Generative AI-driven Antibody Design and Optimization: A Transformers Paradigm

Abstract

The human immune system generates high-affinity antibodies against pathogens and diseases. These antibodies serve as key therapeutic and diagnostic tools for disease classification and treatment. Traditional approaches like display technologies can generate potential antibody leads, but they come with challenges such as expressibility, viscosity, immunogenicity, and pharmacokinetics.. The recent advancements in AI have led to the foundation of the generative AI, which has also impacted the field of bioinformatics. The field of generative AI like transformer-based NLP models have significantly improved the development of better computational tools in protein and antibody design. These models can leverage protein and antibody sequence information to reduce the need for resourceintensive display technology experiments. However, existing models are often trained on multi-species datasets, which can introduce species-specific biases and limit their ability to generate diverse human antibodies. Here, this study proposed; AbSynth, a class of transformer-based antibody language models exclusively trained on 1 million human antibodies sequences dataset to improve generalizability in human antibody design. AbSynth models were tested on the natural antibody 1E6J to improve binding affinity. Of the 400 generated antibodies, 10 showed significant improvement in binding affinity.. Furthermore, AbSynth-generated sequences exhibited 96% humanness in heavy chain sequences under low sampling parameters, demonstrating its potential as an effective tool for designing and isolating diverse, humanized antibody candidates.