovery. At the same time, AI can successfully design human CD47 antibodies that can achieve the atomic precision and high sequence consistency of natural antibodies in terms of structure and sequence. Although AI can generate a large number of candidate antibodies in the design phase, the shortcoming of AI-designed antibodies is the lack of experimental validation, as the current experimental validation process is relatively difficult. Experimental validation of the biological properties and functions of these candidate antibodies is needed to ensure their feasibility and safety in clinical applications. However, the experimental validation process may be constrained by a variety of factors, such as the complexity of experimental conditions, the difficulty in obtaining experimental materials, and the high cost of experiments. Therefore, continuous technological innovation and interdisciplinary collaboration will promote the wide application of AI in antibody design in the future.