Week 3 HW: Cell-Free Protocols

HTBAA Homework #3: Designing a Cell-Free Reaction Composition

Prepared by: Nitin Due: Tuesday, 6 October 2026

0. Scope, and how this relates to my main project

This is a deliberate branch off my HW1/HW2 thread rather than a continuation of it. My main project is a lab-in-the-loop system for discovering aptamer biosensors, and in HW2 I designed a theophylline-riboswitch-gated sfGFP construct to validate the cell-free layer of that pipeline. HW3 fixes the three protein targets (sfGFP, PETase, Reteplase) and fixes the reporter sequence, so the riboswitch design does not transfer directly. What does transfer is the reaction chemistry underneath it: every assay in my pipeline runs in an E. coli lysate CFPS reaction, so a cheaper and more reproducible reaction composition lowers the per-assay cost of my entire HW1 scale estimate. I return to that connection briefly in Section D.

One scoping note on provenance: I could not access coopetition.htgaa.org, which requires an HTGAA Keycloak login. That means I do not have the actual reagent menu, the available stock concentrations, or the current community reagent list. Every volume in Section B is therefore computed against assumed stock concentrations and must be recomputed against the real menu before submission. I have flagged this at the point of use rather than burying it here.

A sequence check worth doing first

I verified the three supplied sequences rather than taking the handout’s stated sizes at face value, because $/g is inversely proportional to molecular weight and a tag changes it:

The cysteine counts confirm the handout’s folding-difficulty argument exactly (0, 1 and 9 disulfide bonds). The length and mass discrepancies for PETase and Reteplase are not errors: the supplied constructs carry N-terminal Strep-tag II (WSHPQFEK) plus a TEV site (ENLYFQ) and linkers, and sfGFP carries a C-terminal Strep-tag II. The handout’s figures describe the untagged mature proteins. This matters for the coopetition objective: if titer is quantified on the tagged construct, the $/g denominator should use the tagged mass (33.2 and 44.1 kDa), which is roughly 13–19% higher than the untagged values and therefore makes $/g look correspondingly better. Worth confirming which convention the scoring uses.

A. Compare cell-free reaction compositions

A1. Why potassium glutamate and magnesium glutamate concentrations differ, and which is more sensitive

The two differ by more than an order of magnitude in every published formulation — potassium glutamate sits at roughly 130–200 mM while magnesium glutamate sits at roughly 4–12 mM — because they do different jobs.

Potassium glutamate sets bulk ionic strength and is there to mimic the E. coli cytoplasm, which is the explicit founding premise of PANOx-SP (Jewett and Swartz, 2004). Glutamate is deliberately chosen as the counter-ion over acetate or chloride because glutamate is the dominant cytoplasmic anion. Its optimum is broad: it is a bulk electrostatic environment, not a stoichiometric participant.

Magnesium glutamate is a catalytic and structural cofactor. Mg²⁺ mediates ribosomal subunit association, stabilises tRNA and mRNA tertiary folds, and is stoichiometrically complexed with nucleotide triphosphates (the true polymerase substrate is Mg-ATP, not free ATP). Too little and ribosomes dissociate and polymerases stall; too much and translational fidelity drops and nucleotides are over-chelated. Its optimum is correspondingly narrow.

Magnesium is clearly the more sensitive of the two, and I want to ground that in experimental behaviour rather than assert it mechanistically. The strongest evidence is what the Olsen et al. (2026) team actually re-optimised: they re-determined magnesium and phosphate concentrations for every new lysate batch, for every change in lysate concentration, and for every new source strain — and did not do the same for potassium glutamate. A parameter that must be retuned per lysate batch is, by definition, the sensitive one. Two further observations support this: E. coli extract optima cluster tightly at 8–12 mM Mg(Glu)₂, whereas a C. autoethanogenum extract optimised at 32 mM with a reported five-orders-of-magnitude swing in luciferase expression across the 8–36 mM range (Krüger et al., 2020) — an enormous response to a modest concentration change; and across the eight formulations Olsen et al. tabulated, magnesium glutamate barely moves (4–12 mM) precisely because it has so little room to move.

Practical consequence for the coopetition: magnesium is the parameter most likely to cause a composition that wins in round 1 to fail reproducibility in round 2, because the 96-well re-test may use a different lysate batch. Spending screening budget on a magnesium titration around the chosen point is better insurance than optimising a dozen cheap salts.

