Artificial intelligence can generate antibody sequences far faster than conventional expression workflows can test them. A fit-for-purpose cell-free triage layer can turn digital designs into early expression and binding evidence, while reserving mammalian systems for the questions only they can answer

Juhi Misra at Nuclera

Computational antibody discovery has crossed an important threshold. Generative models, structure-aware methods and sequence language models can now propose more candidates than most discovery teams can evaluate experimentally. The rate-limiting step is no longer only sequence generation; it is obtaining reliable physical evidence quickly enough to determine which candidates deserve scarce downstream capacity.

A recently published prospective benchmark makes the point. The AIntibody challenge tested 511 artificial intelligence (AI)-designed or AI-ranked antibodies from 29 organisations across affinity maturation, affinity ranking within HCDR3 clusters and out-of-library critical design review design. Several groups produced developable antibodies with sub-100pM affinity, but those successes were exceptions and did not transfer consistently across tasks.1 More broadly, de novo antibody design has delivered striking structural successes, yet the associated experimental campaigns have still required extensive screening and produced strongly target-dependent outcomes.2

These findings are not an argument against AI. They highlight why experimental capacity must be scaled alongside computational design. A digital score is a useful prioritisation signal, not physical proof that a sequence can be produced in the intended format or will bind the intended target. When a laboratory can test only a small fraction of a generated library, selection naturally favours candidates the model already considers promising or familiar. Less certain regions of sequence space may never be explored and the experimental data needed to improve the next generation of models is not available.

Mammalian expression remains the appropriate standard for answering many downstream questions, including Fc-mediated biology, glycan-dependent activity, cell-based potency and manufacturing-relevant behaviour. But it is a resource-intensive way to answer the narrower questions that arise immediately after computational design. Before committing a candidate to that workflow, a discovery team typically needs to know three things: was a valid expression template generated; was an apparently assembled antibody produced, and did it engage the target under defined assay conditions?

Together, these measurements place candidates into four useful categories. A candidate may express and bind; express but show no detectable binding; express poorly or assemble incorrectly; or return an indeterminate result near the assay threshold. Preserving these distinctions is important. An expression failure may reflect template quality, chain imbalance, folding or format incompatibility. By contrast, an expression-positive, binding-negative result provides more interpretable evidence about target engagement, although assay sensitivity, antigen presentation and format effects must still be considered. Treating every unsuccessful outcome as a non-binder collapses these biologically and technically different results into a single, unreliable label.

The first physical gate should be designed as a classification step, not as a compressed version of full lead characterisation. Its purpose is to separate clear binders, non-binders, expression or assembly failures and indeterminate results. This removes obvious failures from the downstream workflow, prevents technically ambiguous candidates from being rejected as biological negatives and routes a smaller, better-evidenced set into mammalian confirmation. Just as importantly, it produces structured experimental labels that can inform the next round of computational design.

A useful first physical gate must shorten the path from sequence to protein, operate across many candidates in parallel and generate interpretable expression and binding results. Enter cell-free protein synthesis (CFPS).

CFPS uses cellular transcription and translation machinery outside the constraints of a living host. Depending on the system, the reaction may use a crude cell lysate or purified components. Because the reaction environment is open, variables such as heavy-to-light-chain DNA ratio, redox conditions, salts and chaperones can be adjusted directly. Linear DNA can also replace plasmids, avoiding cloning and shortening the path from sequence to protein to better match the pace of AI-generated libraries.3

This architecture allows reactions to be miniaturised, analysed in parallel and connected directly to matrix-tolerant binding assays. In one peer-reviewed workflow, linear DNA assembly, bacterial CFPS and a homogeneous proximity assay enabled a single researcher to express and profile hundreds of antibody fragments in less than 24 hours. The study evaluated 135 published human antibodies and 119 antibodies derived from immunised mice.4 It also demonstrated an important limitation: the high-throughput screen used a synthetically dimerised Fab format because full-length IgG assembly was less consistent under the same conditions. Fragment formats can maximise throughput, but candidates must subsequently be reformatted before their behaviour as intact antibodies can be confirmed. Full-length formats preserve heavy–light-chain pairing and native valency, but they are more demanding CFPS products. Two heavy and two light chains must be produced at suitable ratios, folded and paired correctly and connected through multiple intra- and interchain disulphide bonds. Nevertheless, several cell-free systems have successfully produced functional intact antibodies.5–8

The relevant question is not whether CFPS can reproduce every feature of mammalian expression, but whether it can generate sufficiently reliable expression and target-binding evidence to support an initial triage decision. The available evidence is beginning to define where that threshold can be met.

Peer-reviewed work has shown that bacterial CFPS can produce aglycosylated full-length trastuzumab with biophysical and affinity behaviour comparable to mammalian material.5 This establishes the technical feasibility of producing functional intact antibodies in a bacterial cell-free system, although evidence from a single antibody cannot establish performance across diverse sequences and targets.

