Week 1 HW: Final Project Ideas

Idea 1 — Glumen: a de novo gluten-binding protein

Framing questions

1. What do you want to make? →
2. What living system could make it? →
3. How could automation help you make it? →
4. What would it take to make it at a meaningful scale?

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

Describe the target product or output you would like to make. Why are you interested in making it? Your motivation might be utilitarian—for example, addressing a problem in health, sustainability, food, energy, or manufacturing—but it does not need to be. Your project might instead have a social, cultural, artistic, or personal motivation. Clearly articulate what you want to make, who or what it is for, and why you think it is interesting or important.

I’m interested in producing a gluten binding protein to be a key component in a gluten detector - Glumen.

The goal is to make a very reliable gluten detector for individuals with Celiac Disease - something that can be used remotely at resturants, that is compact, fast, and highly reliable for a broad range of foods.

My wife and 2 daughters have celiac disease and eating out is quite challenging. Even food that is nominally gluten free - free of wheat or obvious gluten containing ingredients is not always safe. Hidden ingredient in sauces, cross contamination, lack of knowledge within the kitchen, all contribute to such challenging circumstances.

The human digestive/immune system obviously reacts to even these small amounts of gluten - and we’d like to detect small amount of gluten directly in a non-lab/compact consumer device.

I think the total system for gluten detection will be complicated including some method (e.g. ethanol, heat, and mechnical agitation) to release the gluten from food and to detect and calibrate thea mount of gluten present. But the key piece would be this de novo protein binder that can be used to bind to gluten and provide opportunity for identification. Existing gluten detection systems have limitations across food types particularly for fermenetd and hydrolyzed foods (e.g soy sauce, fried rice, pho, etc).

(2) What living system would you use?

Propose at least one living-system manufacturing chassis that could produce your target—for example, a microbial, mammalian, or cell-free system. At a high level, describe how the system would make your target product. What goes into the system? What biological process or processes occur? What comes out? You do not need to design the complete biological system yet. Make reasonable assumptions where necessary and identify them.

The living system I unerstand could be an E. coli expression system.

We would introduce the DNA that encodes our protein binder into the E. coli cells. Grow those cells and then extract and purify out our protein.

(3) How would you use automation?

Identify at least one automation technology that you would combine with your living system. This could include a large-scale Cloud Lab, a desktop liquid-handling robot such as an Opentrons, a programmable microfluidics platform such as Nuclera’s, or another automation technology. Describe what part of the biomanufacturing workflow you would automate and why. For example, you might automate liquid handling, culturing, assembly, screening, sampling, measurement, optimization, or another step.

While we can design the protein in-silico - using RF Diffusion, Protein MPNN, and predict the structure with Boltz/Protenix or AlphaFold3 - we will not actually know if the protein folds properly, if it will bind to the gluten fragment as expected, and if it has the selectivity needed to not identify other confoudners. I believe an automated wet lab can help experimentally determine effectiveness.

I had been looking at Adaptyv as a potential vendor/CRO lab that could both express the protein and then test the protein folding, binding strength, binding time, selectivity, etc.

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

Perform a back-of-the-envelope calculation estimating what it would take to manufacture your target at a meaningful scale.

(a) How much?

Define a scale relevant to your project and calculate approximately how much of your target product you would need to manufacture. For example, if you are producing a medicine, you might estimate the amount required to treat 1,000 people, everyone in Massachusetts, or everyone in the world. If you are producing a building material, you might calculate the amount required for one structure or an entire neighborhood. If you are producing an art or design artifact, define an appropriate scale yourself. Clearly show your calculation, including units, assumptions, and sources for important numbers.

While we dont know the exact amount of binder we’d need, for planning we can estiamte 250ug of binder per test. Lets say we are looking to equip a device with 20 tests to begin with.

And we want to get 500 trial devices x 20 ==> 10000 test x 250 ug = 2.5 g of binder

Lets assume some loss (20%) - round up to 3.0g.

This is for production.

For the automation run - I think we need small amount of samples to test binding affinity, binding time etc.

Assay volumes

Gluten Concentration

(b) How long?

