Variant Library Screening Strategies in Cell-Free Expression Systems

Cell-free protein synthesis (CFPS) makes high-throughput variant screening practical because it uses open reactions that work with plate-based automation in place of cell culture for each construct. For ML-guided campaigns testing hundreds of variants, its strongest advantages are expression from amplified templates, direct activity assays, and rapid experimental feedback. The decisive system-selection criteria are assay fidelity, support for difficult targets, reagent reproducibility, and cost per variant screened.

Why ML-guided protein design needs faster experimental feedback than cell-based systems can provide

ML-guided protein design depends on functional measurements that distinguish promising sequences from unsuccessful ones. Structure prediction and sequence generation do not remove that requirement. A model needs enough experimental data to learn which changes improve the property under selection, and subsequent rounds need new measurements to refine those predictions. Thornton et al., in ACS Synthetic Biology in 2025, identify protein purification at the scale required for ML training as a binding constraint. Generating candidates is not the same as testing them.

Cell-based expression makes that distinction expensive because each variant enters a workflow of heterologous expression and, conventionally, purification. Each synthesis cycle consumes many days. Cloning, transformation, culture, and downstream processing impose per-variant demands on time, labor, and consumables that conflict with rapid design cycles. Cytotoxicity and inclusion body formation can also cause some variants to fail, so failed expression may tell more about the host than about the intended protein property.

How can CFPS screen variants in plates without cloning and transformation?

CFPS removes the requirement to maintain a living host during protein synthesis. The reaction has no cellular membrane enclosing it, no functional genome to sustain, and no cell whose survival competes with target production. This removes the host cytotoxicity burden and permits direct control over the synthesis environment. Additives go straight into the reaction without having to cross a cellular boundary. Chaperones, disulfide isomerases, detergent micelles, nanodiscs, and liposomes can therefore support targets that are difficult to express inside cells.

That architecture also changes how a library reaches expression. PCR products can substitute for plasmids, removing cloning, transformation, and colony picking from each variant’s path to testing. Extract-based systems can express dozens of open reading frames in a single day, while synthesis workflows can contract from 1–2 weeks to as little as 1–2 days. Because reactions can be assembled in plates, liquid-handling automation acts on a shared reaction format rather than a series of separate cell-culture workflows.

What plate-based CFPS screening looks like in practice — two peer-reviewed cases

Tomková et al., in FEBS Open Bio in 2026, demonstrated an unpurified CFPS platform for screening staphylokinase variants at P. J. Šafárik University in Košice. Their microplate workflow used a chromogenic plasminogen-activation assay to obtain real-time activity readouts. The rankings from crude mixtures were close to those from purified proteins. The authors also described the workflow as modular and adjustable to other enzymes, making the demonstration relevant beyond this particular thrombolytic target.

Thornton and colleagues at the University of Edinburgh supplied the complementary ML-guided demonstration. Their protease variants showed similar activity trends when screened directly in CFPS and when tested as purified proteins at equal concentrations. A first random screen supplied training data, a model proposed targeted variants, and CFPS reduced the time to the next experimental scores from weeks to hours per round. Together, these cases support crude-readout fidelity for their respective assays, rather than a blanket assumption that every protein can be ranked without assay validation.

How the CFPS design-test-learn cycle actually closes — connecting raw activity data to the next round of ML proposals

An ML-guided campaign starts with a sample of the fitness landscape, which measures how sequence changes affect the selected property. The protease study used a 48-variant random first round to obtain that starting information. Model-proposed sequences then returned to the same experimental workflow for expression and scoring, producing data for further learning. CFPS thus serves both the initial exploration and the targeted rounds, without a separate purification-intensive phase to establish the first dataset.

The operational connection is straightforward. Proposed sequences can be amplified, expressed in plates, and assayed without purification where crude readouts preserve the relevant activity rankings. For a model learning those rankings, absolute kinetic constants from purified protein are not required at every screening step. The leading variants can then be selected and purified for characterization. Weekly cycles give more iteration density than monthly cycles because they fit more rounds of experimental learning into the same project timeline. Faster expression becomes useful because it shortens the return path from a prediction to evidence.

Where CFPS screening reaches its limits and how to extend throughput beyond the plate

Plate-based CFPS is well matched to libraries in the dozens to low hundreds per run, especially when an ML model has already narrowed the search. It is not the instrument for every library. When early exploration requires tens of thousands to millions of variants, fluorescence-activated droplet sorting (FADS) provides a complementary throughput tier. FADS encapsulates individual CFPS reactions in droplets and sorts them using fluorescent activity signals. The distinction is the screening architecture, not a choice between cell-free synthesis and an unrelated expression method.

The format should match uncertainty in the fitness landscape. When that landscape is unknown, FADS or random mutagenesis libraries can support broad exploration. Initial hits then define a region for further work. ML-guided plate CFPS can optimize within that region, where each proposed sequence has a stronger basis for selection. This tiered approach keeps plate-based measurements for focused campaigns and avoids putting very large, unbiased libraries through one well at a time.