A2. PEP+NTP versus ribose+NMP formulations

These are two different answers to the same question: where does the ATP come from?

PEP + NTP (the PANOx-SP approach). Nucleotides are supplied directly as triphosphates (ATP ~1.2 mM; GTP/UTP/CTP ~0.86 mM each), and phosphoenolpyruvate at ~33 mM acts as a high-energy secondary phosphate donor, regenerating ATP from ADP via pyruvate kinase. It works well and was the highest-yielding formulation in Olsen et al.’s own head-to-head comparison. It has two problems. Cost: PEP alone runs roughly $2,000–3,000 per litre of reaction, pushing total reagent costs above $4,000/L and accounting for over 75% of material costs. Metabolic dynamics: Olsen et al. measured the time course directly and found ATP surges within 30 minutes and then falls to nearly zero, while inorganic phosphate rises sharply and stays high because Pi is liberated stoichiometrically as PEP is consumed. Accumulated phosphate chelates magnesium and the reaction dies early — the system is energetically front-loaded and short-lived.

Ribose + NMP (the RF_opt approach). Nucleotides are supplied as monophosphates, which are far cheaper, and the lysate’s own surviving kinase activity phosphorylates them up to triphosphates in situ. Energy comes from non-phosphorylated carbon sources — ribose, maltodextrin, glucose — feeding glycolysis to regenerate ATP. The measured behaviour is qualitatively different: RF_opt begins with no exogenous ATP at all, yet reaches comparable ATP within 30 minutes and then sustains roughly 250 µM for the entire reaction, well above the ATP affinity threshold for cell-free expression. Phosphate stays at 10–15 mM, about three-fold lower than PANOx-SP. The outcome is a longer-lived reaction: 2.4 ± 0.3 g/L sfGFP, an 87% yield increase over PANOx-SP, at 99% lower reagent cost.

Two constraints and one caveat that matter for design. NMP concentrations must be capped at about 5 mM or insoluble precipitates form, and NMPs must be added to the reagent mix last. Separately, Olsen et al. found GMP can be replaced by guanine plus ribose with no yield penalty — relevant to Section C below. The caveat: RF_opt buys its cost advantage at the price of robustness. It is more sensitive to lysate amount than PANOx-SP, more sensitive to surface-area-to-volume ratio (because it is oxygen-limited), and needs per-lysate magnesium and phosphate retuning. In a 384-well plate at 20 µL — a high surface-area-to-volume, oxygen-rich geometry — that sensitivity may actually work in our favour, but it is the main reason a cheap composition can fail to reproduce on scale-up to 2 mL tubes.

A3. Compare and contrast two published formulations: PANOx-SP versus RF_opt

What the comparison actually teaches. The headline is not “fewer components” but that removing expensive components and switching the energy strategy raised yield. Olsen et al. explored 1,231 reaction mixtures to get there, and a minimal formulation of only potassium glutamate, NMPs and amino acids was already productive at 0.5 g/L. That is the single most useful prior for this coopetition: the design space rewards subtraction, and reagents that feel obligatory (tRNA, CoA, NAD⁺) are dispensable when the lysate is not dialysed. It also reframes the cost problem — in RF_opt, reagents fall from 75% of total material cost to under 4%, meaning that once you adopt a cheap formulation, lysate and DNA dominate and further reagent trimming has sharply diminishing returns.

Provenance warning on the cost figures

These numbers are widely quoted in mutually incompatible ways, and I think this is worth stating explicitly before anyone builds a cost model on them.

• Olsen et al. is peer-reviewed (Nature Communications, 5 March 2026, 17:3478). The published version reports ~$60/g_protein at the 15 µL scale and ~$39/g at 4 mL with oxygen supplementation. The August 2025 bioRxiv preprint of the same work reports $55/g and $36/g. The figures shifted between preprint and publication; the published values should be cited.
• Smith et al. (2026) is a preprint (bioRxiv, 5 February 2026) and has not been peer reviewed. It reports sfGFP at $422/g versus what it calls Olsen et al.’s $698/g — a 40% improvement, with a 27% titer increase, from six rounds and roughly 29,000–36,000 conditions on Ginkgo’s RAC cloud lab.
• These two sets of numbers are not comparable. Olsen’s ~$60/g counts reagents only; Smith’s $422/g counts total reaction components including lysate and DNA. Quoting “$60/g” and “$422/g” in the same table as though they measure the same thing would be wrong by roughly an order of magnitude. Note also that the 36,000-experiment figure comes from Ginkgo’s own press release, not the preprint’s abstract, and press-release figures should be labelled as such.