Using your proposed living system and automation technology, estimate how long it would take to manufacture that amount. Consider quantities such as production rate, yield, volume, number of parallel experiments or reactors, automation throughput, or whatever parameters are most appropriate for your system. You may discover that the automation technology you would use to prototype your system is very different from what you would use to manufacture at scale. If so, describe both. Again, we are looking for a reasonable engineering estimate. Show your assumptions and calculations.

According to Adaptyv Bio - they can express the protein in about a week, run the measurements by week 2, and have the full data by week 3. 16-21 days in total.

https://www.adaptyvbio.com/services/binding

I believe the in-silico design can be done in a few days. So maybe the full loop is 3-4 weeks in total - from start of a design to getting results from wet-lab.

(5) What don’t you know yet?

After completing your analysis, identify the 2–3 biggest technical unknowns, bottlenecks, or assumptions in your proposed biomanufacturing system. What would you need to learn, measure, or experimentally determine to know whether your idea could actually work?

LOTS. There are a few classes of problem in general with this device.

1. can we separate enough signal of the gluten from the food - this is pure extraction or “gluten release”
2. can we get a binder that can bind to enough of the types of gluten but be selective enough to not bind to the other type of confounders (e.g. maize etc)
3. will the binder survive the diltuion liquid (e.g ethanol)
4. Will the binder provide enough signal - can we attaach a flouresence or something else - so that we can detect the amount of gluten that binds.

I think for this project - we should focus on step 2 - as that is the biomanufacturing part of making a de novo protein

The first test would be to see if we can make a binder that binds strong enough to these small fragments. And test folding, binding affinity etc.

I think the concern will then be is the binder will survive the gluten release step (ethanol, heat, etc.) and if its selective enough amongst the confoudners like maize and those that have similar amino acid seqeunces.

AI use

I’ve been using AI substantially over the past several weeks and months to learn the basics of AI protein design — RFdiffusion, ProteinMPNN, and structure prediction with Boltz/Protenix and AlphaFold3.

For this write-up specifically:

• The idea, motivation, chassis choice, and open questions are mine.
• AI assisted with the estimates in section 4(a) — the per-test binder quantity, the assay volumes, and the resulting ethanol concentrations.
• Claude Code handled formatting and publishing: it converted my plain-text draft into Markdown, built the two tables from my values, and committed and pushed to this repository. It was instructed not to change my wording, and did not. Although it did draft this note :)

Prompt to Claude Code: “can you put Idea 1.txt in very nice attractive markdown. Update the table. DO NOT change the text from my language. Just format nicely.”

Week 2 HW: DNA Design

Project scope and AI assistance

My HW1 project, Glumen, aims to produce a gluten-binding protein for a gluten detector. For HW2, I used an existing computational protein candidate to explore how to move from a protein sequence to a DNA expression design while the gluten-binder work continues.

Much of this write-up was written by OpenAI Codex as an AI-assisted exploration of the topic. Codex also did substantial technical work: retrieving the candidate, generating and checking sequence files, helping inspect the design in VectorBee, and capturing screenshots. This is a learning exercise, not a claim that I independently designed or validated every element. I am still developing my understanding of the terminology and design choices.

We did not have access to Kernel. We used VectorBee to inspect the constructs, with Biopython and DNA Chisel for local sequence generation, optimization, and checks. No Kernel repository was created and no Kernel tools were run. Database similarity searching and biosecurity screening remain incomplete, so this draft does not claim that every assignment requirement has been fulfilled.

The practice candidate was designed to bind human transglutaminase 2 (TG2). It is not a gluten binder, and its binding has not been experimentally validated. Using it gave us a concrete sequence for exploring the DNA-design workflow.

Part 0 — Workspace and tools

The local workspace is HTBAA/HW2/practice-v1. It contains source sequences, an annotated construct, a saved VectorBee workspace, screenshots, and a detailed step-by-step walkthrough. (The data files linked below are published with this page; the full walkthrough is kept in that local workspace.)

The original backbone and modified construct were opened in VectorBee to inspect their maps and sequence annotations. The local folder serves as the homework workspace; it is not a substitute for a findable Kernel repository.

Part 1 — Define the expression target

The intended product is a small protein binder. For this exercise, we used the 90-amino-acid candidate rf_9_dldesign_0, generated in my earlier RFdiffusion and ProteinMPNN work against TG2.

