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

AbOptimizer improves multiproperty optimization of therapeutic antibodies by parallel multisite sequence editing, outperforming baseline methods in silico and enabling affinity-specificity co-optimization in vitro.

• Why it matters: Multiproperty antibody discovery is hard because sequence space is vast, annotations are sparse, and property tradeoffs limit joint optimization.
• What they did: AbOptimizer uses reinforcement learning with surrogate evaluation and Pareto rewards, selecting editing sites and actions (reservation, substitution, optional insertion/deletion) and recommending mutations from protein language models.
• The result: In vitro, AbOptimizer yields affinity-specificity co-optimization for emibetuzumab, improved affinity and specificity with largely maintained thermal stability for siltuximab, and indel-inclusive affinity optimization for trastuzumab.