Although cell-free and mammalian material are not universally interchangeable, CFPS can provide early expression and target-binding evidence before mammalian confirmation. For workflow design, the relevant value is triage: removing clearly binding-negative candidates before they consume downstream capacity.

The value of an upstream gate depends partly on keeping its remit limited. CFPS can support early decisions about expression and target engagement, but it cannot replace mammalian material for questions governed by host-cell processing, cellular function or the properties of the intended drug format.

Bacterial CFPS-derived IgG is commonly aglycosylated.5 It should therefore not be used to infer glycan-dependent Fc-receptor engagement, antibody-dependent cellular cytotoxicity or complement activity. Cell-based potency and mechanism-of-action studies also require assay systems that reproduce the relevant biological context.

Nor does early expression and binding establish that a candidate is developable. Secretion behaviour, long-term stability, aggregation, viscosity, polyspecificity, formulation tolerance, pharmacokinetics, immunogenicity and manufacturing performance remain downstream questions. Clinical-stage antibodies occupy a constrained biophysical space, underscoring the need for multiparametric developability assessment beyond target binding.9 The role of CFPS is therefore routing, not replacement. It helps determine which

candidates warrant mammalian expression and deeper characterisation, while preserving expression failures and indeterminate results as distinct outcomes rather than misclassifying them as non-binders. Candidates that progress should still be rebuilt and tested in the intended mammalian system using acceptance criteria appropriate to the program. Even within those boundaries, the gate produces something valuable beyond a progression decision: a structured record of which designs expressed, bound, failed or remained indeterminate.

The most durable benefit of a triage layer may be the data set it creates. Antibody sequences are increasingly available, but uniformly generated functional labels remain comparatively scarce, expensive and often siloed. The AIntibody benchmark highlighted both the bias in public sequence corpora and the need for larger, more diverse experimental data sets.¹ A well-designed triage campaign can return positive, negative, technical-failure and indeterminate outcomes for an entire candidate panel rather than documenting only the successful molecules. For these data to improve future models, they must be structured at source. Records should capture the antibody format, template and reaction batch, expression and assembly measurements, antigen and assay conditions, raw and normalised signals, replicate behaviour and confidence category. Technical failures must remain distinct from biological negatives; otherwise, ambiguous labels may be recycled as ground truth.

The value lies not only in collecting more data, but in collecting more informative data. Model-guided antibody optimisation already benefits from experimentally labelled, target-specific measurements. In one example, a deep-learning workflow trained using directed-evolution data identified an antibody variant with a 17-fold improvement in affinity.10 The same principle can be applied earlier in discovery: use model uncertainty and sequence diversity to select the next candidate panel; run a rapid physical gate, and retrain using the complete set of outcomes.

This creates a more informative design–build–test–learn cycle. Rather than selecting only candidates that resemble known successes, teams can explore less certain sequence space and generate the physical evidence needed to evaluate it. The objective is not simply to screen more sequences, but to ensure each round improves the next design decision. Faster classification and better labels support a staged workflow in which each platform answers a deliberately limited set of questions.

AI antibody discovery will not be unblocked by replacing one expression host with another. It will be unblocked by matching each experimental system to the question it is best equipped to answer. CFPS can provide a rapid, scalable layer of physical evidence between digital design and mammalian confirmation: does the candidate express, does it bind and how confident is that classification? Mammalian systems can then be reserved for the biological, developability and manufacturing questions that genuinely require the intended cellular context and antibody format.

This sequencing changes the economics of exploration. Larger and more diverse computational panels can be tested, negative results can be preserved rather than inferred and downstream capacity can be committed to candidates supported by evidence.

Cell-free triage is not intended to replace mammalian validation but rather move a defined part of this validation earlier in the discovery process.

References: 1. Visit: doi.org/10.1038/s41587-026-03238-62. Visit: doi.org/10.1038/s41586-025-09721-53. Visit: doi.org/10.1038/s41576-019-0186-34. Visit: doi.org/10.1038/s41467-023-38965-w5. Visit: doi.org/10.4161/mabs.4.2.19202 6. Visit: doi.org/10.1002/btpr.2082 7. Visit: doi.org/10.1038/s41598-018-36691-88. Visit: doi.org/10.1038/s41598-017-12364-w9. Visit: doi.org/10.1073/pnas.1616408114 10. Visit: doi.org/10.1038/s41467-023-36028-8

Dr Juhi Misra is senior product manager for antibody applications at Nuclera in Cambridge, UK. She began her career in academic research and holds a PhD in Preclinical Medicine and Ageing. She now leads product strategy and execution for antibody discovery and cell-free protein expression workflows, drawing on experience in protein biology, antibody validation, high-throughput screening and commercial service delivery. Her current work focuses on translating cell-free expression and characterisation into scalable triage and kinetics workflows that enable biopharmaceutical and machine-learning teams to turn large digital candidate libraries into decision-ready experimental evidence.