Week 1 HW: Final Project Ideas

Project Idea: Cell-Free Manufacturing of chemotherapeutic Etoposide Precursor

(1) What you would like to biomanufacture and why?

The goal is to biomanufacture podophyllotoxin, a natural precursor used in the production of the chemotherapeutic drug etoposide, which is used primarily to treat testicular cancer and small-cell lung cancer. Podophyllotoxin is traditionally obtained from mayapple plants, particularly Podophyllum hexandrum (Himalayan mayapple), which contains higher concentrations of podophyllotoxin than the more common American mayapple, Podophyllum peltatum. The reliance on these plants for extraction creates both a supply challenge and a conservation concern, particularly for P. hexandrum, which has experienced significant pressure from harvesting. There has already been work toward producing podophyllotoxin in alternative biological systems, including engineered tobacco plants. However, I am interested in exploring whether a more controllable biomanufacturing platform, such as a microbial chassis or cell-free system, could eventually provide a more flexible and scalable route to production. My aim is to both: contribute to a more reliable supply of an important pharmaceutical precursor, potentially making its downstream production more accessible and less dependent on plant harvesting, while also reducing pressure on the native plant populations from which podophyllotoxin is traditionally obtained.

(2) What living system would you use?

E. coli would be used as the production host for the enzymes needed to build the cell-free system. The genes encoding the mayapple pathway enzymes, including pinoresinol-lariciresinol reductase (PLR), secoisolariciresinol dehydrogenase (SDH), two O-methyltransferases (OMT1 and OMT3), the ring-closing 2-oxoglutarate dioxygenase (2-ODD), and the three cytochrome P450s (CYP719A23, CYP71CU1, and CYP82D61), would be synthesized and introduced into separate E. coli strains, with each strain expressing one enzyme or a small group of enzymes.

One thing I would have to account for is that plant P450s can be difficult to express as active enzymes in E. coli because they normally require a membrane environment and an electron-donor partner. However, Decembrino et al. demonstrated that these three specific mayapple P450s can be expressed and function in E. coli when co-expressed with the plant-derived NADPH reductase ATR2 from Arabidopsis. Because of this, I would use that approach rather than switching to a eukaryotic host such as yeast or insect cells.

After expression, the E. coli cells would be lysed and the enzymes collected, either as purified enzymes or as semi-crude lysates. These enzyme fractions could then be combined in a single cell-free reactor with the starting substrate, coniferyl alcohol, as well as the cofactors needed for the different reactions. These would include SAM for the methylation steps, NADPH for the P450 reactions, and Fe²⁺, 2-oxoglutarate, and O₂ for the ring-closing reaction.

The idea is essentially to take the pathway out of the cell. Instead of having the engineered E. coli grow and try to produce the final molecule themselves, I would use them to produce the individual enzymes and then combine those enzymes in a controlled cell-free system. The enzymes would act sequentially on the starting substrate until the pathway reaches the desired product, 4’-demethylepipodophyllotoxin, which could then be purified from the reaction mixture.

(3) How would you use automation?

An Opentrons liquid-handling robot would mainly be used for the optimization and screening stage of the cell-free reactor, before moving to a larger-scale system.

I would use the Opentrons to:

• Set up the reactions: dispense different combinations and ratios of the ~8 enzyme fractions, along with different concentrations of the required cofactors, into a 96-well plate. This would allow me to run many small reactions in parallel instead of manually pipetting every condition.
• Take samples: automatically collect small samples from each reaction at set time points so I could see how the pathway is progressing.
• Measure the results: transfer those samples to a plate reader or HPLC/LC-MS system to measure product formation.
• Optimize the conditions: use the results to guide a Design-of-Experiments (DoE) approach, testing different enzyme ratios, cofactor concentrations, pH, temperature, and other conditions. The robot could then set up the next round of experiments based on the results from the previous one. The main reason I would use automation here is that there are a lot of variables to optimize. Once I start combining ~8 enzymes with different enzyme ratios, cofactor concentrations, pH, temperature, and reaction times, the number of possible conditions becomes very large. Having the Opentrons set up these reactions would let me test many more conditions in parallel and make the optimization process much faster and more systematic.

I would use the Opentrons as a prototyping and optimization tool, rather than trying to use it for the final production scale. It is well suited for small-volume, highly parallel experiments, while scaling the optimized reaction to liters or more would require a different system, such as a stirred-tank or flow reactor. I would therefore use the Opentrons to figure out what works, and then transfer those conditions to the larger production system.

