Week 1 HW: Final Project Ideas

AI-guided dynamic engineering of CHO cells for bispecific antibody biomanufacturing

Jennifer Zayetsy Almeyda Tejada

1. What would I like to biomanufacture and why?

My interest in biomanufacturing comes from both my personal experiences and my previous work with cell-based therapies. During college, I became interested in gene therapies, immunotherapies, and the possibility of engineering cells for therapeutic purposes. I later worked at V&C Biotech in Peru, a biotechnology company developing cell-based regenerative therapies, including applications for burn injuries. I was responsible for research and development, which gave me a lot of independence but also made me realize how much more I wanted to learn. Beyond developing the research itself, I became familiar with some of the requirements involved in producing cell-based therapies under controlled conditions, including clean-room environments and standardized manufacturing processes.

Being responsible for R&D was actually one of the reasons I decided to look for additional training outside the company. Working independently helped me identify the questions I could address with the experience I already had, but also the areas where I still needed more exposure. I wanted to learn technologies and approaches that I had not yet had access to, particularly in cell engineering. This eventually led me to research involving CAR-T cells, where I became interested in how genetic design choices can change cellular behavior.

Working with engineered immune cells also made me realize how quickly the experimental design space can grow. Even when the therapeutic goal is clear, there can be many possible genetic designs and combinations to test, and the effect of one modification can depend on the rest of the cellular system. This made me interested not only in engineering cells, but also in how we could explore these design spaces more systematically using high-throughput experiments, automation, and computational models to decide what to test next.

Through these experiences, I started thinking about a cell not simply as a container for a therapeutic gene, but as a system whose endogenous signaling, transcriptional state, metabolism, stress responses, and capacity to synthesize and process proteins can influence the final therapeutic product. This is what led me to the problem I want to explore in this project: whether we can engineer a manufacturing cell whose behavior is not fixed, but can be programmed to change as the needs of the production process change.

There is also a personal motivation behind this interest. I have seen people in my family face cancer, autoimmune disease, and severe chronic illness, including my grandmother’s cirrhosis and the burden associated with advanced medical care. Growing up in Peru also made me aware that newer therapeutic technologies do not always reach patients in different countries with the same speed or accessibility. I do not think that improving manufacturing alone can solve this problem, since pricing, regulation, health-system capacity, distribution, and many other factors are also involved. Still, manufacturing is one part of the problem that can be approached through biotechnology, and it made me interested not only in how we develop new therapies, but also in how we can produce them more efficiently and reproducibly.

For this project, I would like to manufacture a difficult-to-express bispecific antibody using a dynamically engineered CHO-cell chassis. I chose a bispecific antibody because complex antibody formats can present manufacturing challenges related to expression and the formation of unwanted side products in CHO cells [1,2]. In addition, the sequence and chain configuration of a bispecific antibody can directly affect cell-line generation and productivity [1]. This makes a difficult-to-express bispecific antibody a useful model for testing whether engineering the host-cell state can improve manufacturing.

The main problem I want to address is that mammalian manufacturing cells are commonly engineered using static genetic modifications, such as permanent knockouts or constitutive overexpression. A fed-batch culture is not static: nutrient availability, cell density, metabolite concentrations, stress, and other aspects of cellular physiology change throughout the process. Consequently, traits that are useful during early culture, such as rapid proliferation, may not be the same traits that are desirable later, when the cell must sustain production of a difficult-to-express bispecific antibody [3,4].

I therefore want to ask whether we can engineer a CHO cell that does not remain in one fixed engineered state, but instead transitions from a growth-oriented state to a production-oriented state when production becomes the priority. During an early growth phase, the cells could prioritize proliferation and biomass accumulation. Once sufficient biomass has been generated, the system could transition toward a production-oriented state in which more cellular resources are directed toward bispecific-antibody expression and the cellular processes needed to sustain production.

The central synthetic-biology question would therefore be:

Can synthetic gene regulation and AI-guided cell-state modeling be used to program CHO cells to transition from a growth-oriented state to a production-oriented state and improve the manufacturing of a difficult-to-express bispecific antibody?

