Built for AI Antibody Design. Accelerating the DBTL Cycle.

AI antibody design generates candidates faster than wet labs can validate them, creating a bottleneck in the Design-Build-Test-Learn (DBTL) cycle. Teams can model massive in silico libraries, but lack a rapid workflow to generate the physical evidence to classify which sequences express, bind, or fail.

This creates a bottleneck in the DBTL cycle. The Design stage is increasingly scalable, but the Build and Test stages remain constrained by mammalian expression timelines, purification capacity, analytical queue times, and downstream assay costs.

For AI-driven antibody discovery, the wet lab should not be a slow checkpoint at the end of a design round. It should provide timely, reproducible experimental data that helps teams refine models, prioritize candidates, and reduce unnecessary downstream work.

This guide explains the core technologies used in AI antibody design, why AI-generated antibodies can fail during wet-lab validation, why mammalian-first workflows slow the DBTL cycle, and how cell-free antibody screening can provide an early experimental triage layer before CHO scale-up.

What Are The Core Technologies of AI-Driven Antibody Discovery?

AI-driven antibody discovery uses a growing set of computational tools to generate, score, and prioritize candidate sequences. Moving beyond the constraints of natural evolution, modern computational pipelines can help researchers explore larger regions of antibody design space and focus experimental effort on more promising designs, but they do not remove the need for physical validation.

Modern computational antibody design pipelines commonly combine four technology areas.

	Generative models for de novo antibody design1 Deep neural networks (including diffusion frameworks like RFdiffusion) propose entirely novel binding scaffolds and CDR loops from scratch. These approaches expand the pool of candidates that can be considered before experimental screening and can help researchers explore design space beyond naturally sampled antibody repertoires.
	Protein language models (PLMs)2 Architectures like ESM-2 extract statistical patterns from massive immune-repertoire datasets to score sequences, optimize humanness profiles, and guide sequence-level engineering without requiring dense structural data.
	Deep learning structure prediction1,3 Advanced tools, for example AlphaFold 3™ modeling, generate rapid 3D visualizations of antibody-antigen interactions to streamline candidate prioritization, though these theoretical docking interfaces still require empirical wet-lab validation.
	Developability assessments1,4 Machine learning screeners flag sequence-level liabilities, such as aggregation, polyreactivity, or poor solubility, long before physical synthesis begins. These predictions can help teams deprioritize risky candidates earlier before downstream bioprocessing.

The next frontier: connected AI workflows

As AI antibody discovery matures, teams are beginning to connect multiple computational tools into more coordinated workflows^5–7. Generative design, structure prediction, developability assessment, and experiment planning increasingly operate together rather than as isolated steps.

Computational coordination can help teams generate, model, and prioritize candidates to Build and Test next, but does not remove the need for physical validation. It increases the importance of timely, reproducible experimental data.

Table 1. The operational roles of connected AI tools and their wet-lab validation requirements.

Computational tool or workflow	Role in AI antibody discovery	Why wet-lab feedback is critical
Generative design models	Proposes new antibody sequences or CDR designs from scratch.	Proposed sequences must still express and fold as physical molecules and show measurable target binding in vitro.
Structure-prediction tools	Model 3D structures and possible antibody-antigen interactions.	Predicted interfaces must be confirmed experimentally.
Developability agent	Flag potential aggregation, solubility, expression, or polyreactivity risks.	Predicted structural liabilities require rapid, empirical confirmation.
Experiment-planning workflows	Prioritize which candidates should be built and tested next.	Prioritization improves when positive and negative experimental outcomes are returned quickly.

Why Can Antibody Designs Fail in Wet-Lab Validation?

While generative architectures expand discovery bounds, hit-to-lead validation success rates remain highly variable depending on the target and model used⁸. For de novo antibody designs hit success rates can drop below 20%^9–13.

When computationally generated candidates move into the wet lab, failure can occur for three main reasons.

1. Expression and developability failures: A sequence that appears stable in an in silico snapshot may fail to express, fold, assemble, or remain soluble as full-length antibodies. Function relies on dynamic conformational states; ignoring entropic penalties can result in a molecule that aggregates or degrades in solution. Developability issues like severe aggregation, poor solubility, or extreme viscosity have historically led to a 40% candidate attrition rate.¹⁵
2. Binding failures: Even when a candidate expresses, the predicted antibody-antigen interface may not translate into measurable binding. Benchmarking of advanced architectures like AlphaFold 3 reveals up to a 60% failure rate in accurately predicting antibody-antigen docking complexes¹⁶. Protein flexibility, antigen presentation, assay conditions, and conformational dynamics can all affect whether a computational design binds in practice.
3. Data-quality failure: Slow or inconsistent validation workflows can produce data too late, too sparsely, or too variably to support model learning. If results are generated across fragmented workflows, it becomes harder to determine whether a candidate failed due to poor digital design or faulty assay execution¹⁷.

