As AI models become more involved in antibody discovery and design, the need for reliable experimental data is growing. Turning predicted sequences into structured, comparable assay results is an important step towards connecting computational predictions with laboratory evidence.
Antibody design is increasingly shaped by computational methods. Protein language models, machine learning algorithms and generative AI tools are helping researchers explore sequence space, predict molecular properties and propose new antibody candidates. These approaches can narrow the field of possibilities, but their predictions still need to be tested experimentally.
That creates a practical challenge. How do you turn thousands of proposed antibody sequences into meaningful experimental data that can be used to evaluate and improve computational models?
The answer depends on more than producing antibodies at scale. It requires a workflow that connects sequence synthesis, expression and characterization, while generating consistent data that can be interpreted across large candidate sets.
AI models are only as useful as the data available to develop and evaluate them. In antibody discovery, computational models may predict binding affinity, expression behaviour, or other molecular properties based on sequence information. However, these predictions do not always translate directly into experimental performance.
A sequence predicted to bind a particular target may show weak binding in an assay. Another candidate may exhibit promising affinity but prove difficult to express. Differences between predicted and observed properties can reveal limitations in a model and help guide subsequent design decisions.
Experimental validation provides the reference point for these comparisons.
For AI-driven antibody programmes, this means generating datasets that connect antibody sequences with measured experimental outcomes. Rather than relying on isolated assay results, researchers need information collected across sufficiently large candidate sets, using consistent experimental conditions and clearly defined measurements.
The quality of this data matters as much as its quantity. Inconsistent expression conditions, differences in assay formats or incomplete sequence information can make comparisons difficult. A large dataset is not automatically a useful one.1
Building a model-ready antibody dataset involves several connected stages.
It begins with sequence selection. Researchers may submit candidates generated through computational design, sequence optimisation or existing antibody libraries. These sequences are then synthesised and expressed to produce material for downstream testing.
Expression is an important consideration because not every designed sequence produces a usable antibody. Yield, purity and other expression-related properties can vary between candidates. Recording these outcomes alongside sequence information helps researchers understand not only which antibodies perform well, but also which sequences present practical production challenges.
Once expressed, antibodies can be evaluated using relevant analytical and functional assays.
For example, surface-based binding techniques such as biolayer interferometry (BLI) and surface plasmon resonance (SPR) can provide information about antibody-target interactions, including kinetic parameters and binding affinity. Other assays can assess different aspects of molecular behaviour.
Depending on the research objectives, characterization may include thermal stability measurements using differential scanning fluorimetry (DSF), as well as aggregation or self-association assessments using methods such as AC-SINS.
Each assay contributes a different piece of information. Together, these measurements can provide a broader experimental profile of an antibody candidate.
The resulting dataset should retain the connection between each antibody's sequence, expression results, and characterization measurements. This allows researchers to compare candidates systematically and examine relationships between sequence features and experimentally observed properties.2
As computational approaches generate larger numbers of candidate sequences, laboratory validation can become a bottleneck.
Testing a handful of antibodies may be sufficient during early feasibility studies. However, evaluating larger candidate libraries requires workflows that can handle substantial sample numbers without sacrificing experimental consistency.
High-throughput expression and characterization help address this challenge. Automated or parallelised workflows can process many candidates under defined conditions, reducing the manual effort involved in moving between individual samples.
Capacity also affects how quickly researchers can evaluate new design rounds. When computational teams generate successive batches of sequences, delays in experimental validation can slow the feedback process.
Biointron's RushDATA™ platform is designed to support high-throughput antibody production and characterization, with processing capacity of 10,000+ antibodies per day. Depending on the selected expression workflow, the sequence-to-data turnaround can be as fast as 2 days with cell-free expression or 3 days with CHO expression.
These workflows offer different expression approaches, allowing researchers to select an option suited to their experimental requirements. Cell-free expression can support rapid initial evaluation, while CHO expression provides a mammalian cell-based production route.
The appropriate choice depends on the study design, intended measurements, and the level of biological relevance required from the expression system.
Generating experimental measurements is only part of the process. How those results are organised and delivered determines how easily they can be used in subsequent analysis.
For computational researchers, model-ready data should be structured so that individual sequences can be linked to their corresponding experimental results. This includes retaining relevant sample identifiers, assay conditions, measurement units, and quality information.
Clear data organisation makes it easier to compare candidates across experimental batches and identify potential sources of variation. It also supports downstream analysis, including the evaluation of prediction accuracy, identification of sequence-property relationships and selection of candidates for further optimisation.
For example, a research team developing an antibody affinity prediction model could compare predicted binding values with experimentally measured results across a candidate library. Researchers could then examine where predictions align with observations and where discrepancies occur.
Similar analyses can help investigate the relationship between sequence characteristics and expression or stability measurements.
Importantly, experimental datasets do not automatically produce better models. Their usefulness depends on assay quality, dataset composition, appropriate controls and how well the measurements reflect the properties being studied. Researchers also need to consider differences between experimental systems when combining data from multiple sources.
The growing use of AI in antibody discovery is changing what researchers need from laboratory partners. Beyond producing individual antibody candidates, there is increasing demand for experimental workflows capable of evaluating large sequence sets and returning structured results within practical timelines.
Connecting sequence design with high-throughput expression and characterization gives computational teams a more direct way to test their predictions against laboratory evidence.
With its high-throughput capacity and rapid expression workflows, Biointron's RushDATA™ service supports this connection, helping researchers move from antibody sequences to experimental datasets at scale.
As antibody design models continue to develop, reliable experimental validation will remain an essential part of the process. The goal is not simply to generate more data, but to generate data that researchers can interpret, compare, and use to make better-informed decisions about the next generation of antibody candidates.
Liu, Y. T., Zhang, L. L., Jiang, Z. Y., Tian, X. S., Li, P. L., Wu, P. H., Du, W. T., Yuan, B. Y., Xie, C., Bu, G. L., Zhong, L. Y., Yang, Y. L., Li, T., Zeng, M. S., & Sun, C. (2025). Applications of Artificial Intelligence in Biotech Drug Discovery and Product Development. MedComm, 6(8), e70317. https://doi.org/10.1002/mco2.70317
Kumar, A., Kumar, P., Singh, N., Singh, P. K., & Singh, A. (2026). Artificial intelligence across the drug discovery pipeline: From data integration to precision therapeutics. The Journal of Precision Medicine: Health and Disease, 8, 100061. https://doi.org/10.1016/j.premed.2026.100061
Recap of Antibodies After Hours Boston 2026: IPI's DELPHI developability predict……
Key trends from the 17th World Bispecific Summit 2026 in Boston: AI-guided desig……
Biointron’s Antibodies After Hours: From Algorithms to Antibodies was held at Th……
You can also contact us on the Scientist and Science Exchange marketplaces.