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

The capacity of an antibody to recognize and bind an antigen with high specificity and strength (i.e., its binding affinity ) is a key determinant in the design and optimization of neutralizing antibodies against pathogenic targets. Here we introduce Ab-Affinity, a fine-tuned protein language model (PLM) that predicts the binding affinity of an antibody against a target peptide within the SARS-CoV-2 spike protein. Cross-validation results shows that that Ab-Affinity outperforms other PLM-based affinity predictors when evaluated on a held-out validation and independent test sets. When integrated with a genetic algorithm and simulated annealing for sequence diversification and fitness optimization, Ab-Affinity enables the generation of candidate antibodies exhibiting over 100-fold improvement in predicted binding affinity relative to its parental (natural) antibody. More broadly, antibodies designed with Ab-Affinity attain higher predicted binding affinities than those produced by alternative mutational strategies, yielding synthetic antibodies with favorable predicted biophysical characteristics. Molecular docking and molecular dynamics simulations for the strongest synthetic candidates indicate enhanced interactions and increased stability with the target peptide epitope within the spike protein. Experimental validation in vitro for a top-ranked candidate reveals an apparent increase in binding relative to its parental antibody. All model weights, source code, and analysis pipelines are publicly released to allow full reproducibility.