I would approach this at two interacting engineering levels. The first would be the synthetic expression program controlling the bispecific antibody. Different promoter, enhancer, and regulatory architectures could be used to control how strongly and when the product is expressed. Because the product contains multiple components, the architecture could also be used to control their relative expression.

The second level would be the endogenous regulatory state of the host cell. Rather than treating individual genes as isolated engineering targets, I would investigate regulatory programs associated with processes such as proliferation, ER proteostasis, protein folding and secretion, metabolism, stress responses, and survival. Existing CHO transcriptomic and perturbation datasets could be used to identify programs associated with growth and productive phenotypes and to prioritize candidate regulators for experimental testing.

Importantly, I would not assume that existing datasets can directly predict the optimal cellular state for my specific bispecific antibody. Instead, they would provide biologically informed starting hypotheses, while product-specific experimental data generated during the project would progressively refine the model.

Artificial intelligence would therefore have two distinct roles.

First, I would use existing transcriptomic and perturbation data to identify regulatory programs associated with different CHO phenotypes. An AI/ML model could help identify regulatory patterns associated with growth and productivity and prioritize interventions predicted to shift cells toward desirable states. Rather than asking only which individual gene correlates with high production, I would ask what regulatory state the cell should be in and which perturbations might produce that state.

The first AI workflow would therefore be:

Existing CHO omics + perturbation data → AI/ML cell-state model → candidate regulatory programs → experimental validation

Second, once candidate regulatory programs and synthetic expression architectures are defined, AI would guide optimization of the engineered system itself. Experimental data describing expression architecture, host-cell perturbation, switching time, measured cellular state, bispecific-antibody productivity, viability, and product quality would be used to update a predictive model after each experimental round.

The second AI workflow would therefore be:

Expression architecture + host-cell perturbation + switching time + measured cell state → AI/ML optimization model → improved bispecific-antibody production

Ultimately, promising regulatory programs could be incorporated into an inducible synthetic circuit. A first proof-of-concept would use an externally controlled signal to trigger the transition from growth to production. This would allow me to experimentally determine whether temporal control provides an advantage and to optimize when the transition should occur.

If this strategy is successful, a future version could incorporate a sense-and-respond circuit, in which the cell detects an intracellular or extracellular cue and automatically activates a predefined regulatory response [3]. The initial project would therefore focus on demonstrating programmable control with an inducible switch rather than attempting to build a fully autonomous manufacturing cell immediately.

The broader goal would be to treat the manufacturing cell as a programmable biological system rather than a static protein-production host.

2. What living system would I use?

I would use Chinese hamster ovary (CHO) cells engineered to produce a difficult-to-express bispecific antibody as the living manufacturing chassis. CHO cells are widely used for commercial production of recombinant therapeutic proteins because they can grow in suspension, can be adapted to chemically defined media, and can perform the folding and post-translational processing required for complex recombinant proteins [5].

What goes in, what happens, and what comes out

The physical inputs would include an engineered CHO cell line, chemically defined basal medium and feeds, oxygen and other controlled culture inputs, and a genomic expression system encoding the bispecific antibody.

Inside the cell, the antibody components would be transcribed and translated, folded and assembled in the endoplasmic reticulum, processed through the secretory pathway, and secreted into the culture medium.

The final desired output would be correctly assembled bispecific antibody, rather than simply total secreted recombinant protein. This distinction is important because increasing expression would not necessarily improve manufacturing if the additional protein is incorrectly assembled, forms unwanted side products, or has other product-quality problems.

Engineering layer 1: the bispecific-antibody expression program

The first synthetic-biology intervention would control bispecific-antibody expression.

Instead of using only a constitutive expression system, I would test regulatory architectures that allow control over expression strength and timing. Relative expression of the antibody components could also be varied, since the sequence and configuration of bispecific antibody formats can influence their productivity in CHO cells [1].

In the simplest dynamic system, bispecific-antibody expression would remain relatively low during the growth phase and would increase when the cells enter the production phase.