The original candidate sequence is:

DEKAKELLKKAKEAGKKALEEVKKAKEKGDLGAASKAAAEGTEKMGELVQEAVDLGLPQEAIDELKKLLFETGREMMKLVEETREAREKA

Its identity and original source are recorded in the candidate provenance file.

The proposed expression host is E. coli BL21(DE3). The design uses a T7 promoter, so the host must supply T7 RNA polymerase. The protein would remain inside the cell, in the cytoplasm, followed by extraction and purification.

The proposed product is one joined protein:

Start methionine (M) → purification handle (HHHHHH) → short spacer (GSG) → 90-amino-acid candidate

The six histidines, called a His6 tag, provide a handle for affinity purification. The GSG spacer is three amino acids—glycine, serine, glycine—connecting that handle to the candidate. Together with the starting methionine, these additions give a 100-amino-acid protein. The candidate itself is unchanged.

No secretion signal or dedicated tag-cleavage site is included. The proposed purified protein would retain its tag; whether the tag affects binding remains an experimental question.

Part 2 — Start with an existing plasmid and replace one section

I used the pET-28a(+) backbone based on the sequence from SnapGene’s sequence resource, and replaced the original tag and cloning region with the candidate coding sequence.

A plasmid is a circular piece of DNA. Its backbone is the reusable supporting DNA: it lets bacteria maintain the plasmid and provides controls for reading the protein instructions.

The before-and-after comparison is:

We removed the original tag-and-cloning section and supplied our own defined version. We did not add a second set of tags on top of the existing ones. The replacement corresponds to bases 88–250 in the locally oriented source record; those coordinates refer to that record, not necessarily other displays of the same circular plasmid.

The exact sequences, coordinates, and sources are recorded in the parts table.

Part 3 — Assemble and inspect the construct

The expression region follows this order:

Reading controls → ribosome landing site → instructions for our tagged protein → stop-copying signal

In the terminology used on the map:

T7 promoter / lac operator → RBS → tagged candidate CDS → T7 terminator

The final circular plasmid is 5,509 base pairs: 5,369 original bases − 163 removed bases + 303 inserted bases. The 303-base CDS contains 300 bases encoding the 100-amino-acid protein and three bases forming the stop codon.

The annotated sequence was generated locally and inspected in VectorBee. Saved computational checks confirm that:

• The CDS translates into the intended tagged protein.
• There are no internal stop codons.
• Backbone DNA outside the intended replacement is unchanged.
• The proposed assembly fragments reconstruct the final circular sequence.

These checks establish consistency of the sequence design. They do not establish expression, folding, or binding.

Figure 1. The final plasmid in VectorBee, showing the candidate CDS and retained backbone features.

Figure 2. The selected 303-base CDS, with the start codon, His6 tag, GSG spacer, and candidate translation. Overlapping annotations and translation rows describe the same sequence; they are not additional protein copies.

Figure 3. The TAA stop codon tells the cell to stop making this protein. The separate downstream T7 terminator signals the end of copying DNA into RNA.

The annotated GenBank file, saved VectorBee workspace, and check results are retained with the design.

Part 4 — Optimize and quality-check

Codon optimization — completed locally

Because this is a de novo protein candidate, there was no natural gene sequence to use as the starting DNA. We generated a synthetic baseline and used the open-source tool DNA Chisel to optimize codon usage for E. coli while preserving the protein sequence.

A codon is a group of three DNA letters that specifies an amino acid or a stop. Different codons can specify the same amino acid, allowing the DNA to change without changing the protein.

GC content is the percentage of DNA letters that are G or C. Roughly half the letters are G or C in both versions. We tracked it as a sequence-design constraint; higher is not automatically better.

Of the 100 amino-acid-encoding codons, 62 changed, involving 67 nucleotide substitutions. Constraints included limits on overall and local GC content, avoiding long single-base repeats, and excluding selected restriction sites from the coding region.

A codon adaptation index of 1.0 means the selected codons maximize the score against this particular usage table. It does not mean perfect expression. RNA structure, protein folding, and other biological factors remain relevant. The before-and-after codon comparison records the changes.

Sequence similarity search — incomplete

We verified that the designed DNA encodes the intended candidate, but have not searched it against a database of known sequences. I therefore cannot report a closest natural match or conclude that the sequence is novel.

A later step would be to search both the untagged protein and designed CDS, recording the database, settings, closest matches, identity, coverage, and E-values.