Expressing difficult targets — how CFPS handles toxic, membrane, and disulfide-bonded proteins that defeat cell-based workflows

High expression can stress or kill cells and can accompany inclusion body formation, complicating comparisons across a variant library. CFPS removes host survival as a competing requirement. For a screening campaign, that distinction allows expression conditions to be selected for the target.

Membrane proteins require a different intervention within the same open architecture. Detergent micelles, liposomes, or nanodiscs can be added during synthesis to supply a membrane-mimetic environment without engineering a heterologous strain such as E. coli. The environment is supplied to the reaction. Across a library, the same additive conditions can be applied to every well, making plate-based comparisons tractable without individual strain optimization. OpenCFPS™ is positioned for these difficult-target workflows, where direct access to the reaction is central to the expression strategy.

Disulfide-rich and multi-domain proteins need help with folding even after the host is removed. DsbC, a disulfide isomerase, and the DnaK chaperone system can increase soluble yield when supplied to CFPS reactions. Documented examples include proteins with up to 24 disulfide bonds and sizes from 14.3 to 53.2 kDa. The library-screening advantage is that isomerases and chaperones can be added uniformly across wells rather than co-expressed or engineered into separate hosts for each variant.

Scaling confirmed hits from screening to milligram production without switching platforms

A screen must lead to enough material to characterize the proteins it selects. CFPS can preserve continuity between those stages rather than requiring a move into cellular expression after hit selection. Dialysis-mode production of anti-EGFR antibody fragments demonstrates this transfer. The conditions optimized at 30-µL scale, including the critical DsbC concentration, also worked in a 9-mL reaction, where yield scaled linearly and exceeded 1 mg/mL. In that system, the additive conditions established at small scale remained useful for production.

This provides a concrete workflow for confirmed hits. Variants are screened at 30 µL per well, leading candidates are selected, and the established conditions are scaled to milliliter volumes for characterization. Batch yields span 1 µg/mL to 2.3 mg/mL depending on protein complexity, so target-specific performance still governs material output. The payoff is avoiding a platform switch where the screening conditions have already proved transferable.

Milligram quantities of ¹⁵N-labeled bovine pancreatic trypsin inhibitor and human lysozyme C have been produced for characterization. NMR spectra and functional analyses supported proper folding and function. Screening and structural characterization can share an expression system even though they demand different amounts of material and different measurements.

Why reagent reproducibility and lot documentation are not optional in a multi-round screening campaign

A Bristol-Myers Squibb study comparing three batches of recombinant sLAG3 showed how total protein concentration can conceal inactive protein species and produce apparent differences in kinetic binding parameters between lots. Defining reagents by assay-specific concentration rather than total concentration dramatically decreased immunoassay lot-to-lot coefficients of variation. Total protein was not a sufficient quality measure. Although this was a recombinant-protein reagent study rather than a CFPS library screen, it exposes a relevant distinction between the amount of protein present and the amount that functions in an assay.

For crude lysate CFPS, batch-to-batch extract variability creates a related risk. The i-POPFLEX preprint identifies reproducibility across individual laboratories as a challenge and describes cost and labor intensity as drivers toward modular, reconstituted systems with automated assembly. Reagent drift can confound comparisons between screening rounds. If one lot changes expression efficiency, a model may learn a reagent-associated difference as though it were a sequence-associated improvement. More measurements do not correct a systematically shifting experimental baseline.

Lot documentation belongs in the screening data. Commercial reagent names and lot numbers provide traceability, while crude lysate workflows need documented per-lot quality control to support meaningful comparisons. OpenCFPS™ publishes lot-level QC data for this purpose. Its formulation transparency also gives laboratories a basis for auditing changes when reproducibility drifts. Documentation makes reagent changes available for investigation.

How reagent economics determine whether high-throughput CFPS screening is actually accessible at the scale ML campaigns require

Reconstituted CFPS has historically faced a substantial per-reaction cost barrier. The i-POPFLEX preprint reports a modular reconstituted system cost of USD 0.054/µL, described as 40% lower than the previously reported minimum for reconstituted CFPS. That result narrows the economic gap with crude lysate systems. The preprint also says that in-house PURE preparation takes a lot of work, so access depends on more than reagent price. Automated modular assembly addresses a preparation burden that otherwise competes with the screening work itself.

A separate 2025 reagent-optimization preprint reports a minimal 12-component formulation that reduces reagent cost from the dominant share of material costs to a small fraction. It also reports formulations that produce more than 2 g/L protein product. These remain preprint results, but they illustrate a broader shift toward more accessible reaction economics. OpenCFPS™ pricing is positioned within that shift, with high-throughput access as an explicit aim. Without a stated price, however, that positioning does not establish a numerical cost advantage over other systems.

For an engineering team, the useful comparison is cost per variant screened. That comparison includes the cloning, culture, purification, and labor that a suitable CFPS workflow removes. Formulation transparency also has an economic role. An open, documented formulation lets a laboratory look into a yield drop instead of having only a black-box reagent replacement to turn to. With faithful assays, documented reagents, and costs assessed across the complete workflow, CFPS is not a niche workaround but a practical experimental foundation for ML-guided protein engineering.