AI and the World's Hard Problems · Bioengineering & synthetic biology · Nature (computational biology) · 2026-10-26
Multiproperty optimization of therapeutic antibodies through sequence editing via large language model and reinforcement learning
Designing therapeutic antibodies that improve several properties at once is hard because sequence space is vast, property annotations are sparse, and properties often trade off against each other. The authors propose AbOptimizer, a reinforcement learning framework that selects editing sites and actions (reservation, substitution, and optional insertion or deletion) and recommends mutations from protein language models. Tested on three therapeutic antibodies, it improved the optimized properties in silico beyond baseline methods, and selected variants showed affinity–specificity co-optimization for emibetuzumab, improved affinity and specificity with largely maintained thermal stability for siltuximab, and indel-inclusive affinity optimization for trastuzumab.