The RESP AI model accelerates the identification of tight-binding antibodies

Area of Science:

• Biotechnology
• Immunology
• Computational Biology

Background:

• Directed evolution is crucial for identifying high-affinity antibodies but can be iterative and time-consuming.
• Current deep learning methods lack confidence intervals for assessing prediction reliability in antibody design.

Purpose of the Study:

• To present RESP, a novel pipeline for the efficient identification of high-affinity antibodies.
• To overcome limitations of existing methods by incorporating uncertainty quantification.

Main Methods:

• Developed a learned representation trained on over 3 million human B-cell receptor sequences to encode antibody sequences.
• Utilized a variational Bayesian neural network for ordinal regression on directed evolution sequences binned by off-rate.
• Enabled assessment of novel antibody sequences beyond the directed evolution library.

Main Results:

• Achieved a 17-fold improvement in the dissociation constant (K D ) of the PD-L1 antibody Atezolizumab.
• Demonstrated the pipeline's ability to expand the search space for optimal antibody candidates.
• Quantified the likelihood of sequences being tight binders against a target antigen.

Conclusions:

• RESP significantly accelerates the identification of high-affinity antibodies.
• The pipeline's uncertainty quantification enhances the reliability of antibody design predictions.
• This approach shows great potential for facilitating general antibody development and therapeutic discovery.