For AI-driven antibody discovery, reproducible Build and Test data are central to the Learn phase of the DBTL cycle to refine models and guide the next design round.

What is the AI Antibody DBTL Cycle?

The DBTL cycle is the iterative framework used to connect computational design with experimental feedback.

In AI antibody discovery, the DBTL cycle bridges the digital and physical worlds to improve sequence generation with every cycle.

• Design: Generative AI and language models navigate vast sequence-fitness landscapes to generate, score, and prioritize zero-shot antibody candidate sequences. These tools help optimize CDR loops and modeling target-specific binding interfaces in silico.
• Build: Selected digital sequences must be converted into physical molecules. Typically this process includes DNA preparation, construct assembly, and antibody expression.
• Test: Physical antibody variants are experimentally characterized. Early testing may include expression status, target binding, kinetic measurements (association rate, dissociation rate, and affinity), and early antibody developability assessments.
• Learn: High-dimensional experimental data, crucially including both positive hits and negative failures, are fed back into machine learning architectures to refine algorithmic weights and guide the next design round.

To reduce antibody design failures within the wet lab, we must continuously train and refine our models, which is exactly where the Learn phase of the DBTL cycle becomes critical. In an ideal engineering loop, you Design in silico, Build the physical molecules, Test their attributes, and Learn from the data to optimize the next generation.

The Data Starvation Problem

The promise of the DBTL cycle depends on continuous experimental feedback. AI models improve when they receive enough high-quality data to learn from both success and failure.

The mammalian validation wall (speed & cost caps)

Mammalian expression and downstream characterization are powerful and necessary, but they are expensive and capacity-constrained. When large in silico libraries are passed directly into mammalian expression, teams may spend weeks and substantial budget testing candidates that ultimately do not bind or express.

Basic hit validation typically takes 4 to 8+ weeks and requires specialist equipment and reagents^18–21. Furthermore, a cost of $400 to $600 per antibody can force teams to make blind, arbitrary library cuts, taking only a fraction of their computational designs forward into physical validation.

The missing negative data

When experimental validation is slow and expensive, teams naturally focus on candidates most likely to succeed. That can leave many non-binders and non-expressors unmeasured.

For machine learning, this is a problem. Negative data are not simply failed experiments. They help models learn which sequence features, design strategies, or predicted interfaces are less likely to translate into useful physical behavior. Training an AI exclusively on successes is like training a self-driving car only on flawless trips; without crash data, the algorithm cannot learn what to avoid. Negative data are not just failed experiments, they are critical training data.

Fragmented and proprietary datasets

AI antibody discovery depends on access to relevant experimental data, but many high-value datasets are proprietary, campaign-specific, or generated using different workflows. This means discovery teams often need to generate their own ground-truth data for each target, format, or design strategy.

Federated learning and other privacy-preserving approaches may help teams learn across distributed datasets, but their value still depends on the quality and comparability of the underlying experimental data. If expression, binding, and kinetic results are generated using different systems, assay formats, protocols, or reporting standards, it becomes harder to compare outcomes across campaigns.

Whether teams are building internal datasets or participating in collaborative data networks, the Build and Test stages need to generate reproducible, interpretable outputs. Without timely and comparable experimental data, the Learn stage remains weak and AI antibody design remains limited by validation capacity rather than sequence generation.

Restoring the DBTL Cycle with Cell-Free Antibody Screening

To unblock the DBTL bottleneck, pipelines require an upstream triage step to physically validate candidates faster than mammalian biology allows. By improving the Build-Test-Learn part of the cycle we are able to generate improved designs.

Mammalian expression remains essential for downstream confirmation, developability assessment, functional assays, and manufacturing-relevant validation. But it is not always the most efficient first screen for every computationally generated sequence.

Cell-free antibody screening helps answer the first practical questions earlier:

1. Can the candidate be expressed as a full-length IgG?
2. Does the expressed antibody show binding to the intended antigen?
3. Which candidates should be prioritized for deeper characterization?

