The development of therapeutic antibodies typically begins with obtaining naive candidates from animal models after immunization [1]. However, the use of purely animal antibodies carries a high risk of triggering immune responses, often indicated by elevated levels of anti-drug antibodies (ADA) [2,3]. To reduce immunogenicity, naive animal antibodies undergo a process of humanization, wherein they are modified to make them more human-like.
Numerous wet lab methods for antibody humanization have been developed [4,5,6,7,8,9], but in recent years, in silico humanization methods have gained popularity due to advancements in machine learning. Tools like Hu-mAb [10], Sapiens (BioPhi) [11], and AbNatiV [12] utilize various humanness metrics to guide their pipelines, employing both classical machine learning algorithms (Hu-mAb) and deep neural networks (BioPhi, AbNatiV). The primary challenge in humanization lies in preserving the functional properties of the antibody after modifications, such as mutations or non-fuctional framework change.
In this study, we developed a pipeline that iteratively introduces mutations into the original animal sequence while calculating a comprehensive penalty at each step. This penalty is composed of three key components: (1) humanness, (2) preservation of functionality, and (3) proximity to humans and non-animals. The resulting humanized antibody sequences not only meet the formal criteria of “humanized” antibodies established by USAN and INN [13], but also achieve OASis Identity scores (human 9-mer composition metric) [11] comparable to therapeutic antibodies. Furthermore, AF3-predicted structures of these humanized candidates show a lower RMSD compared to their AF3-predicted animal ones. We believe that integrating multiple penalty components allows for more precise selection of mutations, enhancing the degree of humanization while maintaining the functional integrity of the antibody.