My own sanity check, flagged as inference rather than a sourced result. Lysate is fixed at 5,000 nL per 20 µL reaction, so at $4.00/mL it contributes exactly $0.02 per reaction regardless of composition. At a 2.4 g/L titer, a 20 µL reaction yields 48 µg, so lysate alone sets a floor of about $417/g. That this lands within 1% of Smith et al.’s reported $422/g total is striking and suggests their number is almost entirely lysate-dominated at a titer near 2.4 g/L. I have not verified their titer or cost basis, so treat this as a hypothesis to check against their Table 1 — but if it holds, it means the only lever that materially moves $/g in this coopetition is titer, not reagent cost. Reagent trimming below a certain point is optimising a rounding error.

For orientation, the lysate-only floor at other titers:

B. Three reaction compositions

Read this before the tables. Volumes are computed as V = 20,000 nL × C_final / C_stock, rounded to the nearest 25 nL as the specification requires. The stock concentrations are my assumptions, since I could not reach the coopetition reagent menu. The target final concentrations are the real design content and are grounded in the citations given; the nanolitre volumes are arithmetic that must be recomputed against the actual menu stocks. Rounding to 25 nL introduces up to roughly 6% error on the smallest additions (spermidine, putrescine), which is acceptable for those but would not be for magnesium — if the real stock makes magnesium land awkwardly, dilute the stock rather than accept the rounding error, for the sensitivity reasons in A1.

All three designs follow the RF_opt direction (NMP + non-phosphorylated energy, no PEP, no tRNA/CoA/NAD⁺), because A3 establishes that this is both cheaper and higher-yielding, and because A’s cost analysis says titer is what matters.

Design 1 — sfGFP (reducing)

Rationale: an RF_opt-direction composition using the 5 mM NMP cap, maltodextrin and ribose for ATP regeneration, spermidine retained (both Olsen et al. and Smith et al. identify spermidine as yield-positive), and no PEP, tRNA, CoA or NAD⁺. The 1,400 nL of headroom is deliberate: it is the room to run a magnesium titration (A1) without rebuilding the whole composition.

Design 2 — PETase (reducing)

I would submit the Design 1 composition unchanged for PETase in round 1. This is an experimental-design choice, not laziness, and I want to defend it explicitly. TfCut2 has one disulfide bond and the handout states it remains compatible with reducing conditions, so there is no mechanistic reason to switch regimes. More importantly, I have no PETase-specific data, and the handout’s own framing — that a composition performing well for one protein may perform poorly for another — is a hypothesis that only gets tested if someone submits the same composition across two proteins. Submitting an sfGFP-optimal composition for PETase measures the transfer gap directly and tells the community something the all-variants-differ approach cannot. If the transfer gap turns out to be large, round 2 has a clear direction: a lower magnesium variant, since the longer transcript and the single disulfide may shift the optimum.

Design 3 — Reteplase (oxidising / DSB)

Also specified outside the variable budget: BL21 Star (DE3) Δgor lysate, IAM-treated at the low dose (25 µM, not 500 µM), with the IAM quenched by cysteine.

The design decision that matters here, and why. The handout tells you Reteplase needs GSSG/GSH and DsbC. What it does not tell you is that bolting that oxidising module onto a low-cost non-phosphorylated formulation is actively dangerous. Olsen et al. found that adding IAM, GSSG and GSH to RF_opt reduced sfGFP production by 75–97%. They traced the mechanism: glutathione addition drops the reaction pH, and because RF_opt is already more acidic than PANOx-SP, the pH falls below 7.0 — outside the 7.4–7.8 cytoplasmic range where PANOx-SP comfortably sits. Two fixes recovered the yield: raising the HEPES buffer to pH 7.5, and knocking out glutathione reductase (Δgor) in the lysate source strain, which let them use 20-fold less IAM (25 µM rather than 500 µM) while stabilising the oxidising environment. Both are built into the design above. Without the pH 7.5 specification, I would expect this composition to fail badly, and I think that is the most likely silent failure mode for anyone submitting a cheap oxidising Reteplase formulation.