By moving these early questions upstream, teams can reduce the number of non-binding candidates that enter mammalian workflows and focus downstream resources on candidates with experimental evidence of expression and binding.

A cell-free triage workflow can support AI antibody discovery in four ways.

• Faster Build and Test: Cell-free protein synthesis bypasses the time required for mammalian cell culture and enables rapid expression of antibody variants for screening. This significantly reduces the time and cost required to generate positive and negative data for faster model feedback.
• Useful model feedback: By utilizing cell-free expression and binding assays, you can rapidly identify non-expressors, non-binders, and lower-confidence candidates, before they consume premium mammalian downstream capacity. Because the workflow returns both positive and negative outcomes, AI teams can use the data to refine subsequent antibody design rounds. Expression status, binder/non-binder classification, and kinetic ranking all provide useful experimental feedback.
• Consistent early triage data: Decoupled from the noisy, living variables of host cells, CFPS produces exceptionally clean, standardized, machine-readable output optimized for ML ingestion. As AI models become more sophisticated, access to high-quality experimental data becomes increasingly important. In our internal studies, we’ve observed replicate CVs of approximately 3% from SPR kinetic profiling data of a cell-free produced full-length IgG dataset.
• Highly Predictive of CHO Performance: A common concern with cell-free antibody screening is whether material produced outside mammalian cells can support reliable early binding decisions. Nuclera’s internal validation data show that cell-free-produced full-length IgG antibodies generate binding profiles that closely track CHO-derived controls for early hit triage.

Although CHO expression remains important for downstream confirmation, developability assessment, functional assays, and manufacturing-relevant decision-making, these data support the use of cell-free-produced antibodies as an early screening material for binding-driven triage. This allows teams to identify likely binders and non-binders before committing every candidate to mammalian expression.

Recent commentary in Nature Communications highlights that cell-free systems serve as the premier platform for megascale data generation, seamlessly integrating machine learning with protein engineering by screening hundreds of thousands of variants for thermodynamic stability (ΔG) and activity¹⁷.

Cell-free vs CHO Validation: What Should Happen When?

Cell-free antibody screening is best suited for early triage. It helps teams determine which candidates express as full-length IgG, which show binding to the target antigen, and which should be prioritized for further analysis.

CHO expression remains important for downstream confirmation, functional testing, developability assessment, and manufacturing-relevant decision-making.

The practical question is not whether CHO validation matters. It does. The question is whether every computationally generated candidate should enter CHO workflows before there is experimental evidence of expression and binding.

For large AI-generated libraries, a triage-first workflow can help teams use each system more efficiently:

• Use cell-free validation to triage broad in silico libraries.
• Use CHO validation to confirm and develop candidates that pass triage.

How Nuclera Antibody Services supports AI antibody discovery

Nuclera’s Cell-Free Antibody Screening Service provides an early triage layer for AI-generated antibody libraries. We convert large AI-generated libraries into decision-grade binding data from 14 business days.

Customers provide VH/VL sequences and antigen. Nuclera performs DNA preparation, 96-plex full-length IgG cell-free expression and binding screening, and focused SPR on prioritized binders. The output is a data package that helps teams determine which candidates should progress for further downstream validation.

For AI/ML teams, the value is faster feedback with both positive and negative experimental data to refine models. For discovery teams, the value is better prioritization. For downstream teams, the value is fewer unsupported candidates entering high-cost validation workflows.

Accelerate your DBTL loop, feed your models the data they need, and turn the wet lab into your greatest competitive advantage.

Table 2. Mammalian-first validation of AI-generated antibody libraries vs. Nuclera cell-free antibody triage.

Metric	CRO / Mammalian Workflow	Nuclera Cell-Free Antibody Services
Speed to data	Typically 4 to 8+ weeks for expression and binding data.	From 14 business days.
Cost per variant	$400–$600 per antibody for expression and binding data.	$100* per antibody for DNA synthesis, 96-plex ELISA-based expression and binding screening, plus SPR on five focused binders.
Primary role	Downstream confirmation, developability, and manufacturing validation.	Early hit validation. ELISA-based expression and binding assessment. SPR kinetics.
DBTL Impact	Slow loops; negative data are under-collected or delayed.	Restores the flywheel with rapid positive and negative experimental data.

Timelines and costs determined by surveys from 3 CROs. *Introductory pricing.

