Multiproperty optimization of therapeutic antibodies through sequence editing via large language model and reinforcement learning

Discovering antibodies with multiple beneficial properties is essential for therapeutic drug development, but experimentally identified antibodies require further optimization. Joint optimization remains challenging because sequence space is vast, annotations are sparse and property tradeoffs are common. Herein, we propose AbOptimizer, a reinforcement learning framework for multiproperty antibody optimization through parallel multisite sequence editing. AbOptimizer identifies promising editing sites and edit actions, including 'reservation', 'substitution' and optional 'insertion' or 'deletion', before recommending mutations from protein language models, while surrogate evaluation and Pareto rewards shape policy training. We evaluate AbOptimizer on three therapeutic antibodies across multiproperty, data-limited and indel-inclusive optimization. In silico, AbOptimizer improves optimized properties and outperforms baseline methods. In vitro, selected variants show affinity-specificity co-optimization for emibetuzumab, improved affinity and specificity with largely maintained thermal stability for siltuximab, and indel-inclusive affinity optimization for trastuzumab. These results validate AbOptimizer as a practical framework for therapeutic antibody optimization.