Week 1 HW: Final Project Ideas

Revised Idea: Manufacturable APOBEC3A for epigenetic-age measurement

1. What would I like to biomanufacture and why?

I want to biomanufacture APOBEC3A (A3A), the cytidine deaminase that makes bisulfite-free DNA methylation sequencing possible, using E. coli — and to fix the specific reason it is hard to make, which is that the enzyme mutates its own expression plasmid while the bacteria are building it. Methylation patterns are the basis of epigenetic clocks, the main quantitative measure of biological age we currently have, so this is a contribution to aging research, to make a cheaper, more reliable version of the reagent the measurement depends on.

The product: purified recombinant human APOBEC3A — a 199-amino-acid, zinc-dependent single-domain cytidine deaminase (UniProt P31941), sold as an enzyme reagent in microgram to milligram quantities.

What it does. A3A converts cytosine in single-stranded DNA into uracil. Critically, it converts unmodified cytosine and 5-methylcytosine but not TET-oxidized 5-methylcytosine. That selectivity is what makes it useful: DNA sequencing reads the four letters but is completely blind to the methyl tags attached to them. To read methylation you need a chemical step that changes unmethylated letters into a different letter while leaving methylated ones alone, so the methylation pattern shows up as a sequence difference.

For three decades that step was done with sodium bisulfite, which works but chemically shreds the DNA, fragmenting and destroying most of the sample. A3A does the same conversion enzymatically and gently. This is the basis of enzymatic methyl-seq (EM-seq) and ACE-seq, and the published comparison found that EM-seq libraries outperformed bisulfite libraries on coverage, duplication and sensitivity, and worked from as little as 100 pg of input DNA.

Who it is for. Epigenetics core facilities, aging-clock companies, clinical methylation labs, and anyone assembling their own methylation library prep rather than buying a complete kit.

Why I want to make it. Methylation patterns drift with age in a regular enough way that you can estimate someone’s chronological age from a blood sample, and more interestingly, estimate it wrong in a direction that tracks their health and mortality risk. These epigenetic clocks are currently the best quantitative biomarker of biological age we have, which makes them the measuring stick for whether any proposed longevity intervention does anything at all. You cannot evaluate an anti-aging intervention without a way to measure aging.

The specific problem I want to solve. A3A is genuinely hard to produce in bacteria, for a reason that is almost self-referential. The enzyme attacks single-stranded DNA and transcription and replication of a plasmid inside E. coli continuously expose exactly that substrate. So when you ask bacteria to make A3A, the A3A they make deaminates cytosines in its own expression plasmid, producing G→A mutations in the gene that encodes it. It also poisons the host, so the culture grows poorly.

This is documented three independent ways. A PLOS ONE biochemical characterization of E. coli-produced A3A reports that expression was highly toxic to the bacteria and that on induction it caused G→A mutations in the expression vector. Addgene plasmid #187822 carries a standing growth instruction warning that the protein may mutate the vector and that clones should be cultured individually and sequenced after transformation. And a third lab’s published workaround was to insert an L1 intron into the A3A coding sequence specifically to prevent the protein from being made in bacteria at all.

2. What living system would I use?

Chassis: E. coli. A3A is 199 amino acids, cytoplasmic, needs no glycosylation and has no disulfide bonds — it requires a single zinc ion, which comes free in standard media. Yeast buys nothing here and costs growth rate. Cell-free synthesis is the interesting alternative and I discuss it under unknowns.

What goes in:

• A synthetic, E. coli codon-optimized gene encoding His₆-MBP–PreScission–A3A–Gly₆-His₆, ordered as a ~2.3 kb DNA fragment
• A plasmid backbone carrying a tightly repressible promoter — pBAD (araBAD promoter, araC repressor, ampicillin marker) rather than the leaky T7 promoter every published A3A construct uses
• Competent E. coli: a low-mutation propagation strain such as NEB Stable, and an expression strain such as LMG194 or BL21-AI
• Standard growth medium plus glucose (which suppresses the araBAD promoter even further), and L-arabinose as the inducer

What happens, step by step:

1. The synthetic gene is cloned into the backbone by Gibson assembly and the construct is verified by Sanger sequencing.
2. The plasmid is transformed into the expression strain and grown at 37 °C in glucose-containing medium, during which the promoter is held firmly off.
3. At mid-log phase, the temperature is dropped to 18 °C and arabinose is added. The bacteria transcribe and translate the fusion protein for a defined window — hours, not overnight.
4. The maltose-binding protein fused to the N-terminus keeps A3A soluble instead of letting it precipitate into inclusion bodies. This is the strategy both published constructs use and I am keeping it.
5. Cells are harvested by centrifugation and lysed.
6. The fusion is captured on nickel or amylose resin by gravity column, the MBP tag is cleaved off with PreScission protease, and a second nickel pass removes the free tag.

What comes out: purified A3A protein in buffer, assayed for activity and stored frozen or lyophilized.

The engineering change and its rationale. Every published bacterial A3A construct uses a T7 promoter. T7 systems leak: even uninduced, a trickle of protein is produced, and in an overnight culture that trickle has many hours in which to damage the plasmid. The araBAD promoter under glucose catabolite repression is one of the tightest off-switches available in E. coli, and it is the standard tool for expressing proteins that are toxic to their host. My hypothesis is that moving from constant low-level leakage to a short, controlled burst of expression substantially reduces the number of G→A mutations accumulated per expression run, and that this shows up as both cleaner sequence and higher usable yield.

Assumptions I am making:

• That promoter leakage, rather than induced expression itself, accounts for most of the observed plasmid mutagenesis. If deamination happens mostly during the induced window, a tighter promoter delays the problem rather than solving it.
• That the MBP fusion does not substantially inhibit the enzyme in its folded state — the published constructs suggest it does not, since they recover active enzyme.
• That zinc in standard LB or TB is sufficient for full occupancy of the catalytic site, and I do not need to supplement.
• That the host’s own uracil repair pathway is not a dominant variable. It may well be, and strain choice around ung is something I would test rather than assume.

3. How would I use automation?

What I would automate: the expression-condition search, and the activity assay that scores it.

The reason is that my central question is not “does this protein express” but “which combination of conditions minimizes plasmid damage while maximizing usable enzyme” — and that is a combinatorial problem. The variables that plausibly matter:

A full factorial there is 3 × 4 × 3 × 3 × 3 = 324 conditions. Even a sensible fractional design with triplicates is 150–250 cultures. By hand, in 96-deep-well blocks, that is the entire semester and the pipetting error would swamp the effect I am trying to measure. With an OT-2 it is three deep-well blocks, set up reproducibly in an afternoon.

The three workflow steps I would automate:

1. Culture setup and induction. The OT-2 dispenses media, antibiotic, glucose and inducer across 96-deep-well blocks, inoculates from a source plate, and performs the timed inducer additions. Blocks go in a shaking incubator between steps. This is the step where manual error is most likely and most damaging, because a mis-pipetted inducer concentration looks exactly like a real effect.
2. Lysis and small-scale capture. Automated addition of lysis buffer and magnetic nickel beads, wash steps, and elution in plate format. This gives semi-purified enzyme from all 96 wells in parallel without 96 gravity columns.
3. The deamination assay. The OT-2 sets up the kinetic reaction — a 5′-FAM / 3′-quencher single-stranded oligo carrying one internal cytosine, plus UDG and Endonuclease IV — across a 384-well plate with the controls interleaved. The plate reader runs the kinetics. Deamination followed by cleavage separates fluorophore from quencher, so the rate of fluorescence increase is the enzyme’s activity, read for every condition at once.

What stays manual: cloning, and the Sanger sequencing of post-expression plasmids that gives me the mutation-rate number. Sequencing goes out to a service provider.

Why this is the right automation choice. The honest answer is that the OT-2 is not making the protein — the bacteria are. The OT-2 is making the search tractable. Without it I would test maybe eight conditions and probably pick wrong; with it I can run a proper designed experiment and have the statistics to defend the result. That distinction — automation as a way to explore design space rather than as a production tool — is what I take to be the actual lesson of this part of the assignment.

