Efficient inference of non-polyreactive antibody variants dependent on local fine-tuning

Antibody polyreactivity, which is characterized by broad, low-affinity off-target binding, is an undesirable drug property associated with rapid clearance. Existing in vitro screening methods often suffer from interassay variability, while current in silico models lack the generalizability needed to efficiently identify non-polyreactive variants within the localized sequence spaces typical of drug‑optimization campaigns. Here, we address both challenges through refining low‑ and high‑throughput assays suitable for generating standardized datasets and quantifying the impact of localized protein language model (PLM) tuning in a specific drug‑optimization context. We demonstrate that range‑normalized summation scores derived from multi-concentration monoclonal antibody ELISA provide highly reproducible ground-truth measurements with inter-experiment Pearson r ≥ 0.99. We verified a previously developed yeast-display high-throughput approach for polyreactivity data generation and screened a library of 240,000 single-chain variable fragment heavy‑chain complementarity-determining region variants, then used this data to fine-tune two PLMs. As expected, local fine-tuning dramatically improved model performance on a test set in the relevant sequence space, while performance on an out-of-distribution set of 80 clinical antibodies was largely unchanged. The top‑performing base and tuned models were tasked with inferring 18 low or non-polyreactive variants. The observed inference success rates were 0% and 66.6%, respectively, demonstrating the practical utility of local tuning in a drug development context. Together, this work provides a reproducible, integrated framework combining robust in vitro assay methodology with locally tuned in silico models to efficiently resolve polyreactivity during antibody optimization.