Safety screening — incomplete

We did not have access to Kernel’s screening tool and have not run an equivalent biosecurity screen. The local checks address sequence integrity and selected design constraints; they do not substitute for screening against defined biological hazards.

No screening pass or absence of flags is claimed.

Part 5 — From design to physical DNA

For a first prototype, my proposed approach is to obtain a synthetic insert and assemble it with the retained backbone. The saved design specifies a 363-base insert including overlap regions and a 5,206-base retained backbone. The overlapping ends allow the two pieces to join; those shared bases are not counted twice in the final plasmid.

This approach reuses most of the existing vector and introduces only the new coding region. Gibson et al. describe joining DNA fragments through overlapping ends, supporting this proposed assembly strategy. Fragment joining has been checked computationally; no physical assembly has been performed. Gibson et al., 2009

The coding sequence is short and has moderate GC content, but repeated codons and other local features could still affect synthesis. A supplier has not assessed or accepted the design. Primer design and laboratory conditions also remain outside the completed work.

Carr et al. describe error correction for synthetic DNA. The practical lesson for this construct is that synthesis can introduce errors. I would verify the resulting sequence and assembly junctions before interpreting protein experiments; error correction during synthesis would not replace final verification. Carr et al., 2004

For repeated production of the same protein, the goal would be to retain a verified production clone and scale expression and purification. The gene would not need to be chemically resynthesized for every batch. Producing many different binder candidates would require a separate strategy for DNA synthesis, assembly, and identification.

Part 6 — What remains uncertain?

Even a positive result against TG2 would not establish gluten binding. Moving back to Glumen would require an actual gluten-binder candidate, an updated construct, and appropriate binding and selectivity tests.

Open-source follow-up to explore later

We have already used open-source tools for sequence generation, optimization, and checking. If time permits, the next priorities are:

NCBI documents local BLAST+ searching. SecureDNA provides open-source components and screening documentation. We still need to establish the appropriate access and setup, and should not assume its screening scope is identical to Kernel’s.

These are proposed follow-ups, not completed results.

AI use and prompts

OpenAI Codex wrote much of this document and performed substantial portions of the computational workflow. My role included providing the project context, requesting the exploration, reviewing the explanations, and asking for clarification of concepts such as the tag, spacer, coding sequence, and GC content. The document is an account of AI-assisted exploration rather than independent experimental work. No DNA was ordered and no expression or binding experiment was performed as part of this exercise.

The original design-work prompt, as recorded in the walkthrough, was:

“ok if you were going to go ahead and help with this assignemtn can you get th binder sequence or a sample one? maybe we can start with one we have and then we can improve it . make it later? and contuie with the rest of the steps . make sure each step is super clear what you do what happend and that I can follow - take screenshots if need be. I want to follow step by step and see the full restults”

For this write-up, I asked:

“ok can you suggest good answers here - that I can review - say we didnt have Kernel but show the parts we did do and what we could do with open source if we have time (which we shoudl look at later)”

I also requested VectorBee screenshots and a plain-English explanation of the original plasmid and the section replaced. For finalization, I asked:

“lets go ahead and finalize this draft Md - and acknwoledge much of this was written by Ai as an exploration of this topci and didnt have access to Kernel”

Week 3 HW: Cell-Free Reaction Design

Part A — Compare cell-free reaction compositions

Each question below shows my own working answer — arrived at with substantial AI research and help — followed by an AI cleaned-up version. See the note on AI use at the bottom.

Part B — publishing 20 µL reaction compositions for sfGFP, PETase, and Reteplase on coopetition.htgaa.org, plus the Community Reagent contribution, is not included in this write-up yet.

Q1 — Why does the concentration of potassium glutamate and magnesium glutamate differ between compositions? Which is more sensitive?

My answer

Short answer

a/ Why they differ. The concentrations differ because (a) they serve and help catalyze different reactions, and (b) different recipes bind differently and require different amounts of K⁺ or Mg²⁺.

b/ Which is more sensitive. Magnesium is more sensitive, as we can see in Table II — it shows that if you overdo the concentration the yield falls off substantially as compared to potassium for the same increase %.

Longer answer

CFPS requires various reagents and buffers to enable the protein generation process to take place, and potassium glutamate and magnesium glutamate both offer various chemical behaviors (providing charged particles) that facilitate that process.

