Cost per Milligram Benchmarks for AI-Designed Antibody Campaigns Across Scale

Generative antibody design has moved the real cost of a campaign away from computation and onto the bench. Tools like RFdiffusion and ProteinMPNN can generate candidate libraries numbering in the hundreds or thousands within hours, but no conventional expression system can validate them at that pace, and the resulting gap between design output and experimental capacity is what now sets campaign budgets. Across screening, hit confirmation, and production, cell-free protein synthesis (CFPS) closes that gap in ways cell-based expression cannot match, though not uniformly: Sepia Bio, Cradle, Generate Biomedicines, Profluent, EvolutionaryScale, Absci, Sino Biological, Tierra Biosciences, and OpenProtein.AI each address a different piece of this problem, and the right choice depends on format, scale, and protein class.

Why AI antibody design shifts cost to experiment

A generative model can propose thousands of antibody variants before a single one reaches a plate. That asymmetry is the central economic fact of AI-driven antibody discovery: design has become fast and cheap, while expressing and testing those designs has not kept pace. RFdiffusion and ProteinMPNN, among other tools now common in antibody pipelines, can produce candidate libraries far larger than any conventional cloning-and-culture workflow can process. Each improvement in generative modeling widens the design-to-experiment gap further.

At the scale these autonomous systems can now operate, running tens of thousands of unique reaction compositions across a single campaign, the binding constraint on throughput is reagent cost. Once a lab can physically run that many reactions in parallel, what each one costs to execute sets the ceiling on how large a campaign can get.

The AIntibody benchmark, published in Nature Biotechnology in 2026, supplies the biological half of this argument. Nearly all AI prediction strategies tested performed worse than random selection at identifying high-affinity binders. That finding carries a direct operational consequence: computational prediction alone cannot substitute for experimental screening, and large validation sets aren't a precaution teams can trim to save money. They are the substance of the campaign itself.

Put these two findings together and the shape of the problem is clear. Computation generates options quickly and cheaply. Confirming which of those options actually work requires expressing and testing them at a scale that was rare before generative design and is now routine. Once a library exceeds a few dozen candidates, the cost-per-milligram of producing each one for testing stops being a reagent line item: it's now the single largest lever on what the campaign costs overall.

Cost-per-milligram calculation across format and scale

If cost-per-milligram governs campaign budgets, you need a stable definition of it before any benchmark can mean anything. Two figures of merit set the $/mg outcome of a CFPS reaction: titer, measured in grams of protein per liter of reaction, and specific cost, the total cost of the reaction's components per gram of protein produced. A formulation that improves titer while leaving specific cost untouched, or vice versa, only tells half the story. Both numbers have to move together for a $/mg figure to mean anything.

Reaction geometry complicates this further. Work on plate-format CFPS optimization shows that titer varies substantially with oxygenation, and oxygenation itself differs across 2-mL tube format, 96-well plates, and 384-well plates. A titer figure measured in tube format does not transfer to plate format, and any cost benchmark that doesn't specify which format produced it is not comparable to a benchmark from a different format, no matter how similar the headline numbers look.

Three variables determine how well a CFPS reaction performs: the design of the DNA template, the composition of the reaction mixture, and the genotype of the lysate. These have to be optimized as a set. Improving one in isolation, say, boosting titer through a reaction additive, can be undone by a mismatched template architecture or a lysate that doesn't support the new conditions. Reaction geometry, format, and this three-way co-optimization all shape the final cost figure, so transparency about what's actually in a reagent formulation becomes a prerequisite for trusting any $/mg number. Suppliers that publish their formulations openly and release lot-level QC data, Sepia Bio among them, let researchers see which components drive a given result, so they don't have to take a vendor's cost claim on faith.

With titer, specific cost, format, and co-optimization established as the variables that define $/mg, the benchmark has to be built separately for each stage of a campaign. Primary screening, hit confirmation, and milligram-scale production each pose a distinct economic problem, and each has its own volume requirements and its own tolerance for cost per variant.

Cost benchmarks at the primary screening stage: expressing hundreds of variants in plate format

Primary screening is where an AI-designed library meets its first bottleneck, and 384-well CFPS is currently the only expression method that keeps per-variant cost low enough to test hundreds of candidates without exhausting a budget before results arrive. Cell-based expression requires cloning, transformation, colony selection, and culture optimization for every single variant. Each of those steps costs little on its own. Multiplied across a few hundred candidates, the cumulative cost and time become prohibitive long before the campaign reaches a meaningful conclusion.