(4) How much would you need to make, and how long would it take? (a) How much?

I would set the production target based on enough etoposide to cover the two FDA-approved US indications I am considering: testicular cancer and small-cell lung cancer (SCLC) for one year.

1. To estimate the number of patients

• Testicular cancer: 9,720 new US cases/year
• SCLC: 13% of 226,650 annual US lung cancer diagnoses ≈ 29,465 cases/year
• Total ≈ 39,185 patients/year or 39,200 patients/year

2. To Estimate etoposide required per patient

For a rough estimate, I used standard treatment regimens and an average adult body surface area of 1.9 m².

• Testicular cancer: 100 mg/m² × 1.9 m² × 3 doses/cycle × 4 cycles = 2,280 mg ≈ 2.3 g/patient

• SCLC: 100 mg/m² × 1.9 m² × 4 doses/cycle × 5 cycles = 3,800 mg ≈ 3.8 g/patient

If we use around 3 g etoposide per patient as a rough average,

39,200 patients × 3 g = 117,600 g

then we will have to produce 117,600 g of the drug or 118 kg etoposide/year total.

3. Convert this to the amount of aglycone that needs to be produced

The cell-free pathway would produce the etoposide aglycone rather than etoposide itself. The final sugar-attachment step would be carried out separately.

If I assume an overall 50% yield after the enzyme cascade, purification, and final chemical glycosylation:

118 kg ÷ 0.50 = 236 kg

THen the estimated production target would be around 236 kg precursor/year.

(b) How long?

For the production time, I would use the results from Decembrino et al. as a reference point. Their work reached 137 mg/L for an early pathway intermediate and reported up to 98% yield for a five-enzyme cascade to deoxypodophyllotoxin. Since the full pathway I am proposing would involve additional steps, I would use more or less 100 mg/L as a deliberately rough starting assumption for the final aglycone.

If we made the following assumptions about production

Titer: ~100 mg/L Reactor volume: 1,000 L Batch time: ~2 days

In a cloed batch reactor that would be, per batch:

100 mg/L × 1,000 L = 100,000 mg → 100 g aglycone/batch

Comparing that with the annual target:

236 kg ÷ 0.100 kg/batch = 2,360 batches

At one reactor running continuously:

2,360 batches × 2 days = 4,720 days

4,720 ÷ 365 ≈ 12.9 years

One 1,000 L reactor would take ~13 years. To produce the same amount in one year we would need around 13 reactors running in parallel.

(5) What don’t you know yet?

There are still a few important things I would need to figure out before knowing whether this system would actually work at the scale I am proposing:

1. Real titer and yield for the full pathway: The literature numbers I used are from a 5-enzyme partial cascade that reaches deoxypodophyllotoxin, rather than the full ~8-enzyme pathway to the aglycone. I don’t know yet how much the additional steps would reduce the overall yield. For example, if the actual titer were closer to 10 mg/L instead of my assumed 100 mg/L, the number of reactors needed would increase by roughly 10x.

2. P450 stability and reusability: I am assuming that the enzymes, particularly the P450s, would remain active throughout a roughly 2-day batch. I don’t know if that is realistic in a cell-free system. I would need to test their stability and determine whether they can be used for a full batch, or whether an approach such as enzyme immobilization would be needed to extend their lifetime.

3. Efficiency of the final glycosylation step: My 50% overall process-yield assumption is mainly a placeholder for estimating the scale of the system. I would need to determine the actual yield of converting the purified aglycone into etoposide and understand how much of this step is already established at industrial scale. This could significantly change the amount of aglycone I would actually need to produce upstream.

Overall, the biggest unknown is whether the full pathway can reach a useful titer while keeping the enzymes stable enough for a practical cell-free process. Those are the first things I would want to test experimentally before treating the reactor and production estimates as realistic.

AI

AI Client: ChatGPT & Claude

Prompts:

