Artificial intelligence (AI) is changing how therapeutic antibodies are discovered and developed. Advances in machine learning, protein structure prediction, and generative AI have made it possible to design and evaluate large numbers of antibody sequences computationally, expanding the range of candidates researchers can explore.
Yet designing a promising antibody sequence is only the beginning. An antibody predicted to bind strongly to its target may perform differently when produced and tested in the laboratory. Its actual binding affinity, expression level, stability, and other properties must be determined experimentally before it can advance through development.
This gap between computational predictions and experimental results has become an important challenge in AI-driven antibody discovery. As computational design becomes faster and more scalable, experimental validation must keep pace. Increasingly, researchers are turning toward integrated workflows that combine AI-based design with high-throughput laboratory testing, allowing experimental results to inform subsequent rounds of optimization.
Traditional antibody discovery relies on established approaches such as hybridoma technology, phage display, and single B cell screening to identify antibodies with desirable binding properties. These methods have contributed to the development of numerous therapeutic antibodies, but identifying and optimizing candidates can require extensive screening and multiple rounds of laboratory work.
AI introduces a complementary approach. Rather than relying exclusively on experimental screening to explore antibody diversity, computational models can analyze existing sequence, structural, and experimental datasets to predict potentially useful antibody properties and generate new candidates.
Recent developments include protein language models trained on antibody sequences, deep learning methods for predicting antibody structures, and generative models that propose new amino acid sequences with specified characteristics.
A recent review describes how these methods are being applied across antibody development, including sequence design, epitope–paratope interaction prediction, affinity optimization, and developability assessment.1
These capabilities can help researchers prioritize candidates before committing resources to laboratory testing.
One important application of AI is the prediction of antibody structure and interactions with target antigens.
Antibody binding is largely determined by the complementarity-determining regions (CDRs), which form much of the antigen-binding surface. Computational models can predict how these regions adopt three-dimensional structures and interact with a particular epitope.
Generative AI extends this capability by proposing new antibody sequences rather than analyzing only existing ones. Depending on the model, researchers can explore sequence variants, optimize specific CDR regions, or design candidates against a target structure.
AI models can also support early developability assessment by identifying sequence characteristics associated with aggregation, poor solubility, or other potential liabilities.
However, computational predictions remain approximations of biological behavior. Antibody–antigen interactions depend on molecular flexibility, solvent conditions, and structural changes that may not be fully represented in computational models.
Consequently, a promising computational design must still demonstrate its predicted properties experimentally.
The growing ability to generate antibody sequences has introduced a new bottleneck: testing them.
AI models can produce hundreds or thousands of candidate sequences in a design cycle, but each candidate requires experimental evaluation to establish whether it behaves as predicted. This typically involves recombinant antibody expression followed by binding and biophysical characterization.
Unlike computational screening, these steps require physical materials, laboratory equipment, and carefully controlled assays. When performed separately or in small batches, they can slow the transition from computational design to experimentally supported candidates.
The challenge is not simply producing antibodies more quickly. It is generating enough reliable experimental data to determine which predictions are correct, where models fail, and how subsequent designs should be improved.
In another recent review published in Cancer Medicine, Ören Varol and Varol identify insufficient experimental validation as a significant limitation in AI-driven therapeutic discovery.2 They note that models trained on incomplete or biased datasets may produce predictions that appear convincing computationally but perform poorly in biological systems. Thus, computational predictions are useful hypotheses, but experimental measurements are needed to establish their biological validity.
A candidate antibody may receive a favorable computational binding score yet fail to demonstrate the expected interaction in vitro.
Several factors can explain discrepancies between predictions and experimental outcomes. Protein structure models may not fully capture conformational flexibility, while predicted antibody–antigen interfaces may differ from those observed under experimental conditions.
Even when an antibody binds its intended target, it may exhibit other undesirable characteristics that limit its suitability for development.
These include:
Low expression yield: The designed antibody may be difficult to produce efficiently in recombinant expression systems.
Poor stability: The molecule may be susceptible to unfolding or degradation under relevant conditions.
Aggregation: Antibody molecules may form aggregates that complicate manufacturing and formulation.
Nonspecific binding: A candidate may interact with unintended proteins or biological components.
Unfavorable biophysical properties: Poor solubility, self-interaction, or other developability liabilities may affect downstream development.
Ammar et al. (2026) emphasize that optimizing a single antibody property does not necessarily improve overall developability. For example, sequence changes intended to enhance affinity may introduce unwanted effects on stability, solubility, or expression.1
This makes experimental characterization important not only for confirming target binding, but also for evaluating whether an antibody possesses a suitable combination of properties for further development.
Experimental validation provides the measurements needed to compare computationally designed candidates and assess their suitability for further development.
The appropriate assays depend on the stage of discovery, the antibody format, and the objectives of the computational model.
Before an AI-designed antibody can be characterized, its sequence must be converted into a physical protein.
Researchers typically obtain DNA encoding the antibody heavy and light chains, incorporate the sequences into expression constructs, and introduce them into a suitable production system, such as Chinese hamster ovary (CHO) or HEK293 cells.
Measuring recombinant expression provides an early indication of whether a designed sequence can be produced efficiently. Low expression may also reveal limitations that were not apparent during computational design.
Binding assays establish whether an antibody interacts with its intended antigen and help quantify the interaction.
Enzyme-linked immunosorbent assays (ELISA) can provide initial binding information, while techniques such as biolayer interferometry (BLI) and surface plasmon resonance (SPR) can characterize binding kinetics.
