Enhancing antibody affinity through experimental sampling of non-deleterious CDR mutations predicted by machine...

Area of Science:

• Biotechnology
• Computational Biology
• Immunology

Background:

• Machine learning (ML) is increasingly used for antibody optimization.
• Public antibody-antigen datasets are small and biased, hindering accurate prediction of binding affinity changes (ΔΔG) from mutations.
• Predicting deleterious vs. non-deleterious mutations is a viable alternative for antibody affinity maturation.

Purpose of the Study:

• To develop and validate an ML model for classifying antibody mutations.
• To integrate this model into a workflow for antibody affinity maturation.
• To assess the practical utility of ML in enhancing antibody binding.

Main Methods:

• Developed a Random Forest classifier (AbRFC) using expert-guided features.
• Integrated AbRFC into a computational-experimental workflow.
• Validated AbRFC on an in-house dataset, avoiding public data biases.

Main Results:

• AbRFC accurately predicted non-deleterious mutations on a bias-free validation set.
• Experimental screening of <10^2 designs identified affinity-enhancing mutations.
• Achieved up to 1000-fold increase in binding affinity for two SARS-CoV-2 antibodies.

Conclusions:

• ML-driven classification of non-deleterious mutations is a powerful strategy for antibody affinity maturation.
• This approach overcomes limitations of traditional ML models reliant on large, unbiased binding affinity datasets.
• The developed AbRFC workflow effectively enhances antibody binding to SARS-CoV-2 RBD.