• What does automation actually mean in a biomanufacturing project? How could I make automation meaningful rather than just having a robot dispense things? Could I use a cell-free manufacturing system instead of a living organism? What would be a good cell-free + Opentrons project? I want something with a strong medicine/cancer angle.
• I want something analogous to the Madagascar periwinkle/cancer-drug story but with a simpler pathway. Are there other plants that help other diseases that I could try to make the active compound? What plant-derived drugs have relatively simple biosynthetic pathways?
• Are there plant-derived medicines that could be made with ~1–5 enzymes? Which of those would actually work for a cell-free system? Which ones would have a good Opentrons automation angle? What about medicines from plants other than the obvious cancer ones?
• Could I make the active compound rather than the whole plant extract? What are some plant → active compound → disease examples?
• What about L-DOPA? What about 5-HTP?What about galantamine? What about colchicine? What about artemisinin? What about camptothecin? What about podophyllotoxin? What about etoposide? What about Taxol/paclitaxel?What about vinblastine/vincristine from Madagascar periwinkle? What about cyclopamine?
• What about marine-derived compounds instead? Could I use a marine enzyme to make a pharmaceutical compound?
• What about a marine brominase/halogenase? What about a marine-derived enzyme that makes a pharmaceutical precursor?
• Moving toward mayapple Is mayapple a good one? What exactly does mayapple make? Is podophyllotoxin itself the drug, or is it a precursor? What cancer drug comes from podophyllotoxin? How complicated is the podophyllotoxin pathway? Can I isolate a 3–5 enzyme section instead of doing the whole pathway? Is there a small enzyme module within the mayapple pathway? Could I make deoxypodophyllotoxin instead of the complete drug? Could I use the pathway toward etoposide rather than trying to make etoposide itself?
• What did the 2015 mayapple paper actually do? Since that 2015 paper, has the tobacco idea gone anywhere? Did they actually make the compound in tobacco? How much did they make? Did later papers improve the tobacco system? What happened after the 2015 pathway-reconstruction paper? Did they identify bottlenecks? Did they improve precursor supply? Did they eventually make more of the etoposide-related compounds? Could the tobacco system be replaced by a microbiological chassis? Could it be replaced by a cell-free system? Would it be correct to say that a microbial or cell-free system could improve on tobacco? How much of the actual drug is now made using tobacco versus the original plant? Is tobacco actually being used commercially to make etoposide? Or is tobacco still just a research chassis? Has anyone moved the pathway into microbes/yeast? Has anyone actually gotten closer to microbial production of etoposide? help me with the scale production

Bibliography:

• Decembrino, D., Ricklefs, E., Wohlgemuth, S., Girhard, M., Schullehner, K., Jach, G., & Urlacher, V. B. (2020). Assembly of Plant Enzymes in E. coli for the Production of the Valuable (−)-Podophyllotoxin Precursor (−)-Pluviatolide. ACS Synthetic Biology, 9(11), 3091–3103. https://doi.org/10.1021/acssynbio.0c00354
• Decembrino, D., Raffaele, A., Knöfel, R., Girhard, M., & Urlacher, V. B. (2021). Synthesis of (−)−deoxypodophyllotoxin and (−)−epipodophyllotoxin via a multi-enzyme cascade in E. coli. Microbial Cell Factories, 20(1), 183. https://doi.org/10.1186/s12934-021-01673-5
• Kuo, H., Wei, Z., Lu, P., Huang, P., & Lee, K. (2014). Bioconversion of Pinoresinol into Matairesinol by Use of Recombinant Escherichia coli. Applied and Environmental Microbiology, 80(9), 2687–2692. https://doi.org/10.1128/aem.03397-13
• Lau, W., & Sattely, E. S. (2015). Six enzymes from mayapple that complete the biosynthetic pathway to the etoposide aglycone. Science, 349(6253), 1224–1228. https://doi.org/10.1126/science.aac7202
• American Cancer Society. (2025). Cancer facts & figures 2025. American Cancer Society report
• Conservation status, anticancer compounds and pharmacological Aspects of Royle: a review. (2022). Indian Journal of Ecology. https://doi.org/10.55362/ije/2022/3637

Week 2 HW: DNA Design

Part 0: Set up in Kernel

The location of my files in Kernel is in the HTBAA26_Genspace repository, file name Paulina Liceaga

Part 1: Define your expression target

The target molecule is a single enzyme found in the Himalayan Mayapple plant, OMT3 or Pluviatolide O-methyltransferase, involved in the biosynthesis of etoposide, a chemotherapeutic compound of the topoisomerase inhibitor family (PubMed:26359402).