At manufacturing scale, the automation is different. Once a winning condition is fixed, there is no search left and liquid handling is irrelevant. Production automation would be a stirred-tank bioreactor with automated feed, dissolved-oxygen, pH and temperature control running a fed-batch program, feeding an AKTA or equivalent chromatography skid with automated column steps and in-line UV monitoring. If I needed to scale the search rather than the production — for instance, screening hundreds of A3A sequence variants rather than culture conditions — the right tool becomes a cloud lab such as Emerald or Strateos, which gives access to throughput and instrumentation a community lab will never have.

4. How much would I need to make, and how long would it take?

(a) Defining the scale

Target: enough A3A to run one epigenetic-age methylation assay for every adult in Chicago. I picked this because I run a longevity community here and a city-scale biological-age screening program is a concrete thing I can picture.

Chicago has roughly 2.7 million residents, of whom about 78% are adults — call it 2.1 million, rounded to 2 × 10⁶ assays.

(b) Mass of enzyme required

The enzyme loading per reaction is not published by NEB, so I am estimating it from typical in-vitro deamination conditions: A3A is generally used at roughly 0.1–5 µM in a 50–100 µL reaction.

Taking 2 µM in 100 µL as a central case:

n = 2\ \mu\text{M} \times 100\ \mu\text{L} = 200\ \text{pmol}

A3A monomer mass is 199 residues ≈ 23 kDa = 23,000 g/mol, so:

m = 200 \times 10^{-12}\ \text{mol} \times 23{,}000\ \text{g/mol} \approx 4.6\ \mu\text{g per reaction}

Round to 5 µg per reaction (plausible range 0.5–20 µg; this is my single largest source of uncertainty and it moves the final answer by a factor of ten in either direction).

2 \times 10^{6}\ \text{reactions} \times 5\ \mu\text{g} = 1 \times 10^{7}\ \mu\text{g} = \mathbf{10\ grams}

Ten grams of protein covers every adult in Chicago. That number is worth sitting with. At NEB’s conversion-module list price of about $28 per reaction, those same 2 million assays represent roughly $56 million of reagent spend — for ten grams of material. This is why enzyme reagents are a high-margin, low-volume business, and why scale-up is not the hard part of this project.

For comparison, my bottom-up estimate of total worldwide enzymatic methylation conversion is 200,000–1,000,000 reactions per year, which is 1–5 g of A3A per year globally. Ten grams is several years of world supply.

(c) Culture volume required

Assumptions, stated so they can be attacked:

• MBP-A3A fusion is ~68 kDa, of which A3A is ~23 kDa, so A3A is 34% of fusion mass
• Soluble fusion yield in shake flask, current published state: 20 mg/L → 6.8 mg/L A3A-equivalent
• Recovery through capture, cleavage and polish: 50% → ~3 mg/L purified A3A (baseline)
• If my promoter change delivers the ≥2× improvement I am targeting: 7 mg/L
• High-cell-density fed-batch typically gives 10–20× shake-flask biomass; I assume a conservative 50 mg/L at manufacturing scale

(d) How long

At ChiTownBio (bench scale). A realistic cycle is 4 × 1 L flasks per week: day 1 inoculate, day 2 induce and harvest, day 3 lyse and purify, days 4–5 assay and reset. That is 4 L/week.

\frac{1{,}429\ \text{L}}{4\ \text{L/week}} \approx 357\ \text{weeks} \approx 6.9\ \text{years}

Seven years of continuous flask work to supply one city once. This is the honest conclusion of the exercise: a community lab cannot manufacture this, and should not try. What a community lab can do is determine the conditions under which someone else manufactures it well — which is exactly what the project is actually about.

For the course project itself I need perhaps 5 mg of enzyme to run all the assays, which is under 2 L of culture, or about two weeks of bench work.

At manufacturing scale. A single 250 L working-volume stirred-tank fed-batch run at 50 mg/L produces 12.5 g — the whole target in one batch.

Total: roughly 4–6 weeks from inoculum to vialed product, with QC dominating. The biology is a week; the paperwork is a month.

If instead I used a 20 L benchtop fermenter — the largest thing plausibly accessible to a small operation — each batch yields about 1 g, so ten batches at ~10 days each is about 3.5 months of continuous operation.

The gap between seven years and six weeks is entirely about cell density and vessel size, not about anything biological. That is the clearest thing this calculation taught me.