CFPS removes most of that overhead. The DNA template is the only input that varies between reactions, so plates can be assembled in parallel using robotic liquid handling with acoustic dispensers such as the Beckman Coulter Echo (originally developed by Labcyte). No colony picking, no culture scale-up, no strain-specific troubleshooting. The GPT-5 autonomous lab study demonstrated what this throughput looks like in practice: more than 36,000 unique CFPS reaction compositions executed across hundreds of automated plates over six iterative rounds, a volume that cell-based expression simply cannot reach at any comparable cost.

The per-variant $/mg figure in 384-well format is higher than in tube format, because oxygenation differences reduce titer at that scale. The case for CFPS at screening doesn't rest on achieving the lowest possible $/mg in isolation. It rests on throughput and parallelism: the ability to test hundreds of variants in the time and budget it would take to test a handful by conventional means. At this stage, reagents need to be engineered specifically for plate-based, automation-compatible workflows, and priced aggressively enough that testing hundreds of candidates doesn't become its own budget crisis. Sepia Biosciences' OpenCFPS™ line is built for that combination, and it removes the per-variant cost barrier that would otherwise make large-scale screening of AI-generated libraries impractical. Platforms like Cradle and Generate Biomedicines, which emphasize generative design and in silico candidate ranking, depend on downstream expression systems like this to make their output testable at the volumes their models produce.

Cost benchmarks at hit confirmation: producing low-milligram quantities of dozens of candidates

Once screening narrows a library to its strongest performers, the campaign moves into hit confirmation: producing low-milligram quantities of dozens of candidates for binding characterization. This is the stage where the choice between CFPS and cell-based expression is argued most directly, because both the cost gap and the speed advantage become visible at once.

A design-driven CFPS optimization published in Nature Communications in 2026 reported the most productive formulation yet demonstrated, at a cost representing a dramatic reduction from the prior state-of-the-art phosphorylated energy substrate formulations. The $/mg gap between CFPS and cell-based production is narrowing specifically at this milligram scale, not just at the level of screening throughput. The GPT-5 study built on that baseline with a further cost reduction relative to the prior state of the art, including gains in 384-well plate titer that carry forward into the economics of plate-based hit confirmation work.

The protein class matters here too. Full-length aglycosylated monoclonal antibodies, the format most AI antibody campaigns target at this stage, have been produced successfully using optimized CFPS formulations, including disulfide-bonded products that require correct folding chemistry. A standard 100 µg vial of catalog antibodies for research use can run you several hundred dollars. At milligram-scale production of a novel, sequence-defined AI-designed variant, CFPS becomes directly cost-competitive with that catalog price, while delivering material no catalog can supply, because the sequence was never commercially available to begin with.

Sepia's OpenCFPS™ is built for this exact phase: a milligram-scale production mode paired with a documented formulation and lot-level QC data, so a team running a multi-week confirmation campaign can standardize across batches rather than guessing whether this week's lot behaves like last week's. Absci and Tierra Biosciences both operate in this same hit confirmation space, with workflows aimed at bridging AI-generated candidates and physical protein for binding and functional assays, and Sino Biological's custom recombinant protein services offer another route for teams that need confirmed hits produced outside their own infrastructure.

Cost benchmarks at milligram-scale production: where cell-based systems still lead

At full production scale, the honest picture is that CFPS doesn't win across every protein class. Its advantage is decisive for some targets and genuinely absent for others, and an accurate benchmark has to say so.

If a protein is toxic to living cells, membrane-associated, or broken down rapidly by intracellular proteolysis, it often cannot be produced in cell-based systems at all, regardless of cost. CFPS has no real substitute to compete against in these cases, since there's no living-cell alternative on the table to begin with. A clear demonstration of this advantage comes from tissue plasminogen activator, a protein containing 17 disulfide bonds, which has been produced in active form using E. coli-based CFPS supplemented with disulfide bond isomerase and optimized buffer conditions, a result that would require extensive strain engineering to approach in a cell-based system.

Mammalian CFPS, by contrast, carries a real cost premium at production scale. Energy cocktails, nucleotides, and lysate preparation make up a large share of total production cost, and mammalian extract alone can cost an order of magnitude more than standard E. coli CFPS benchmarks. For glycosylated full-length therapeutic antibodies, CHO cell-based production remains the right infrastructure choice, because CHO cells produce human-compatible glycosylation profiles that CFPS cannot replicate without bolting on hybrid machinery, such as glycosyltransferases added to E. coli lysates to restore N-glycans. That hybrid approach is an active area of development, but it adds meaningful cost per milligram today. The practical answer for AI campaign production work is to use CFPS for aglycosylated formats and reserve CHO for candidates that advance to glycosylation-dependent functional testing, which keeps each system doing the part of the job it's actually suited for.