The accession code for this protein in UniProt is A0A0N9HMN6 and its genetic sequence is as follows:

MEMAPTMDLEIRNGNGYGDSGEELLAAQAHIYNHIFNFISSMALKCAVELNIPEILHNHQPKAVTLSELVQALQIPQAKSACLYRLLRILVHSGFFAITKIQSEGDEEGYLPTLSSKLLLKNHPMSMSPCLLGLVNPTMVAPMHFFSDWFKRSDDMTPFEATHGASLWKYFGETPHMAEIFNEAMGCETRLAMSVVLKECKGKLEGISSLVDVGGGTGNVGRAIAEAFPNVKCTVLDLPQVVGNLKGSNNLEFVSGDMFQFIPPADVVFLKWILHDWNDEECIKILKRCKEAIPSKEEGGKLIIIDMVVNDHNKGSYESTETQLFYDLTLMALLTGTERTETEWKKLFVAAGFTSYIISPVLGLKSIIEVFP

The enzyme would be expressed in an E.coli chassis and it’ll be produced in the cytoplasm. OMT3 is a soluble enzyme with no known membrane association or secretion signal — so no targeting sequence is required, which meaningfully simplifies this design compared to the P450 enzymes in the same pathway.

Part 2: Genetic parts

For the transcription start, I would use an inducible promoter, such as a T7- or lac-based promoter typical of a pET-style E. coli expression system. An inducible promoter lets the culture grow to a healthy density first, without the metabolic burden of constantly producing OMT3, and then lets expression be triggered on command at the time of the researcher’s choosing — giving more control than a constitutive promoter would.

For the translation start, I would use a ribosome binding site (RBS) of moderate-to-strong strength, a stronger RBS is a reasonable choice to maximize protein yield.

The coding sequence is the OMT3 gene itself: 1,119 base pairs (372 amino acids), including its start and stop codons, sourced from UniProt accession A0A0N9HMN6. This is the actual gene of interest, the sequence that gets transcribed and translated into the functional enzyme.

For targeting, no signal peptide or compartmentalization tag is needed, since OMT3 is a soluble, cytoplasmic enzyme with no known membrane association or secretion signal in its native plant context.

For the transcription stop, I would include a standard terminator (such as a T7 or rrnB terminator) immediately after the stop codon, to ensure a clean end to the transcript and prevent RNA polymerase from reading through into downstream vector sequence.

Finally, for the vector context, I would reuse an existing standard E. coli expression backbone (a pET-type vector, for example), which provides a ColE1-family origin of replication and a kanamycin resistance marker for selection. There’s no need to design a backbone from scratch here — an established, well-characterized backbone already provides everything this construct needs for replication and selection in E. coli.

Part 3 : Assembly

I originally added a T7 lac promoter with a BBa_B0034 Ribosome binding site and a T7 terminator but I then decided to substitute these individual parts for an already assembled backbone, so the construct consists of the pET-28a(+) backbone with the OMT3 CDS positioned for insertion at the vector’s MCS. After removing the redundant standalone promoter/RBS/terminator parts, the construct reduces to two elements: the backbone and the CDS. Kernel’s interface did not allow me to drag the CDS into the interior of the backbone’s MCS region, so in my current Kernel construct the CDS sits next to the backbone rather than literally inserted within it.

I confirmed the reading frame of the OMT3 CDS is intact (1,119 bp = 372 codons + stop, matching the UniProt-reported protein length exactly), with no accidental internal stop codons.

Part 4: PENDING

Part 5: Build

How would you build it? Direct synthetic gene synthesis of the codon-optimized OMT3 CDS as a gene fragment, followed by a standard one-step assembly (Gibson assembly, or restriction digestion/ligation using the BamHI/XhoI sites native to the pET-28a(+) MCS) into the vector. There’s no need for oligo-pool/chip-based synthesis, which is better suited to large combinatorial libraries rather than a single ~1.1 kb ORF.

What does synthesis constrain? A single moderate-GC, soluble-protein CDS like OMT3 is generally straightforward to synthesize, without the long hydrophobic repetitive stretches that make membrane-protein genes (like the pathway’s P450 enzymes) harder to synthesize. The main thing worth checking after codon optimization is whether any new restriction sites were accidentally introduced that would interfere with the BamHI/XhoI cloning strategy, or extreme local GC content that could complicate synthesis — neither is expected to be a major issue here.

Prototype vs. scale? Prototyping is this single synthesized CDS, cloned into the pET-28a(+) backbone and tested in a small BL21(DE3) culture. Producing DNA at meaningful scale (e.g., enough plasmid to seed a larger fermentation run) would not require re-synthesizing anything — it would instead mean standard plasmid maxi-prep amplification from a scaled-up overnight culture, since gene synthesis is a one-time design step rather than a per-batch production step.

Readings-

Part 6: What don’t you know yet?

I’d want to run an induction time-course (with and without IPTG) and quantify OMT3 protein levels (e.g., via SDS-PAGE or a His-tag-based Western blot) to confirm how much of the preotein is being expressed, and if it will express correctly to begin with.