5. What don’t I know yet?

1. Is promoter leakage really the dominant cause of the plasmid mutagenesis?

This is the load-bearing assumption of the whole project. The alternative explanations are at least as plausible: the damage may happen mostly during the induced window regardless of how tight the off-switch is, in which case a better promoter delays the problem rather than solving it. Or the dominant variable may be plasmid copy number, or the host’s uracil-repair capacity — E. coli normally excises uracil from DNA using uracil-DNA glycosylase, so the strain’s ung status could matter more than the promoter does.

How I would find out: run the pET control and the pBAD construct side by side, sequence 24+ plasmids from each across three independent expression runs, and count G→A transitions per kilobase. Add an arm that varies ung status and an arm that varies induction duration at fixed promoter, so I can separate “how long the enzyme existed” from “which switch controlled it.” If the pET control turns out not to mutate much under my conditions, I have no contrast and need a different project — so I would run that control first, before building anything else.

2. How much enzyme does one reaction actually need?

My entire scale calculation rests on an estimate of 5 µg per reaction derived from generic in-vitro deamination conditions, because NEB does not disclose the loading in the EM-seq conversion module. The plausible range spans 0.5–20 µg, which moves the Chicago-scale requirement from 1 gram to 40 grams and changes the economics by the same factor.

How I would find out: titrate my own purified enzyme against a fixed amount of substrate DNA in the fluorogenic assay and find the minimum loading that gives complete conversion in a standard incubation. That gives me a real number for my own enzyme, which is the one that matters, and it is a cheap experiment — one plate.

3. Can enzyme purified without chromatography equipment be clean enough to use?

A3A’s customers put it directly into precious DNA samples. If my preparation carries contaminating nucleases from the E. coli lysate, it will degrade the customer’s DNA and the product is worthless no matter how active the enzyme is. ChiTownBio has no FPLC, so I would be working with gravity columns and magnetic beads, which give lower resolution than a proper chromatography system.

How I would find out: incubate my preparation with a plasmid or a defined DNA ladder under reaction conditions, with no deamination substrate present, and run a gel — any smearing or loss of band intensity means contaminating nuclease activity. Compare against the commercial enzyme as a clean reference. This is a standard reagent QC test and it is the one that decides whether a community-lab preparation can ever be a product rather than a demonstration.

Sources:

Idea: Biomanufacturing a defined recombinant extracellular matrix protein

1. What would I like to biomanufacture and why?

I would like to biomanufacture a recombinant human extracellular matrix (ECM) protein that could be used as a defined coating or component of cell culture matrices.

More specifically, my first target would be a roughly 51 kDa protein made from several functional domains of human fibronectin: FNIII9–10 linked to FNIII12–14. I like this as a starting molecule because the two parts do different useful things. FNIII9–10 contains the RGD cell-binding motif and the PHSRN synergy site, which are involved in integrin-mediated cell adhesion. FNIII12–14 is a heparin- and growth-factor-binding region of fibronectin. The idea is therefore to make one relatively small recombinant protein that combines a cell-attachment function with the ability to interact with signaling molecules in the extracellular environment (Martino & Hubbell, 2010; Ben Abla et al., 2021).

This is interesting to me because a surprising amount of cell biology still depends on extracellular matrices that come from animals. Matrigel, for example, comes from a mouse tumor and is a complex mixture of many proteins and signaling molecules. It works extremely well for many cell culture applications, but its composition is not completely defined and varies between batches (Aisenbrey & Murphy, 2020). That makes it difficult to know exactly what signals cells are receiving.

I am not proposing that this one fibronectin protein would replace Matrigel by itself. Matrigel contains many other components, including laminins and collagen IV. Instead, I see this protein as one defined building block that could eventually be combined with other recombinant or synthetic components to make more controlled extracellular matrices.

The immediate users I have in mind are researchers growing mammalian cells for tissue engineering, organoids, drug discovery, or cell manufacturing. Longer term, similar defined ECM components could also be relevant to things like cultured meat, where growing large numbers of adherent cells without relying on animal-derived materials is important.

This particular domain combination is not completely new; related FNIII9–10/12–14 recombinant proteins have already been studied in biomaterials. For this project that is actually helpful because it gives me a technically realistic starting point. What interests me is whether these kinds of proteins can be manufactured cheaply and reproducibly and then systematically modified to create better matrix components.