Engineering layer 2: the host-cell state

The second intervention would control the endogenous state of the CHO cell.

I would initially use transcriptomic and perturbation datasets to identify regulatory programs associated with productive CHO phenotypes. Candidate regulators could then be tested experimentally using CRISPRi or CRISPRa to determine whether decreasing or increasing their activity shifts cells toward the desired state.

There is already evidence that host-cell perturbations can affect the balance between growth and productivity. For example, an arrayed CRISPR screen of 187 candidate genes in CHO cells producing difficult-to-express complex antibody formats found that MYC depletion increased product expression by more than 40% through increased cell-specific productivity, while the engineered cells showed lower cell densities [2]. This provides an example of the type of growth-production trade-off that a dynamic system could potentially exploit.

Other regulatory programs could also be investigated. ER proteostasis and secretory capacity could represent one candidate module, while stress-response or cell-survival pathways could represent others. These would be starting hypotheses rather than predetermined solutions.

Engineering layer 3: dynamic control

The key difference from static host-cell engineering would be when these programs are active.

The first proof-of-concept would use an externally controlled inducible switch:

Growth phase → external induction → production phase

During the growth phase, the cells would prioritize proliferation and biomass accumulation. At a selected time or cell density, an external signal would increase bispecific-antibody production and activate the selected production-supporting regulatory program.

The switching time itself would therefore become an experimental variable rather than being chosen arbitrarily.

If this approach provides an advantage over static engineering, a later version could implement a sense-and-respond system:

Biological/process signal → synthetic sensor → regulatory circuit → production state

Potential signals could include cell density, metabolite accumulation, growth state, or cellular stress. Dynamic CHO engineering using synthetic sense-and-respond programs has been proposed as a way to adapt cell behavior to changing conditions during bioprocessing [3].

The sense-and-respond system would therefore be a future extension, rather than something I would need to construct during the initial 12-week proof-of-concept.

What would I measure?

The primary manufacturing outcome would be correctly assembled bispecific-antibody titer in g/L.

I would also measure:

• viable cell density;
• percent viability;
• growth rate;
• cell-specific productivity;
• expression stability;
• timing and reproducibility of the state transition;
• aggregation and other relevant product-quality attributes.

Higher total expression would not automatically count as an improvement if a large fraction of the bispecific antibody is incorrectly assembled, if viability decreases substantially, or if product quality deteriorates.

3. How would I use automation?

Automation would be especially useful during the discovery phase because dynamic engineering creates a multidimensional design space.

Instead of testing only different expression constructs, the experimental variables could include:

expression architecture × host-cell perturbation × switching time × induction level

For example, testing 12 expression architectures, 10 host-cell perturbations, and 3 switching conditions would already produce 360 possible combinations before biological replicates, controls, or additional induction levels are included.

I would use automated liquid handling to standardize repetitive steps such as reagent normalization, plate setup, cell seeding where feasible, media addition, sampling, dilution, and preparation of analytical assays.

An AI-guided design-build-test-learn loop

I would organize the experimental workflow as a design-build-test-learn (DBTL) cycle, with AI contributing at two different stages.

AI role 1: Modeling and prioritizing cellular states

At the beginning of the project, I would integrate existing CHO transcriptomic and perturbation datasets describing growth, productivity, stress responses, and other relevant phenotypes.

Rather than assuming that these datasets can directly reveal the optimal state for my specific bispecific antibody, I would use them to generate biologically informed starting hypotheses. An AI/ML model would identify regulatory patterns associated with different phenotypes and prioritize candidate regulatory programs for experimental testing.

The workflow would therefore be:

Existing CHO omics + perturbation data → AI/ML cell-state model → candidate regulatory programs → CRISPRi/CRISPRa testing

This computational analysis could begin during the first weeks of the project while the experimental system and automation workflow are being prepared.

AI role 2: Optimizing the engineered system

The second AI application would use data generated specifically in this project.

Each experimental condition would generate information about the bispecific-antibody expression architecture, host-cell perturbation, switching time, cellular phenotype, productivity, viability, and product quality.