Whether the N-terminal His-tag fusion affects OMT3 folding or activity. I assumed the tag would be a useful, low-impact purification handle, but some enzymes lose activity when fused to even small N-terminal tags. I’d want to express both the tagged and an untagged version and compare enzymatic activity (e.g., an in vitro assay measuring pluviatolide-to-bursehernin conversion) to confirm the tag isn’t interfering.

Whether the real location of the ori/KanR/lacI elements inside the pET-28a(+) block, as I’ve documented them from the standard reference map, actually matches Kernel’s specific copy of this part exactly. Since I couldn’t verify this directly within Kernel’s annotation tools, there’s a chance this specific library part has minor sequence differences from the commercial reference I cited. I’d want to BLAST the full Kernel sequence against the canonical pET-28a(+) reference to confirm an exact match before trusting my documented coordinates fully.

Week 3: Designing a Cell-Free Reaction Composition

A. COMPARE CELL-FREE REACTION COMPOSITIONS

A1. Why does the concentration of potassium glutamate and magnesium glutamate differ between compositions? Which is more sensitive?

The concentrations differ because each formulation was optimized for its own conditions. The lysate strain and preparation method change how much Mg²⁺ and K⁺ the extract already carries over, and the energy system matters too. Jewett & Swartz (2004) had to lower Mg²⁺ when switching from PEP to pyruvate, because PEP binds Mg²⁺ more tightly (11.3 mM total in Cytomim vs about 23 mM in PANOx-SP). Olsen et al. (2026) likewise re-tuned Mg²⁺ and phosphate for each new lysate batch. The two ions also do different jobs. K⁺ is the main cation that sets ionic strength, so it is needed at roughly 80-362 mM across formulations. Mg²⁺ is a cofactor for ribosomes and polymerases, needed at about 8-10 mM.

Magnesium glutamate is the more sensitive of the two. We can observe this in Jewett & Swartz (2004), Table II, PANOx-SP gave maximum yield at 20 mM Mg-glutamate, but yield fell to 39% at 28 mM. So in simple terms, there is a sweet spot on how much magnesium glutamate you should use, use too little and you will not be rising to the full yield potential, too much and the yield will fall. In Cytomim, yield dropped to 67% at 12 mM and 41% at 16 mM, compared with the 8 mM optimum.

Potassium glutamate tolerated far larger changes in the experiments: PANOx-SP kept 92% of maximum yield at 225 mM against a 175 mM optimum, and Cytomim kept 95% at 200 mM against a 130 mM optimum. K⁺ is still required, since Cytomim yield fell to 20% with no added K-glutamate, but it has a broad plateau while Mg²⁺ has a narrow peak.

A2. Explain how a PEP+NTP reaction formulation is different from a Ribose+NMP reaction formulation.

A PEP+NTP formulation supplies energy directly. Phosphoenolpyruvate (PEP) is a phosphorylated energy substrate that transfers its phosphate to ADP to regenerate ATP, and the NTPs (ATP, GTP, CTP, UTP) are added already phosphorylated. In PANOx-SP (Jewett & Swartz, 2004), this means 33 mM PEP plus 1.2 mM ATP and 0.85 mM each of GTP, CTP and UTP.

A ribose+NMP formulation instead relies on the lysate’s endogenous metabolism. It provides sugars (glucose and ribose) and nucleoside monophosphates (AMP, CMP, GMP, UMP), and the lysate generates the energy and converts NMPs to NTPs. Olsen et al. show that this works: their optimized formulation (RFopt) reaches ATP levels comparable to PANOx-SP within 30 minutes despite starting with no added ATP.

A3. Choose two cell-free reaction compositions from the literature, compare and contrast the choice of reagents and the concentration of reagents.

I compared PANOx-SP (Jewett & Swartz, 2004) and the optimized formulation RFopt (Olsen et al.). Both use glutamate as the main anion (following the logic that glutamate better resembles the cell’s interior than acetate), both include all 20 amino acids and a crude E. coli lysate, and both depend on tuning magnesium glutamate carefully.

While PANOx-SP acquires energy in the form of PEP and nucleotides as NTPs. RFopt uses sugars and NMPs and lets the lysate’s metabolism do the conversion.PANOx-SP also includes expensive cofactors such as NAD, CoA, folinic acid and tRNA. RFopt drops all of them except nicotinamide. PANOx-SP also adds purified T7 polymerase, whereas RFopt omits it because a BL21 Star (DE3) strain induced with IPTG is used for the lysate (so we can assume it already comes with the polymerase).

