Stacey Willard, Vice President of Marketing and Product Management at Nuclera, explores the validation bottleneck for antibody engineering and why generative AI may be the answer to this challenge. 

Antibody engineering has developed a pronounced asymmetry. AI can now generate and rank antibody candidates at a pace that conventional experimental workflows struggle to match. Diffusion models, protein language models and structure predictors can propose and rank thousands of target-specific variants in the time it once took to plan an immunisation campaign. The bottleneck is increasingly not designing what to test but physically producing enough candidates to discover which designs actually work. A designed antibody must still become a molecule before validation can begin, and for most laboratories that means cloning, transfection, culture, harvest and purification: a chain of operations paced by the doubling time of a mammalian cell, not by the algorithm that proposed the candidate.

 Why mammalian expression slows antibody validation

Mammalian cell-based validation is a major reason experimental testing cannot keep pace with design. Moving from sequence to a characterised antibody typically takes four to eight weeks, with throughput constrained by transfection capacity, expression cycles, purification and biosensor availability.

Generative architectures expand the accessible design space, but wet-lab success rates remain highly variable by target and model. Diffusion-based methods report in vitro success near one in five for designed miniprotein binders, while zero-shot antibody generation has historically fallen far below that^1,2_.

Compounded, variable hit rates and labour-intensive validation mean libraries are cut before sequences reach the wet lab, or limited budget is spent on sequences that do not express or bind. A group holding 500 designs and the capacity to characterise 30 discards most of them on prediction alone, losing viable candidates without record.

Yet the first questions when validating AI-generated antibody candidates are not difficult ones: does this sequence express as an intact IgG? Does it engage the antigen? And does it warrant more resource-intensive mammalian expression? These questions do not necessarily require the same experimental infrastructure as later-stage assessment of effector function, developability or manufacturability. Cell-free expression offers a way to move these early questions closer to the point of sequence generation.

The implication is important – rather than selecting a small number of candidates for mammalian expression before experimental evidence exists, researchers can introduce a rapid physical screening step earlier in the workflow.

How cell-free protein synthesis works

Cell-free protein synthesis (CFPS) removes the host and keeps its machinery³. A reaction combines the transcription and translation apparatus, supplied as a cell extract or assembled from purified parts, with a supply of energy and building blocks. The chemical environment is tuned by the experimenter and can be adjusted while synthesis is still running. Templates can be PCR products rather than plasmids, removing cloning altogether, so the time taken to move from sequence file to testable protein can happen in hours.

CFPS is well established for fragment workflows including scFv, Fab and VHH, which carry few disulfides and express readily in bacterial extracts. Published workflows have combined template generation, expression and binding measurement for hundreds of fragments in a single day, including every antibody then under emergency authorisation for Covid-19⁴. A fragment hit must still be reformatted as an IgG before development begins, and reformatting alters valency and sometimes apparent affinity.

The efficiency gain can be significant. A 96-variant panel moves from DNA to binding classification within a working week, rather than requiring four to eight weeks for a mammalian route. At a lower cost per variant, the entire panel can be screened rather than a predicted subset. The selection decision no longer has to be made before experimental evidence exists.

Applications of CFPS in antibody engineering

A human IgG1 contains 16 disulfide bonds, including the links holding two heavy and two light chains in the correct arrangement, which must form between the right partners. The bacterial cytoplasm is chemically reducing, which works directly against disulfide formation. A cell-free reaction, by contrast, is chemically tuneable, and engineered systems now assemble full-length IgG with correct chain pairing routinely, giving monodisperse size-exclusion profiles and the expected banding under reducing and non-reducing SDS-PAGE.

Bacterial cell-free systems do not glycosylate, which raises the recurring question: can cell-free antibodies report the same behaviour as mammalian-produced material?