These measurements may include the association rate (k_on), dissociation rate (k_off), and equilibrium dissociation constant (K_D).
Such results are particularly valuable when evaluating AI models trained to predict antibody affinity or optimize antigen-binding interactions.
Importantly, strong target binding does not automatically imply favorable specificity or biological activity. Additional experiments may be required to assess cross-reactivity, epitope recognition, and functional effects.
Antibody candidates must also possess suitable biophysical properties to support subsequent development.
Differential scanning fluorimetry (DSF), for example, can provide information about thermal stability. Other assays evaluate polyreactivity, self-interaction, and aggregation-related risks.
These measurements help researchers identify potential liabilities before investing in more extensive development.
For AI-driven antibody engineering, combining binding data with expression and developability measurements also enables models to optimize several properties rather than focusing exclusively on affinity.
A promising direction in AI antibody discovery is the development of closed-loop workflows, in which computational design and laboratory testing operate as an iterative process.
Instead of treating experimental validation as the final step, researchers use experimental results to guide subsequent computational decisions.
This approach is can be described as a Design-Build-Test-Learn (DBTL) cycle.
Design: AI models generate or prioritize antibody sequences based on predicted structural, binding, and developability properties.
Build: Selected sequences are synthesized and expressed as recombinant antibodies.
Test: Laboratory assays measure binding, expression, stability, and other relevant characteristics.
Learn: Experimental results are analyzed and used to evaluate, recalibrate, or retrain computational models, informing the next design cycle.
The cycle can then be repeated as researchers refine their candidates.
Closed-loop discovery is an important strategy for improving the translational relevance of AI-driven therapeutics, as experimental feedback should be incorporated throughout discovery rather than used only to confirm final computational predictions. Larger and more diverse experimental datasets can support improvements in computational models.
Increasing experimental throughput can help researchers evaluate more antibody candidates, but larger datasets are not necessarily more informative.
For experimental results to support reliable model development, measurements should be generated under well-controlled conditions, with consistent assay procedures and clearly documented sample information.
Data should also remain linked to the corresponding antibody sequences and experimental conditions. Without this information, comparing candidates or integrating results into computational workflows becomes more difficult.
Negative results can be particularly informative. An antibody that fails to express, demonstrates weak binding, or exhibits undesirable developability properties can reveal limitations in a model's predictions.
However, these outcomes are often underrepresented in published datasets, potentially introducing bias into model development.2
Well-structured datasets that include both successful and unsuccessful candidates can therefore provide a more complete picture of the relationship between antibody sequence and experimental behavior.
As AI-based approaches expand the number of antibody candidates available for testing, experimental workflows must become more scalable.
High-throughput antibody expression and characterization can help address this challenge by enabling multiple candidates to be evaluated under coordinated experimental conditions.
Integrating gene synthesis, recombinant expression, binding assays, and early developability assessment into a single workflow can also reduce delays between experimental stages.
For researchers developing AI antibody design models, this integration offers another advantage: the ability to generate multidimensional datasets that connect individual sequences with their measured properties.
Biointron has developed RushData, an integrated high-throughput experimental platform designed to support AI-driven antibody discovery and optimization.
RushData combines gene synthesis, rapid CHO-based antibody expression, and experimental characterization to help researchers move from computationally designed sequences to measured antibody properties.
Built around Biointron's 1-day CHO expression workflow, the platform is designed to evaluate thousands of antibody candidates in parallel. Depending on project requirements, available characterization includes binding measurements and early developability assessments such as thermal stability, polyreactivity, and self-interaction.
Experimental results are organized into structured datasets that can support candidate comparison, computational analysis, and subsequent AI model development.
By coordinating antibody production and characterization, RushData addresses a practical challenge in AI antibody discovery: obtaining experimental feedback at a scale and speed that can support iterative design.
The value of this approach extends beyond faster candidate screening. Reliable experimental measurements allow researchers to examine where computational predictions agree with observed antibody behavior and where additional optimization is needed.
AI is expanding the possibilities for antibody design, allowing researchers to explore sequence diversity and evaluate molecular properties at a scale that would be difficult to achieve through experimental screening alone.
Yet the future of AI-driven antibody discovery will depend not only on increasingly sophisticated models, but also on the quality and accessibility of experimental data.
Computational design can help determine which antibodies to investigate. Laboratory testing establishes how those antibodies actually behave.
Bringing these capabilities together through closed-loop discovery may enable more efficient candidate selection, earlier identification of developability risks, and continuous refinement of antibody design models.
As the field advances, high-throughput experimental platforms that connect antibody sequences with reliable biological measurements will play an increasingly important role in translating computational predictions into validated antibody candidates.
Learn how Biointron's RushData platform supports AI-driven antibody discovery through rapid antibody expression, high-throughput characterization, and structured experimental data generation.
Ammar, M., Samsonov, M., Gurylina, E., & Bayzigitov, D. (2026). Artificial intelligence advancements in monoclonal antibody development technology. Frontiers in Immunology, 17, 1802038. https://doi.org/10.3389/fimmu.2026.1802038
Ören Varol, T., & Varol, M. (2026). Closing the translational gap: Closed-loop AI discovery frameworks for experimental validation and clinical implementation in cancer therapeutics. Cancer Medicine, 15, e72193. https://doi.org/10.1002/cam4.72193
Biointron announces limited-time RushData launch pricing for high-throughput ant……
Learn how direct antigen labeling improves single B cell antibody screening by r……
August 2026 brought another mix of acquisitions, licensing agreements, strategic……
You can also contact us on the Scientist and Science Exchange marketplaces.