Goran Rakočević1,2
1Absci, Vancouver, Washington, USA
2School of computing, Union University, Belgrade
grakocevic [at] absci.com
Abstract
Advancements in antibody engineering are crucial for developing effective and safe therapeutic candidates. Traditional approaches often involve limited screening of sequence space, resulting in drug candidates with suboptimal binding affinity, developability, or immunogenicity. However, recent breakthroughs in deep learning and generative artificial intelligence (AI) offer promising solutions to overcome these challenges.
In our work, we utilized deep contextual language models trained on high-throughput affinity data to quantitatively predict binding of unseen antibody sequence variants. Our approach spans a wide range of binding affinities, demonstrating the potential to optimize antibody engineering. Additionally, we introduced a “naturalness” metric that measures similarity to natural immunoglobulins. We found that naturalness is associated with measures of drug developability and immunogenicity, allowing us to optimize it alongside binding affinity using a genetic algorithm.
Additionally, we explored generative AI-based antibody design, and achieved successful design of all complementarity-determining regions (CDRs) in the heavy chain of the antibodies. Our designed antibodies exhibit high binding rates, surpassing randomly sampled antibodies from the Observed Antibody Space. Moreover, these AI-designed binders display high diversity, low sequence identity to known antibodies, and favorable naturalness scores, indicating desirable developability profiles and reduced immunogenicity.
Collectively, our findings demonstrate the immense potential of deep learning and generative AI in revolutionizing antibody optimization and design. By leveraging large-scale data, predictive models, and high-throughput experimentation, we can accelerate and improve our antibody engineering capabilities. The integration of deep contextual language models and the incorporation of naturalness into the design process provide intelligent screening approaches. Similarly, the application of generative AI enables us to efficiently and precisely design antibodies from scratch, outperforming traditional methods in terms of speed and quality.
Keywords: artificial intelligence, antibody design
Acknowledgement: I would like to thank all of my coauthors and collaborators: Amir Shanehsazzadeh, Sharrol Bachas, Matt McPartlon, George Kasun, John M. Sutton, Andrea K. Steiger, Richard Shuai, Christa Kohnert, Jahir M. Gutierrez, Chelsea Chung, Breanna K. Luton, Nicolas Diaz, Simon Levine, Julian Alverio, Bailey Knight, Macey Radach, Alex Morehead, Katherine Bateman, David A. Spencer, Zachary McDargh, Jovan Cejovic, Gaelin Kopec-Belliveau, Robel Haile, Edriss Yassine, Cailen McCloskey, Monica Natividad, Dalton Chapman, Joshua Bennett, Jubair Hossain, Abigail B. Ventura, Gustavo M. Canales, Muttappa Gowda, Kerianne A. Jackson, Jennifer T. Stanton, Marcin Ura, Luka Stojanovic, Engin Yapici, Katherine Moran, Rodante Caguiat, Amber Brown, Shaheed Abdulhaqq, Zheyuan Guo, Lillian R. Klug, Miles Gander, Joshua Meier, Anand V. Sastry, Andrew Stachyra, Borka Medjo, Vincent Blay, Alexander Brown, Nebojsa Tijanic, Rebecca Viazzo, Rebecca Consbruck, Hayley Carter, Jacob Shaul, Randal S. Olson, Sean McClain, Matthew Weinstock, Gregory Hannum, Ariel Schwartz, Roberto Spreafico