2. What living system would I use?

My first manufacturing chassis would be E. coli, probably a recombinant protein expression strain such as BL21(DE3).

The input to the system would be E. coli carrying a plasmid containing a synthetic DNA sequence encoding the fibronectin fragment. The bacteria would be grown in nutrient media and induced to express the recombinant protein. The cells would use their normal transcription and translation machinery to convert the DNA instructions into the protein.

After growth and expression, the bacterial cells would be harvested and broken open. I would initially include an affinity tag on the recombinant protein so that it could be separated from the other bacterial proteins using affinity chromatography. The output would be purified FNIII9–10/12–14 protein.

There is already evidence that this is a reasonable chassis. Recombinant FNIII9–10 has been expressed in E. coli and retained cell-adhesive activity (Ben Abla et al., 2021), and recombinant proteins combining FNIII9–10 and FNIII12–14 have also been reported in biomaterials research (Licht et al., 2019).

One assumption I am making is that my particular construct will express in a soluble and correctly folded form at a useful yield. That is definitely not guaranteed. Protein solubility is one of the things I would want to test early.

3. How would I use automation?

I would use an Opentrons OT-2 mainly during development rather than for the final large-scale manufacturing step.

One of the problems with recombinant protein production is that there are a lot of variables that can affect the final yield: construct design, induction conditions, media, temperature, expression time, purification conditions, and so on. Testing these one at a time by hand would be slow.

I would start by making a small set of related constructs. For example, I would compare FNIII9–10 alone, FNIII12–14 alone, the combined FNIII9–10/12–14 protein, and a few variants of the combined protein. The OT-2 could set up 24- or 96-well expression experiments, make media and reagent mixtures, transfer samples, and prepare assays after expression. I could also use it to test different protein concentrations in a binding assay.

A simple first experiment might contain four protein designs × four expression conditions × three biological replicates = 48 cultures. That is manageable by hand, but it is already enough pipetting that automation would improve consistency. As the design space gets larger, automation becomes much more valuable.

This connects to Petzold and Mukhopadhyay’s discussion of automation in biomanufacturing. They argue that one of the important uses of high-throughput automation is exploring the very large parameter spaces involved in optimizing biological production systems (Petzold & Mukhopadhyay, 2025).

I would not use an OT-2 to manufacture kilograms of protein. Once I had selected a construct and process, I would move from automated micro-scale screening to a controlled fermenter or bioreactor. In other words, the automation used to discover the process would be different from the equipment used to run the process at scale.

4. How much would I need to make, and how long would it take?

(a) How much?

For a first meaningful scale, I will define the goal as producing enough recombinant protein to coat 10,000 standard six-well cell culture plates.

A standard six-well plate has approximately 9.6 cm² of growth area per well, or: 6 wells × 9.6 cm²/well = 57.6 cm² per plate

Fibronectin coating protocols commonly use approximately 1–5 µg/cm². I will use 2 µg/cm² as a reasonable starting assumption. Protein needed per plate: 57.6 cm² × 2 µg/cm² = 115.2 µg

For 10,000 plates: 115.2 µg/plate × 10,000 plates = 1,152,000 µg = 1.152 g of purified protein

I would add approximately 25% extra to account for handling losses, QC samples, and unusable material: 1.152 g × 1.25 = 1.44 g

So my manufacturing target would be roughly 1.5 grams of purified recombinant protein. This calculation is intentionally rough. The optimal coating density for this new protein would have to be experimentally determined, and it might be substantially different from full-length fibronectin.

(b) How long?

The biggest unknown in this calculation is yield because I do not yet know the expression yield of my exact construct. For a first engineering estimate, I will assume that after optimization I can recover 50 mg of purified usable protein per liter of E. coli culture. I am treating this as an assumption rather than a measured value for this construct. Recombinant protein yields in E. coli vary enormously depending on the protein and expression process.

At 50 mg/L: 1,500 mg ÷ 50 mg/L = 30 L of culture That means one approximately 30 L production run, for example in a 50 L fermenter, could theoretically make enough protein for the 10,000-plate target.