These data would be used to train and iteratively update a predictive model relating:

Engineering design → measured cellular state → bispecific-antibody manufacturing phenotype

The model would then prioritize the next experiments, including both combinations predicted to perform well and experiments that would reduce uncertainty in poorly explored regions of the design space.

Therefore, the two AI tasks would be complementary:

1. Cell-state modeling: use existing biological data to identify which regulatory programs may distinguish growth- and production-oriented states and prioritize initial interventions.
2. Design optimization: use product-specific experimental data generated during the project to learn which combinations of bispecific-antibody expression architecture, host-cell regulation, and switching time produce the best manufacturing phenotype.

The complete DBTL cycle

During the design stage, existing data and the initial cell-state model would be used to select expression architectures, host-cell perturbations, and switching conditions.

During the build stage, automated liquid handling could normalize DNA and CRISPR reagents, prepare selected combinations according to predefined plate maps, distribute reagents, and standardize cell-plating procedures.

During the test stage, automation could assist with media additions, feeding, sampling, dilution, and assay preparation. For each condition, I would record the synthetic architecture, regulatory perturbation, switching condition and timing, cell growth, viability, bispecific-antibody titer, and correctly assembled product. Selected conditions could also undergo transcriptomic and product-quality analysis.

During the learn stage, the new experimental data would update the AI/ML optimization model and guide the next experimental round.

AI would therefore not replace the biological experiments. Its purpose would be to use existing information to make better initial choices and then learn from product-specific experimental data as they are generated.

In silico planning before physical experiments

Before physically running an experiment, computational tools could also be used to check automated plate layouts and liquid-handling protocols for problems such as insufficient reagent volumes, incorrect transfers, missing controls, or inappropriate tip use.

The biological design space could then be explored computationally before committing cells and reagents. This would be particularly useful because exhaustive experimental testing would quickly become impractical as additional regulatory targets and switching conditions are introduced.

Prototype automation versus manufacturing automation

The automation used during discovery would be different from the automation used during manufacturing.

During prototyping, the priority would be parallelization and rapid learning through multiwell plates, automated liquid handling, standardized measurements, and repeated DBTL cycles.

Once a stable engineered system is selected, the process could move toward controlled stirred-tank bioreactors. At this stage, automation would monitor and control variables such as temperature, pH, dissolved oxygen, agitation, feed addition, and the timing or signal used to trigger the production state.

The bispecific antibody would then be harvested and purified through downstream operations such as clarification, capture chromatography, polishing, concentration, and formulation.

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

Back-of-the-envelope production target

Commercial monoclonal-antibody manufacturing provides a useful reference point for the scale of mammalian-cell biomanufacturing. Modern CHO fed-batch processes can reach monoclonal-antibody titers in the multi-gram-per-liter range, with values as high as approximately 10 g/L reported for highly optimized processes [3]. Difficult-to-express complex antibody formats, including bispecific antibodies, can have lower manufacturing yields than conventional monoclonal antibodies [2].

Because I am using a difficult-to-express bispecific antibody as a model product rather than selecting a specific approved therapy with a defined clinical dose, I would use the scale of monoclonal-antibody manufacturing as a reference and choose 1 kg of purified bispecific antibody as an illustrative manufacturing target.

Assuming an overall downstream recovery of 70%:

1 kg purified bispecific antibody

↓

70% downstream recovery

↓

1 kg / 0.70 = 1.43 kg required upstream

For an illustrative difficult-to-express baseline, I will assume a usable bispecific-antibody titer of 1 g/L:

1,430 g / 1 g/L = ~1,430 L of culture

If dynamic engineering increases usable bispecific-antibody titer to 3 g/L:

1,430 g / 3 g/L = ~477 L of culture

Therefore, for the same 1 kg of purified bispecific antibody:

Baseline at 1 g/L → ~1,430 L of culture

Dynamic engineering target at 3 g/L → ~477 L of culture

This corresponds to approximately a 67% reduction in required culture volume.

