A closer look at Saunders et al. (2026) and how well-characterized ichorbio reference antibodies supported the development of a robust experimental framework for antibody polyreactivity assessment.

While remarkable advances in antibody engineering have accelerated therapeutic discovery, developability remains a major bottleneck during lead optimization. Among the various liabilities that can compromise clinical success, polyreactivity, the tendency of antibodies to bind multiple unrelated antigens with low affinity, has emerged as one of the strongest predictors of poor pharmacokinetic behavior, increased systemic clearance, and downstream development risk.

Current strategies combine experimental screening with computational prediction, yet both approaches present important limitations. Conventional ELISA- and yeast display-based assays often suffer from assay-to-assay variability, while existing machine learning models frequently struggle to accurately predict polyreactivity within the narrow sequence spaces typically explored during antibody affinity maturation.

In their recent publication in mAbs, Saunders and colleagues from Absci introduce an integrated experimental and computational workflow that substantially improves the identification of non-polyreactive antibody variants. By combining highly reproducible in vitro assays with locally fine-tuned protein language models, the authors demonstrate a practical strategy for accelerating antibody optimization while reducing experimental burden.

A central contribution of the study is the development of a standardized ELISA workflow capable of generating highly reproducible quantitative polyreactivity measurements.

The authors evaluated 32 monoclonal antibodies against four commonly used polyspecificity reagents, ovalbumin (OVA), insulin, DNA, and polyspecificity reagent (PSR), across multiple concentrations and independent experimental days. Instead of relying on individual ELISA measurements, they introduced a Range-Normalized Summation (RNS) score that integrates dose-dependent responses while correcting for plate-to-plate variability.

The resulting assay demonstrated exceptional reproducibility, with Pearson correlation coefficients exceeding 0.99 between independent experiments. Moreover, RNS scores showed strong agreement across different reagents and correlated well with previously published clinical antibody datasets, supporting their use as robust “ground truth” measurements for future machine learning applications.

By establishing a standardized experimental framework, the study provides an important foundation for generating expandable datasets suitable for AI-driven antibody developability prediction.

To evaluate computational prediction, Saunders et al. generated a yeast surface display library comprising approximately 240,000 antibody variants containing localized heavy-chain CDR mutations.

Using iterative flow cytometric selection against OVA, PSR and insulin, the authors generated a large experimental dataset that served both as validation data and as training material for multiple protein language models.

The most significant finding emerged after local fine-tuning of an ESM-based protein language model using antibody variants closely related to the parental lead sequence.

Whereas general-purpose models showed only moderate predictive performance, locally fine-tuned models achieved dramatic improvements across accuracy, recall and F1 score. More importantly, these improvements translated into practical antibody engineering outcomes.

When challenged to identify new low-polyreactive variants, the untuned model failed to produce a single successful candidate. In contrast, the locally tuned model correctly identified 12 of 18 experimentally validated variants exhibiting low or no detectable polyreactivity, corresponding to a 66.6% success rate.

These results reinforce an increasingly important concept in therapeutic antibody engineering: model performance depends not only on model architecture, but also on training with sequence data that closely reflects the optimization campaign under investigation.

Together with Briakinumab and Herceptin, these controls enabled plate-to-plate normalization, calibration of RNS values, and continuous quality control across all experimental runs.

Their inclusion contributed to the remarkable reproducibility reported throughout the study and highlights the importance of reliable, well-characterized reference antibodies for standardized antibody developability workflows.

As machine learning increasingly depends on high-quality experimental data, the value of reproducible reference reagents becomes even more significant. Consistent controls help ensure that datasets collected over time, and potentially across laboratories, remain comparable, enabling more reliable model training and validation.

Rather than positioning artificial intelligence as a replacement for experimental screening, Saunders et al. demonstrate the value of integrating robust laboratory assays with modern protein language models.

Their workflow establishes a practical feedback loop in which standardized experimental measurements generate high-quality datasets that improve computational prediction, while increasingly accurate AI models reduce the number of antibody variants requiring experimental evaluation.

For antibody discovery programs, this approach offers the potential to shorten optimization timelines, reduce screening costs, and prioritize candidates with improved developability characteristics earlier in the pipeline.

Equally important, the study emphasizes that successful AI applications depend fundamentally on the quality of the experimental data used for training. Standardized assays supported by reliable reference reagents remain essential for producing predictive models capable of guiding therapeutic antibody engineering.

As the field continues to move toward data-driven antibody design, studies such as this illustrate how rigorous experimental workflows and trusted research reagents together provide the foundation for next-generation developability platforms.

Citation: Saunders M, Comyn S, Medjo B, et al. Efficient inference of non-polyreactive antibody variants dependent on local fine-tuning. mAbs. 2026;18(1):2692763.