Potassium has one positive charge and magnesium has two — and therefore they have different binding behaviors and serve different purposes.

Imagine trying to fold a long ribbon whose different sections keep pushing one another away.

RNA has a version of that problem: it has many negatively charged parts, which repel one another. Positive particles help soften those electrical pushes. This allows RNA to adopt useful shapes and interact with other molecules. That matters because a ribosome — the machine that assembles protein — is itself made largely of folded RNA and proteins. Potassium helps establish suitable electrical surroundings for these interactions.

Magnesium:

• Keeping the machinery’s shape — it helps stabilize folded RNA and the ribosome.
• Helping chemical reactions happen — many enzymes need magnesium together with ATP.
• Supporting the right molecular connections — the ribosome must attach to molecules, move, and release them as it builds a protein.

When you add magnesium:

• Some associates with RNA, ATP, and other molecules.
• Some remains free in the liquid.
• Changing the ingredients changes that distribution.

For example, PEP binds magnesium more strongly than pyruvate. When the researchers changed the energy ingredient, they therefore retested the magnesium concentration.

For potassium, the authors also tested its concentration within each complete mixture. Its preferred amount depends on the surrounding salts and other ingredients.

So in short, the concentration differs between compositions because of what each of potassium and magnesium binds to — the contents of the solution.

AI cleaned-up answer

1. Why do the concentrations differ between compositions?

Different recipes contain different ingredients and different starting salt levels. Potassium maintains the ionic environment the protein-making machinery needs. Magnesium supports RNA folding, ribosome structure, and every reaction that runs on ATP — but it also binds other components in the mixture, so only part of what you add stays available. Change the ingredients and you change that distribution. For example, PEP binds magnesium more strongly than pyruvate, so switching the energy substrate requires re-tuning magnesium.

2. Which is more sensitive?

Magnesium, clearly, in the formulations tested in Table II:

A comparable percentage increase costs about a third of the yield for magnesium, versus a few percent for potassium. The practical implication: in a screen, magnesium deserves a narrow, finely-spaced range around the optimum, while potassium can be scanned coarsely.

Source: Jewett & Swartz (2004), Table II and accompanying discussion.

Q2 — Explain how a PEP+NTP reaction formulation is different from a Ribose+NMP reaction formulation.

My answer

PEP — helps replenish ATP so energy-dependent tasks can continue.

NTPs — supply the nucleotide form needed to make RNA; ATP and GTP also support energy-consuming tasks.

PEP + NTP: the NTPs supply the building blocks for making RNA from DNA. ATP also powers energy-consuming processes. As ATP is used and ADP is produced, PEP uses that ADP and helps regenerate ATP so the reaction can keep working.

Ribose — a sugar that enzymes in the lysate process to help regenerate ATP from ADP.

NMPs — nucleotides with one phosphate group. Enzymes transfer additional phosphates from ATP onto NMPs, turning them into NTPs with three phosphate groups. RNA polymerase then uses these NTPs to make RNA by following the DNA template.

Ribose + NMP: ribose helps regenerate ATP from ADP. Enzymes use ATP to add phosphates to NMPs, converting them into NTPs. RNA polymerase then uses those NTPs to make RNA from DNA.

Deeper dive

When ATP powers a task, it is chemically changed. Two common outcomes are:

• ATP → ADP — ATP loses one phosphate group.
• ATP → AMP — ATP loses two phosphates together. Attaching an amino acid to tRNA uses this route.

So ADP comes from ATP being used by reactions in the mixture. Some ATP, ADP, and AMP can also arrive with the lysate.

If we keep consuming ATP without replenishing it, energy-dependent steps slow down or stop — even if amino acids and ribosomes are still available. PEP helps turn used ADP back into usable ATP.

AI cleaned-up answer

The two formulations differ in how far upstream they start, and that single choice drives both cost and reaction behavior.

PEP + NTP is the direct route. The nucleotides arrive already in the triphosphate form RNA polymerase needs, and PEP regenerates ATP from ADP in a single enzymatic step. It is fast and predictable, but PEP is consumed stoichiometrically, releases inorganic phosphate that accumulates and inhibits the reaction, and both PEP and NTPs are costly.