The 1 and 3 g/L titers are illustrative engineering assumptions rather than reported values for a particular bispecific antibody. The purpose of the calculation is to show how improving the amount of usable product produced per liter could affect manufacturing scale.

This does not mean that manufacturing cost would decrease by 67%. Final cost also depends on media, facilities, purification, quality control, failures, fill-finish, regulatory requirements, and other factors.

Importantly, the relevant titer would be usable, correctly assembled bispecific antibody, rather than simply total secreted protein.

Proposed 12-week proof-of-concept timeline

Because the course runs from October through December, I would frame the project as a 12-week proof-of-concept rather than an attempt to develop a manufacturing-ready cell line.

Computational analysis, experimental preparation, and automation development could overlap rather than occurring strictly sequentially. In particular, the first AI analysis could begin immediately using existing datasets while the experimental system is being prepared.

A 12-week proof-of-concept would be deliberately limited in scope. The goal would not be to generate a GMP-ready manufacturing cell line within the duration of the course, but to test whether dynamic control of CHO cell state provides a measurable advantage over static engineering for production of a difficult-to-express bispecific antibody.

The timeline assumes access to a parental producer line, basic analytical assays, and the necessary cell-culture and automation infrastructure. Development of a stable manufacturing cell line, scale-up, process validation, and GMP implementation would require substantially more time.

What is the benchmark?

The main comparison would use the same bispecific antibody and CHO background, allowing the experiment to test whether temporal control itself provides an advantage.

If the dynamically engineered cells outperform the baseline but not the optimized static cells, then dynamic control may not justify its additional complexity.

If the dynamic system produces more usable bispecific antibody, maintains better cell fitness, or improves the balance between growth and productivity compared with the static system, that would provide evidence that when an engineering intervention is active matters in addition to which intervention is used.

Indicators of success

The primary manufacturing indicator would be correctly assembled bispecific-antibody titer.

Cell-performance indicators would include viable cell density, percent viability, growth rate, and cell-specific productivity.

Genetic and regulatory indicators would include stability of the engineered system and reproducibility of the growth-to-production transition.

Product-quality indicators would include aggregation and selected glycosylation or other relevant quality measurements.

AI/modeling indicators would include predictive performance on held-out experimental conditions and whether the model can prioritize improved designs using fewer experiments than an exhaustive search.

Automation indicators would include the number of designs screened per cycle, protocol failure rate, missing-data rate, and reproducibility across plates.

The project would therefore be treated as a multi-objective optimization problem, rather than simply maximizing antibody concentration.

5. What do I not know yet?

Unknown 1: What cellular states should I actually engineer?

The central hypothesis assumes that different cellular states may be useful during growth and production, but I do not yet know what the optimal states are.

Existing transcriptomic data may identify genes and regulatory programs associated with productive phenotypes, but correlation does not demonstrate that changing those programs will improve bispecific-antibody production.

I would therefore use existing datasets to generate hypotheses and CRISPRi/CRISPRa perturbations to test them experimentally.

Unknown 2: Is there enough existing data for the initial AI model?

Existing CHO transcriptomic and perturbation datasets come from different cell lines, products, culture conditions, and experimental designs. I therefore would not expect a model trained only on public data to perfectly predict the optimal state for my specific bispecific-antibody producer.

Instead, existing datasets would be used to identify broad regulatory patterns and prioritize the initial design space. Product-specific data generated during the DBTL cycles would then progressively update the model.

This makes the AI component iterative rather than assuming that an existing dataset already contains the answer.

Unknown 3: What should trigger the transition?

The simplest proof-of-concept would use an externally controlled inducible signal because this would allow the timing of the transition to be varied experimentally.

A more advanced system could respond autonomously to a biologically relevant cue such as cell density, metabolite concentration, growth state, or cellular stress [3].

I do not yet know which signal would provide the most robust and reproducible transition under manufacturing conditions.

Unknown 4: Which regulatory programs should change?

Growth, metabolism, ER proteostasis, secretion, stress responses, and apoptosis are all plausible targets, but their effects may depend on the product and culture conditions.