The conserved N297 glycan in CH2 is required for Fcγ receptor and C1q engagement, so an aglycosylated IgG1 is effectively effector-null⁵. Antigen recognition is Fv-mediated and is often preserved in the absence of Fc glycosylation. For many early-stage binding assessments, the absence of Fc glycosylation does not prevent meaningful comparison of antigen-binding behaviour, even though it becomes critical for later-stage antibody biology. A minority of antibodies carry N-X-S/T sequons in the variable domains, where the glycan can modulate stability and occasionally affinity⁶.6 In the large majority of cases the equivalence holds.

Cell-free antibodies should therefore not be considered universally equivalent to their mammalian-produced counterparts. Rather, they can provide a rapid and informative readout for early questions such as expression and antigen binding, while mammalian systems remain necessary for assessing glycan-dependent biology, effector function and later-stage developability.

A campaign is bound by antigen quality as much as by antibody supply, and the targets where antibodies are most wanted are often hardest to present correctly. Multipass membrane proteins, ion channels and GPCRs lose native conformation the moment they leave a lipid environment. CFPS can also supply these antigens, with nanodiscs added to the reaction so a membrane antigen inserts co-translationally into a defined bilayer rather than aggregating. Antigen variants for epitope mapping, or orthologue panels for cross-reactivity, can be made in parallel on the same footing.

Why automation determines data quality

What turns cell-free expression into a screening method is the automation around it: standardised reaction assembly, liquid handling, and readouts can be collected under identical conditions across a whole panel. Removing cloning, culture and purification eliminates most of the steps at which preparations drift apart, making replicate variation in the low single figures achievable.

The same property addresses a second, less visible problem. Because characterisation is expensive, teams test only the highest-ranked candidates. Therefore, the failures that would teach a model which designs to avoid go unrecorded. A screen with a low enough cost per variant to run across an entire panel returns the complete result, expressed or not and binding or not, which is the dataset closed-loop programmes depend on.

This creates an important feedback loop. Generative models do not only need successful examples but they also need information on which sequences fail to express, bind, or behave unexpectedly. Lowering the cost and time required to generate the data therefore has implications beyond candidate selection – it can improve the quality of the data used to guide the next design cycle.

What cell-free expression cannot answer

The goal is not to displace mammalian validation but to make it more efficient. Effector function, glycan-dependent activity, behaviour under formulation stress and manufacturability belong to material made in the mammalian system that will make the drug.

Generative design creates candidate volume. Cell-free expression provides a physical testing process fast enough to evaluate that volume, while automation makes the resulting data consistent enough to support confident decisions and iterative learning. Together these allow the wet laboratory to return evidence closer to the rate at which algorithms generate candidates, reserving mammalian capacity for candidates already supported by experimental results.

If antibody engineering is to realise the potential of generative design, the industry will need to rethink not only how candidates are generated, but when and how they are experimentally tested. Closing that validation gap may be less about replacing existing workflows than about putting the right experiment at the right point in the design cycle.

References

1. Watson, J. L. et al. De novo design of protein structure and function with RFdiffusion. Nature 620, 1089–1100 (2023).
2. Boitreaud, J. et al. Zero-shot antibody design in a 24-well plate. Preprint at https://doi.org/10.1101/2025.07.05.663018 (2025).
3. Silverman, A. D., Karim, A. S. & Jewett, M. C. Cell-free gene expression: an expanded repertoire of applications. Nat. Rev. Genet. 21, 151–170 (2020).
4. Hunt, A. C. et al. A rapid cell-free expression and screening platform for antibody discovery. Nat. Commun. 14, 3897 (2023).
5. 5.Mimura, Y. et al. Glycosylation engineering of therapeutic IgG antibodies: challenges for the safety, functionality and efficacy. Protein Cell 9, 47–62 (2018).
6. van de Bovenkamp, F. S. et al. Variable domain N-linked glycans acquired during antigen-specific immune responses can contribute to immunoglobulin G antibody stability. Front. Immunol. 9, 740 (2018).

Author biography

Stacey Willard is Vice President of Marketing and Product Management at Nuclera, Cambridge, UK, where she leads product strategy for the company’s protein-access technologies. Her background spans life-science tools, protein biology workflows and translational research, with experience turning emerging technical capability into working products.