Once the strain and process were established, my rough timeline for one batch would be about 3–4 days. I would allow roughly a day for inoculum and fermentation, another day for harvesting, cell disruption and initial purification, and another one to two days for final purification, buffer exchange, concentration and quality control.

So, in a simplified estimate: 30 L culture × 50 mg/L purified yield = 1,500 mg = 1.5 g 1.5 g ÷ 115.2 µg/plate ≈ 13,000 plates That gives some margin above my target of 10,000 plates. This calculation also shows why optimization matters. If the actual purified yield were only 10 mg/L, I would need about 150 L of culture. If I could achieve 250 mg/L, I would need only about 6 L. Yield therefore changes the economics and physical scale of the manufacturing process dramatically.

The scale-up itself would introduce new problems. Conditions that work in a 1 mL culture or a shake flask do not automatically behave the same way in a 30 L fermenter because oxygen transfer, mixing, heat transfer, nutrient availability, and other parameters change with scale. This is one of the reasons biomanufacturing infrastructure has to connect small-scale process development with larger-scale manufacturing rather than treating scale-up as an afterthought (Asin-Garcia et al., 2025).

5. What don’t I know yet?

The three biggest unknowns for me are:

Can E. coli produce enough correctly folded, soluble protein? The domains have precedent in E. coli, but the combined construct could still aggregate or express poorly. I would need to measure soluble yield across different constructs and expression conditions.

Does the purified protein actually retain both functions? Producing a protein is not enough. I would want to test whether the FNIII9–10 portion still supports integrin/cell adhesion and whether FNIII12–14 still binds heparin or relevant growth factors after being expressed as one recombinant molecule.

How much protein is actually needed on a culture surface? My scale calculation assumes 2 µg/cm² based on conventional fibronectin coating conditions. The engineered fragment could require more or less. This number has a direct effect on manufacturing cost and therefore would be one of the first things I would measure.

For me, the most interesting part of the project is that these unknowns connect molecular design to manufacturing. A protein could work extremely well biologically but be impossible to manufacture cheaply. Conversely, a protein could express at very high yield but not provide the biological function that cells need. The goal would be to optimize both at the same time.

References

Aisenbrey, E. A., & Murphy, W. L. (2020). Synthetic alternatives to Matrigel. Nature Reviews Materials, 5, 539–551. doi: 10.1038/s41578-020-0199-8.

Asin-Garcia, E., Fawcett, J. D., Batianis, C., & Martins dos Santos, V. A. P. (2025). A snapshot of biomanufacturing and the need for enabling research infrastructure. Trends in Biotechnology, 43(5), 1000–1014. doi: 10.1016/j.tibtech.2024.10.014.

Ben Abla, A., Boeuf, G., Elmarjou, A., Dridi, C., Poirier, F., Changotade, S., Lutomski, D., & Elm’selmi, A. (2021). Engineering of bio-adhesive ligand containing recombinant RGD and PHSRN fibronectin cell-binding domains in fusion with a colored multi affinity tag. International Journal of Molecular Sciences, 22(14), 7362. doi: 10.3390/ijms22147362.

Licht, C., Rose, J. C., Anarkoli, A. O., et al. (2019). Synthetic 3D PEG-Anisogel tailored with fibronectin fragments induce aligned nerve extension. Biomacromolecules, 20(11), 4075–4087. doi: 10.1021/acs.biomac.9b00891.

Martino, M. M., & Hubbell, J. A. (2010). The 12th–14th type III repeats of fibronectin function as a highly promiscuous growth factor-binding domain. The FASEB Journal, 24(12), 4711–4721. doi: 10.1096/fj.09-151282.

Petzold, C. J., & Mukhopadhyay, A. (2025). From bench to biofactory: High-throughput technologies and automated workflows to accelerate biomanufacturing. Current Opinion in Biotechnology, 94, 103320. doi: 10.1016/j.copbio.2025.103320.

AI use

I used OpenAI ChatGPT (GPT-5.6 Sol) to help turn an initial idea about recombinant extracellular matrix proteins into a specific biomanufacturing proposal, identify relevant literature, check the back-of-the-envelope calculations, and edit the response for clarity. I also asked ChatGPT to make the proposed product specific enough to manufacture and to help estimate the amount needed for a meaningful cell-culture application.