An antigen-specific protein language model guides CDR optimization for antibody design.

Ab-Affinity, a fine-tuned protein language model for SARS-CoV-2 spike peptide binding, predicts antibody affinity and enables >100-fold improvement over parental antibodies.

• Why it matters: Neutralizing antibody design depends on achieving high binding affinity, but optimizing antibody sequences against specific spike epitopes remains difficult and often relies on less accurate affinity predictors.
• What they did: The model was trained to predict binding affinity for antibodies targeting a SARS-CoV-2 spike protein peptide, then evaluated via cross-validation on held-out validation and independent test sets. It was combined with a genetic algorithm and simulated annealing to diversify sequences and optimize predicted fitness.
• The result: Ab-Affinity outperformed other PLM-based affinity predictors on validation and independent tests, generating candidates with over 100-fold higher predicted affinity than parental antibodies; docking and dynamics supported stronger epitope interactions, and in vitro testing showed increased binding for a top candidate.

AI summary of the abstract · may contain errors · not medical advice · read the paper.