AI Improves Antibody Design with Preference-Based Expression Ranking

A significant challenge in antibody design is the scarcity of labeled experimental data, which hinders the development of effective models for ranking antibody expression. To address this, a new unified preference-based learning framework has been introduced. This framework intelligently integrates the limited available quantitative expression data with a vast amount of weakly positive supervision derived from immunization data. The core of this approach involves adapting Direct Preference Optimization (DPO) for use with protein language models. This adaptation includes a union-masked log-likelihood approximation and IMGT-based alignment, which together enable efficient training on antibody sequences of varying lengths. The method was evaluated on a diverse internal dataset of 1,254 labeled sequences and 4 million unlabeled camelid-derived antibodies. Results consistently showed that this preference-based learning framework outperformed existing baseline methods across most metrics, demonstrating its scalability and effectiveness in optimizing antibody expressibility even in data-constrained environments.