The specific values — 4 mM GSSG, 1 mM GSH (a 4:1 oxidised:reduced ratio), 5 µM purified DsbC, 25 µM IAM, Δgor — are Olsen et al.’s, not invented. There is also a direct functional precedent: they expressed vtPA, a truncated tissue plasminogen activator closely related to Reteplase, in oxidised RF_opt and showed it was enzymatically active using the IPR-AMC fluorogenic assay — the same substrate and the same assay the handout specifies for Reteplase here. That is about as close to a validated starting point as this target has. Note, though, that even with the pH fix, glutathione addition still reduced yields by up to 62% in their hands, so Reteplase should be expected to be the lowest-titer and highest-$/g of the three targets regardless of composition.

C. Community reagent proposal: guanine

Proposal: add guanine (the free nucleobase) to the reagent inventory.

How I decided on it. I worked backwards from where the remaining cost actually sits. Section A establishes that once PEP is removed, NMPs become the most expensive reagent group in the minimal formulation — Olsen et al. put them at $62/g_protein — and that lysate then dominates everything else. So the only reagent-side lever left with real headroom is making nucleotides cheaper. Olsen et al. tested exactly this and report a specific, actionable result: most individual replacements of nucleoside monophosphates with their corresponding nucleosides or nucleobases decreased yield, but GMP could be replaced by a combination of guanine and ribose with no loss of protein yield. They name this in their discussion as “one potential path” to lower cost. Ribose is already on the menu under energy regeneration; guanine, as far as I can tell from the handout’s category table, is not. So the single missing piece needed to test a published, yield-neutral cost reduction is one cheap nucleobase.

Why it is a good coopetition reagent specifically: it is a targeted substitution with a published positive result behind it rather than a speculative addition, it attacks the largest remaining reagent cost, nucleobases are substantially cheaper than nucleotides, and it is a clean single-variable experiment (swap GMP for guanine, hold everything else) that fits the staged screening design well. If it works, adenine, cytosine and uracil are the obvious follow-ups — though Olsen et al.’s data suggests those will not transfer, since guanine was the surprising exception rather than the rule. That asymmetry is itself worth knowing.

Caveat I would post with it: Olsen et al. observed insoluble precipitates with many nucleotide alternatives and explicitly flagged precipitate formation as needing further investigation. Guanine has poor aqueous solubility at neutral pH, so this needs a solubility check at the intended stock concentration before it goes into an Echo acoustic dispenser — nanolitre acoustic transfer is unforgiving of precipitates, and a reagent that crashes out in the source plate will silently wreck a whole 384-well run.

A second idea for the forum (more speculative, and I would label it as such). Olsen et al. found dissolved oxygen was the largest single lever they identified — switching from a 20% to a 100% O₂ feed raised sfGFP from under 2 g/L to 3.7 ± 0.2 g/L, a 45% improvement, and they note RF_opt-style formulations are especially sensitive to surface-area-to-volume ratio. The menu already carries catalase and pyruvate oxidase under oxygen-related components. Whether an oxygen-carrier additive would help in a sealed 384-well plate is a reasonable question, but I have not found a CFPS paper testing one, so I would raise it as a question rather than a recommendation.

D. Connection back to the main project

Three things from this homework feed my HW1 pipeline.

The cost model changes. My HW1 estimate costed the design phase on oligonucleotide synthesis and treated the assay as negligible. The arithmetic in Section A says the assay floor is set by lysate and titer, not reagents — roughly $417/g at a 2.4 g/L titer. For a pipeline running thousands of 20 µL reporter assays, that is the number to plan around, and it is a correction to HW1.

The coopetition is a working instance of my proposed loop. Smith et al. is GPT-5 proposing compositions, a cloud lab executing them, and results returning to the model over six rounds — structurally the same closed loop I proposed in HW1, with a Pydantic schema doing the design-to-protocol validation step I described abstractly. Their reported sub-1% error rate for AI-designed plates reaching the lab is a useful benchmark for the protocol-generation step I hand-waved.