In the case of RFopt, it uses about twice the K-glutamate but 2.5 times less Mg-glutamate, and 2.5 times more of each amino acid. Its lysate fraction is also higher (30% vs 24%). RFopt makes more than double the protein at roughly 1/70th of the cost per gram, by using fewer, cheaper components (12 vs about 20).

Use of AI

AI client: Claude (Anthropic), Claude Sonnet 5.5, web chat, October 2026. I used it to understand the papers and my design choices, and to draft and pressure-test answers to Section A and the sfGFP design. I checked its claims against the papers and the coopetition interface, and I can explain and justify the design.

Prompts used (paraphrased)

Readings

“Can you help me find these readings?” (six citations). Used to confirm the papers exist and find where to read them.

Understanding the papers

Uploaded Jewett 2004 and Smith 2026: “give me the keynote version, explain the key concepts like I’m 15.” Used to get an overview of CFPS, titer vs cost, Cytomim, and the GPT-5 lab. “Doesn’t that mean pyruvate didn’t work and it was the other stuff?” Checked the AI’s explanation against the paper’s own text (pyruvate only worked together with the new environment). Pasted the oxidizing/reducing passage: “explain please.” Used to understand why reteplase needs IAM, GSSG/GSH, DsbC and Δgor while sfGFP does not. Uploaded Olsen’s preprint: “which was the ideal formulation? what protein? help me understand Table 1.” Used to identify RFopt and read the table (cost per gram = $/L ÷ g/L).

Assignment scope and Section A

Pasted the instructions: “help me understand the questions first, should I find literature to validate them?” Used to separate Section A (answered from the papers) from Section B (design). “Help me with A1-A3,” then uploaded Jewett’s Table II: “what do the numbers mean” Used to verify the Mg vs K sensitivity claim against the table. “Explain A2 more clearly.” Used to understand PEP+NTP vs ribose+NMP. “Why is there an extra phosphate if PEP gives it up?” Used to understand how PEP’s phosphate becomes free phosphate (PEP → ATP → spent ATP). “So ribose+NMP is naturally metabolized vs PEP is introduced?” Used to check my understanding. The corrected version is that PEP delivers energy directly, while ribose+NMP makes the lysate generate it.

sfGFP design

Uploaded the reference CSV: “Wouldn’t two variables be fuzzier?” Used to decide to change one thing at a time. “Why not the optimal GFP recipe as base?” Used to compare RFopt with Smith’s recipe. “GFP Ribose has no spermidine, how can that be?” Used to find that Olsen saw no benefit while Smith found only a correlation, so I treat spermidine as a hypothesis. “Wouldn’t it be simpler to test first?” Used to plan a baseline comparison.

Fitting the design to the interface

Screenshots: “is this the RFopt to change?” and “there are two versions of RFopt, which one?” Used to identify the reducing version for sfGFP. The error “Total 22,375 nL exceeds the 20,000 nL reaction”: “should I adapt it so it adds up properly and then make the changes?” This showed the AI’s first mix didn’t fit in 20 µL, so I restarted from the sfGFP Ribose reference column, which does fit.

REFERENCES AND RESOURCES

• Smith, A. A., et al. (2026). Using a GPT-5-driven autonomous lab to optimize the cost and titer of cell-free protein synthesis (Ginkgo Bioworks × OpenAI). bioRxiv preprint. DOI: 10.64898/2026.02.05.703998. https://www.biorxiv.org/content/10.64898/2026.02.05.703998v1

• Olsen, M. L., Copeland, C. E., Sundberg, C. A., Aw, R., Shaver, Z. M., Rao, G., Swartz, J. R., Karim, A. S., &

• Jewett, M. C. (2026). Design-driven optimization of low-cost reagent formulations for reproducible and high-yielding cell-free gene expression. Nature Communications. DOI: 10.1038/s41467-026-69605-8.

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

• Silverman, A. D., Karim, A. S., & Jewett, M. C. (2020). Cell-free gene expression: an expanded repertoire of applications. Nature Reviews Genetics, 21, 151–170. DOI: 10.1038/s41576-019-0186-3.

• Rasor, B. J., Vögeli, B., Landwehr, G. M., Bogart, J. W., Karim, A. S., & Jewett, M. C. (2021). Toward sustainable, cell-free biomanufacturing. Current Opinion in Biotechnology, 69, 136–144. DOI: 10.1016/j.copbio.2020.12.012.

Final Project