For example, CRISPR screening has shown that MYC depletion can increase cell-specific productivity for difficult-to-express complex antibody formats while reducing cell density [2]. This is useful evidence for a growth-production trade-off, but it does not mean that MYC will necessarily be the optimal intervention for my specific bispecific antibody.

Candidate regulators would therefore be treated as hypotheses to test rather than predetermined solutions.

Unknown 5: Can dynamic engineering actually outperform static engineering?

A more complicated genetic circuit is not automatically a better manufacturing system.

This is why I would directly compare:

baseline bispecific producer → optimized static producer → dynamic producer

If constitutive engineering produces the same improvement with less complexity, there may be little reason to use a dynamic circuit.

The experiment therefore directly tests whether temporal control itself adds value.

Unknown 6: Does the system improve correctly assembled bispecific antibody rather than simply total expression?

This distinction is particularly important for a bispecific antibody. Increasing transcription or total secreted protein would not necessarily improve manufacturing if the additional protein is incorrectly assembled or forms unwanted side products. The sequence and configuration of bispecific formats can themselves affect production in CHO cells [1].

I would therefore distinguish between total expression and the amount of usable, correctly assembled bispecific antibody, while also monitoring relevant product-quality attributes.

Unknown 7: Will the engineered behavior survive scale-up?

A multiwell plate and a stirred-tank bioreactor differ in oxygen transfer, mixing, nutrient delivery, cell density, and metabolite accumulation.

This could be particularly important for dynamic or sense-and-respond circuits because the signal used to trigger the state transition may behave differently at scale.

The 12-week project would therefore focus on proof-of-concept validation rather than claiming that the system is already ready for industrial scale-up.

What would I learn even if the target is not reached?

Even if the engineered cells do not reach the illustrative 3 g/L target, the project could reveal whether growth and production represent sufficiently different cellular states to justify dynamic control, which regulatory programs distinguish those states, and whether synthetic switching provides an advantage over static engineering for bispecific-antibody production.

It could also reveal that the optimal solution is simpler than expected. For example, a static regulatory perturbation may perform as well as a dynamic circuit, or the main limitation may remain the bispecific-antibody expression architecture rather than the host-cell state.

Either outcome would provide information for another design cycle.

What interests me most is therefore not only increasing the amount of bispecific antibody produced, but learning whether we can control when different cellular programs are active to make a mammalian manufacturing cell more programmable.

References

1. Hussain H, Ozanne AMS, Patel T, Vito D, Ellis M, Hinchliffe M, Humphreys DP, Stephens PE, Sweeney B, White J, Dickson AJ, Smales CM. Sequence and Configuration of a Novel Bispecific Antibody Format Impacts Its Production Using Chinese Hamster Ovary (CHO) Cells. Biotechnology and Bioengineering. 2025;122(2):435–444. doi:10.1002/bit.28879.

2. Bauer N, Oswald B, Eiche M, Schiller L, Langguth E, Schantz C, Osterlehner A, Shen A, Misaghi S, Stingele J, Ausländer S. An arrayed CRISPR screen reveals Myc depletion to increase productivity of difficult-to-express complex antibodies in CHO cells. Synthetic Biology. 2022;7(1):ysac026. doi:10.1093/synbio/ysac026.

3. Donaldson J, Kleinjan DJ, Rosser S. Synthetic biology approaches for dynamic CHO cell engineering. Current Opinion in Biotechnology. 2022;78:102806. doi:10.1016/j.copbio.2022.102806.

4. Donaldson JS, Dale MP, Rosser SJ. Decoupling Growth and Protein Production in CHO Cells: A Targeted Approach. Frontiers in Bioengineering and Biotechnology. 2021;9:658325. doi:10.3389/fbioe.2021.658325.

5. Kim JY, Kim YG, Lee GM. CHO cells in biotechnology for production of recombinant proteins: current state and further potential. Applied Microbiology and Biotechnology. 2012;93(3):917–930. doi:10.1007/s00253-011-3758-5.