One hypothesis I now want to test. A composition optimised for maximum sfGFP titer is not necessarily optimal for a riboswitch biosensor, which needs low background and high dynamic range rather than maximum absolute signal. Magnesium is the obvious suspect, since Section A1 establishes it as the sensitive parameter and Mg²⁺ also stabilises RNA tertiary structure — the very folding the riboswitch depends on. I flagged this intuition in discussion before doing the reading and could not source it then; I still cannot source it for riboswitches specifically, so it remains my hypothesis. But it is now a sharper and more testable one: titrate magnesium against ON/OFF ratio rather than against titer, and expect the two optima to diverge.

E. What I do not know

1. The actual reagent menu and stock concentrations. Every nanolitre figure in Section B is provisional until recomputed against coopetition.htgaa.org. The target concentrations are the defensible part.
2. RF_opt’s complete component list. I have its character — 12 components, NMP-based, maltodextrin and ribose, HEPES, spermidine, no tRNA/CoA/NAD⁺ — and several exact values, but not the full Table 1 with every concentration. My Design 1 is RF_opt-directional, built on PANOx-SP’s well-documented salt concentrations with RF_opt’s energy strategy substituted in. Pulling the real Table 1 would improve it.
3. Whether my lysate-floor inference is right. The $417/g versus $422/g coincidence is suggestive, not established. It needs checking against Smith et al.’s Table 1 before anyone relies on the conclusion that titer is the only lever that matters.

AI use disclosure

I used Claude (Anthropic, Claude Opus 5 as a research and drafting assistant for this assignment, continuing directly from the same conversation and AI-assisted workflow described in earlier HW.

References

Jewett, M.C. and Swartz, J.R. (2004) ‘Mimicking the Escherichia coli cytoplasmic environment activates long-lived and efficient cell-free protein synthesis’, Biotechnology and Bioengineering, 86(1), pp. 19–26. https://doi.org/10.1002/bit.20026

Krüger, A., Mueller, A.P., Rybnicky, G.A., Engle, N.L., Yang, Z.K., Tschaplinski, T.J., Simpson, S.D., Köpke, M. and Jewett, M.C. (2020) ‘Development of a clostridia-based cell-free system for prototyping genetic parts and metabolic pathways’, Metabolic Engineering, 62, pp. 95–105. https://doi.org/10.1016/j.ymben.2020.06.004

Olsen, M.L., Copeland, C.E., Sundberg, C.A., Aw, R., Shaver, Z.M., Rao, G., Swartz, J.R., Karim, A.S. and Jewett, M.C. (2026) ‘Design-driven optimization of low-cost reagent formulations for reproducible and high-yielding cell-free gene expression’, Nature Communications, 17(1), 3478. https://doi.org/10.1038/s41467-026-69605-8 (Peer-reviewed. Preprint version: bioRxiv, https://doi.org/10.1101/2025.08.01.668204, August 2025 — cost figures differ slightly between the two; published values cited here.)

Rasor, B.J., Vögeli, B., Landwehr, G.M., Bogart, J.W., Karim, A.S. and Jewett, M.C. (2021) ‘Toward sustainable, cell-free biomanufacturing’, Current Opinion in Biotechnology, 69, pp. 136–144. https://doi.org/10.1016/j.copbio.2020.12.012

Silverman, A.D., Karim, A.S. and Jewett, M.C. (2020) ‘Cell-free gene expression: an expanded repertoire of applications’, Nature Reviews Genetics, 21, pp. 151–170. https://doi.org/10.1038/s41576-019-0186-3

Smith, A.A., Wong, E.L., Donovan, R.C., Chapman, B.A., Harry, R., Tirandazi, P., Kanigowska, P., Gendreau, E.A., Dahl, R.H., Jastrzebski, M., Cortez, J.E., Bremner, C.J., Che, J.C., Thavarajah, W., Smith, D. and Thomas, T.F. (2026) ‘Using a GPT-5-driven autonomous lab to optimize the cost and titer of cell-free protein synthesis’, bioRxiv [preprint], 5 February. https://doi.org/10.64898/2026.02.05.703998 (Not peer-reviewed. The 36,000-condition figure comes from Ginkgo Bioworks’ accompanying press release, not the preprint abstract.)