References

1. Vecchietti, L. F., et al. Artificial intelligence-driven computational methods for antibody design and optimization. mAbs 17, 2528902 (2025). https://doi.org/10.1080/19420862.2025.2528902
2. Chungyoun, M. F. & Gray, J. J. AI models for protein design are driving antibody engineering. Curr. Opin. Biomed. Eng. 28, 100473 (2023). https://doi.org/10.1016/j.cobme.2023.100473
3. Dell’uomo, D. et al. HyperBind2: Multi-Shot Learning Enables Progressive Improvement in Computational Antibody Discovery. bioRxiv preprint https://doi.org/10.1101/2025.11.06.687005 (2025).
4. Parkinson, J. & Wang, W. For antibody sequence generative modeling, mixture models may be all you need. Bioinformatics 40, btae278 (2024). https://doi.org/10.1093/bioinformatics/btae278
5. Ock, J. et al. sdAbs-LLM: Generative Large Language Models For de novo Antibody Design and Agentic Evaluation. bioRxiv preprint arXiv:2026.04.18.716776v1 (2026). https://doi.org/10.64898/2026.04.18.716776
6. Rapp, J. T., Bremer, B. J. & Romero, P. A. Self-driving laboratories to autonomously navigate the protein fitness landscape. Nat. Chem. Eng. 1, 97–107 (2024). https://doi.org/10.1038/s44286-023-00002-4
7. IBM Research. Agentic AI for screening antibody:antigen binding affinity. ACS Spring 2026 Poster Abstract (2026). Available here.
8. Kosonocky, C. W., et al. Closing the loop: Experimentally validated methods in artificial intelligence-driven protein design. Current Opinion in Structural Biology 98, 103272 (2026). https://doi.org/10.1016/j.sbi.2026.103272
9. Levine, S. et al. Origin-1: a generative AI platform for de novo antibody design against novel epitopes. Preprint at bioRxiv https://doi.org/10.64898/2026.01.14.699389 (2026).
10. Biswas, S. De novo design of epitope-specific antibodies against soluble and multipass membrane proteins with high specificity, developability, and function. Preprint at bioRxiv https://doi.org/10.1101/2025.01.21.633066 (2025).
11. Tang, J. Rapid De Novo Antibody Design with GeoFlow-V3. Preprint at bioRxiv https://doi.org/10.1101/2025.10.20.682964 (2025).
12. Boitreaud, J. et al. Zero-shot antibody design in a 24-well plate. Preprint at bioRxiv https://doi.org/10.1101/2025.07.05.663018 (2025).
13. Dell’uomo, D., Satz, A. & Averso, B. HyperBind2: Multi-Shot Learning Enables Progressive Improvement in Computational Antibody Discovery. Preprint at bioRxiv https://doi.org/10.1101/2025.11.06.687005 (2025).
14. Ock, J. et al. sdAbs-LLM: Generative Large Language Models For de novo Antibody Design and Agentic Evaluation. bioRxiv preprint https://doi.org/10.64898/2026.04.18.716776 (2026).
15. Jain, T. et al. Biophysical properties of the clinical-stage antibody landscape. Proc. Natl. Acad. Sci. U.S.A. 114, 944–949 (2017). https://doi.org/10.1073/pnas.1616408114
16. Hitawala, F., Gray, J. What does AlphaFold3 learn about antigen and nanobody docking, and what remains unsolved? mAbs. 17, 2545601 (2025).
17. Voyvodic, P. L. et al. LDBT instead of DBTL: combining machine learning and rapid cell-free testing. Nat. Commun. 16, 9782 (2025). https://doi.org/10.1080/19420862.2025.2545601
18. Moldovan Loomis, C. et al. AI-based antibody discovery platform identifies novel, diverse, and pharmacologically active therapeutic antibodies against multiple SARS-CoV-2 strains. Antibody Ther. 7, 307–323 (2024). https://doi.org/10.1093/abt/tbae025
19. Mullen, T. E. et al. Accelerated antibody discovery targeting the SARS-CoV-2 spike protein for COVID-19 therapeutic potential. Antibody Ther. 4, 185–196 (2021). https://doi.org/10.1093/abt/tbab018
20. Kelley, B. Developing therapeutic monoclonal antibodies at pandemic pace. Nat. Biotechnol. 38, 540–545 (2020). https://doi.org/10.1038/s41587-020-0512-5
21. Rodriguez-Conde, S. et al. Suitability of transiently expressed antibodies for clinical studies: product quality consistency at different production scales. mAbs 14, 2052228 (2022). https://doi.org/10.1080/19420862.2022.2052228

ALPHAFOLD is a trademark of DeepMind Technologies Limited.