Antibody Humanization via Protein Language Model and Neighbor Retrieval

Abstract

Antibody (Ab), also known as immunoglobulin (Ig), is an essential macromolecule involved in human immune response and plays an increasingly vital role in drug discovery. However, the development of antibody drugs heavily relies on humanization of murine antibodies, which often necessitates multiple rounds of sequence optimizations through laborious experimental processes. In recent years, the remarkable capabilities of machine learning have revolutionized the field of natural sciences and have also demonstrated promising applications in the field of antibody humanization. Here, we present Protein-LAnguage-model-knN (PLAN), a machine learning model leveraging protein language model and information retrieval for improving humanization of antibodies. Further, we propose DE, a computed value shows a positive correlation with antigen-binding affinity. Our in silico experimental results demonstrate that 1) the PLAN-humanized sequences average humanness score reaches 0.592, improving over the best existing method by 44.7%; 2) a 63% overlap between the PLAN-proposed mutations and the mutations validated through wet lab experiments, which is 16.7% higher than the best existing result; 3) comparable antigen-binding affinity after DE guided back mutation.