Ribose + NMP pushes work onto the lysate’s own metabolism. Ribose is catabolized through the enzymes already present in the extract to regenerate ATP, and that ATP is then used by kinases in the lysate to phosphorylate NMPs up to NTPs before transcription can proceed. More steps have to work, so the formulation depends more heavily on extract quality — but the inputs are dramatically cheaper, and the slower, metabolically-buffered energy release avoids the phosphate accumulation problem.

The trade-off: PEP + NTP buys speed and simplicity with reagent cost; Ribose + NMP buys cost and sustained energy supply by trusting the lysate to do more of the chemistry. This is exactly the trade-off Q3 quantifies.

Q3 — Choose two cell-free reaction compositions from the literature; compare and contrast the choice of reagents and the concentration of reagents.

My answer

Let’s use Olsen (2026) — we can compare PANOx-SP and RF₀ₚₜ.

All concentrations in mM unless otherwise noted.

Like question 2, we have PEP + NTP vs. Ribose + NMP.

We have:

• 2 vs. 5 mM amino acids
• 57 vs. 75 mM HEPES buffer
• 130 vs. 362 mM potassium glutamate

In RF₀ₚₜ we did not add ammonium glutamate. We also did not add spermidine and putrescine. And there are some different metabolic additives between the two.

The RF₀ₚₜ is double the yield at 1% the cost.

We have more protein building material — with the more amino acids. It seems that the engine Ribose + NMP outperformed the PEP + NTP.

AI cleaned-up answer

I compared PANOx-SP (Jewett & Swartz 2004) and RF₀ₚₜ, the cost-optimized formulation, as evaluated by Olsen et al. (2026).

The energy and nucleotide system is the headline difference. As discussed in Q2, PANOx-SP runs on PEP and NTPs; RF₀ₚₜ runs on ribose, glucose, and NMPs.

RF₀ₚₜ also raises several concentrations:

• Amino acids: 2 → 5 mM each — more protein-building material, and a larger buffer against depletion of whichever residue is limiting.
• HEPES: 57 → 75 mM — greater capacity to resist the pH drift that comes with sustained metabolic activity.
• Potassium glutamate: 130 → 362 mM — a substantially different ionic and metabolic environment.
• Magnesium glutamate: unchanged at 8 mM — consistent with Q1: magnesium is the sensitive parameter and sits near its optimum in both.

RF₀ₚₜ also removes components: ammonium glutamate, spermidine, putrescine, tRNA, folinic acid, NAD, CoA, and oxalic acid are all dropped. In their place it adds phosphate (15 mM) and nicotinamide (4 mM). Several of the removed supplements are among the most expensive per-reaction line items, so removal is doing real work on the cost side as well as the chemistry side.

Result:

The complete RF₀ₚₜ formulation therefore produced roughly 2.2× more protein at about 1.5% of the reagent cost — a ~150-fold improvement in cost per gram.

Important caveat: because many ingredients changed simultaneously, this comparison does not establish which individual changes caused the improvement. The formulations were co-optimized as whole systems, and the components likely interact (the elevated potassium, for instance, may only be beneficial in the context of the ribose-driven energy system). Attributing credit would require a component-by-component ablation.

AI use

Consistent with Week 1, here is how AI was used on this assignment.

My answers are substantially supported by AI help. I’m new to biology, and I used AI heavily to get through these papers — to explain the chemistry, to check my understanding, and to help me pull the numbers out of the published formulations. The “My answer” sections are my own working answers in my own words, written after that back-and-forth; they appear here essentially as I drafted them, with spelling corrected and Markdown formatting applied, but the wording and the reasoning are unchanged. The understanding behind them was built with AI, not independently of it.

The “AI cleaned-up answer” sections are exactly what the label says: Claude restructured my answers into tighter prose, pulled the specific numbers out of my working notes into tables, and in a few places added framing I had not written (the screening implication in Q1, the comparison table and phosphate-accumulation point in Q2, the ablation caveat in Q3). I reviewed each one against the sources.

Claude Code handled formatting and publishing — converting my plain-text notes into this Markdown page and committing it to this repository.

References

• 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.
• Olsen et al. (2026). Cost-optimized cell-free formulation (RF₀ₚₜ).
• Smith et al. (2026). AI-designed cell-free formulations.

Final Project