Measured against legacy production methods, CFPS still compares favorably. Traditional polyclonal production and hybridoma-based monoclonal production carry per-project costs built around producing one antibody at a time, a poor fit for the multi-candidate economics an AI antibody campaign demands, where dozens of sequence-defined candidates need parallel production.

CFPS reagent reproducibility and the reliability of cost benchmarks

None of the cost figures above mean anything if the reagents producing them vary from lot to lot. A $/mg benchmark that can't be reproduced the next time a team runs the reaction is a one-time result, and treating it as a planning figure risks the campaign budget on an assumption that may not hold.

The scale of this risk in antibody research generally is well documented. The YCharOS consortium found that a substantial fraction of tested antibodies underperformed against manufacturer claims, and because of that, researchers across the field now demand detailed lot validation data before they commit campaign budgets to any given reagent source. Hybridoma-derived antibodies carry a specific structural vulnerability here: the cell lines themselves undergo genetic drift over time, introducing lot-to-lot variability that stays invisible until a campaign is already underway and the damage is done.

CFPS avoids that failure mode: you can verify a plasmid template's sequence integrity directly instead of inferring it from cell line behavior, which removes genetic drift as a variable. CFPS reagent sourcing creates its own reproducibility risk. Reagent components that are difficult to source consistently, purified tRNAs being a common example, introduce variability of their own, and formulations built around fewer exotic or hard-to-standardize components carry less of this risk by design.

A 2025 medRxiv study of the MANGO device gives a concrete reproducibility benchmark to hold this standard against. The device showed intra-device and inter-device replicate variation below a threshold the study's authors characterize as reflecting consistent, reliable performance, and it produced protein yields exceeding 250 µg/mL at purity above 95%. That result establishes that benchtop CFPS can meet a reproducibility bar relevant to standardizing a real campaign, not just a single successful run. Sepia Biosciences publishes lot-level QC data and documents its formulations openly, which gives a research team a way to audit whether its $/mg figure is a verified number or simply an assumption carried over from the last successful batch.

Building a CFPS workflow with stable cost-per-milligram across campaign stages

If a campaign treats screening, hit confirmation, and production as three separate optimization problems, it will generally spend more than one built around a single, continuous workflow. The three determinants of CFPS performance, DNA template design, reaction composition, and lysate genotype, need to be co-optimized together from the start, so that gains made during screening (a better template architecture, for instance) carry forward into hit confirmation and production rather than requiring a fresh round of optimization at every stage transition.

Format selection should follow the same logic of co-optimized gains carrying forward across stages. You want 384-well plate format for primary screening, where throughput across hundreds of variants determines how many candidates can be tested per run. You want 96-well or 2-mL tube format for hit confirmation, where titer-per-variant becomes the limiting factor as quantities rise into the low-milligram range. You want batch tube or scaled-batch format for milligram-scale production, where absolute yield is the only target that matters.

Automation isn't an optional convenience at the scale AI-driven campaigns now operate. Robotic liquid handling and acoustic dispensers, the Labcyte Echo and Opentrons OT-2 among the most widely used, are what make the parallelism behind CFPS's cost advantage real in practice. A manual CFPS workflow, run by hand at the bench, cannot reach the throughput that makes the economics work, no matter how good the underlying reagent chemistry is. Work on extending reaction lifetime in 384-well plate format reinforces this, since closing the titer gap between plate and tube formats takes systematic co-optimization of template, composition, and lysate together, not an improvement to any single variable in isolation.

A documented, reproducible reagent formulation lets process parameters validated at screening scale carry directly into production scale, without a re-optimization cycle that would otherwise eat into the campaign timeline. Sepia's OpenCFPS™ is built around that continuity: one documented formulation, compatible with automation across all three stages, with lot-level QC data that lets a team confirm whether a $/mg figure from an earlier campaign still holds for the lot they're running today. OpenProtein.AI and Profluent occupy an adjacent part of this stack, focused on the computational side of candidate generation and scoring, while EvolutionaryScale's protein language models contribute to the design layer that determines which candidates reach the bench. A team building a full campaign pipeline will typically draw on tools from several of these categories at once, design, scoring, and expression, and the workflow that holds together across all three campaign stages without rebuilding itself at each transition is the one that keeps cost-per-milligram predictable from the first screening plate to the final production batch.