AI use disclosure

Prompt used: “Help me organize and refine my idea for HTBAA Homework #1 on mammalian-cell biomanufacturing. I want to explore whether synthetic biology can be used to engineer CHO cells producing a difficult-to-express bispecific antibody so that they dynamically transition between growth- and production-oriented states. I am interested in combining synthetic gene regulation, host-cell engineering, CRISPRi/CRISPRa perturbations, transcriptomic analysis, AI-based modeling of cellular states, AI-guided experimental design, and laboratory automation. Help me identify a meaningful manufacturing problem, determine whether existing CHO data could support the modeling approach, compare static and dynamic engineering strategies, check the biological and engineering assumptions using published literature, work through scale calculations, and improve the clarity and organization of the proposal.”

I used ChatGPT to help organize my ideas, compare possible project directions, check technical assumptions and calculations, identify and verify relevant sources, and edit portions of the text for clarity. I wrote, reviewed and revised the final proposal and sources and can explain the experimental design, assumptions, and calculations presented here.

Week 3 HW: Cell-Free Reaction Design

Literature-guided design and optimization of cell-free reaction formulations

Jennifer Zayetsy Almeyda Tejada

1. Objective

The goal of this project is to design 20-µL cell-free protein synthesis (CFPS) reactions for sfGFP, PETase/TfCut2, and Reteplase that reduce reagent cost while maintaining or improving production of functional protein. The three targets provide an increasing challenge in folding: sfGFP is a robust reporter with no disulfide bond, TfCut2 contains one disulfide bond, and Reteplase contains nine disulfide bonds and therefore requires substantially stronger control of the oxidative folding environment.

The design strategy used here is intentionally different from simply copying one published formulation. We first benchmark the formulations already represented in the course literature table, then identify the biochemical and economic rationale behind their differences. Finally, we use those observations to construct target-specific candidate formulations. The numerical concentrations are therefore literature-constrained hypotheses, not experimentally validated optima.

Inventory note. The formulations below are the proposed designs we derived from the literature benchmark. Components that are available in the current inventory are used directly. If a proposed component is not represented in the inventory or is not recognized by the Reagent Supplement interface, it is retained as a documented proposal and should be listed in the submission notes rather than silently replaced by another reagent.

2. Experimental constraints

Each reaction has a total volume of 20,000 nL (20 µL).

Fixed components:

The remaining 11,000 nL must be composed of available reagents and nuclease-free water. Reagent volumes are constrained to multiples of 25 nL.

The 10X base buffer contributes approximately 30 mM HEPES, 150 mM potassium glutamate, 2.6 mM magnesium glutamate, 1 mM 17-amino-acid mix, 1 mM tyrosine, and 1 mM cysteine to the final reaction. Therefore, concentrations reported below are final concentrations after accounting for the fixed base buffer.

3. Literature benchmark

3.1 PANOx-SP versus RFopt: two fundamentally different design philosophies

The literature table assembled by Olsen et al. provides a useful progression [1] from conventional phosphorylated-energy formulations to low-cost formulations. PANOx-SP uses PEP as the principal energy substrate and supplies ATP-generating capacity through a conventional phosphorylated-energy system. RFopt instead removes PEP and uses a combination of non-phosphorylated substrates, NMPs, ribose, and other low-cost components.

In Olsen et al.’s direct benchmark, PANOx-SP produced the highest sfGFP yield among the previously published formulations tested in their BL21 Star (DE3) lysate, but non-phosphorylated systems had substantially lower reagent cost per gram of protein. Their subsequent optimization produced RFopt at 2.4 ± 0.3 g/L sfGFP and approximately $60/g protein in reagent cost, compared with $4,083/g for PANOx-SP. This is a 99% reduction in reagent cost while increasing yield relative to PANOx-SP.

The important lesson is that yield and cost are not optimized by the same reagent choices. PEP can support high initial energy regeneration, but it is expensive and generates phosphate. Non-phosphorylated systems can lower cost and sustain ATP availability by relying on the metabolic capacity remaining in the lysate.

The values above should be interpreted as formulation comparisons rather than universal optima because lysate composition, reaction geometry, and oxygen transfer also affect performance.

3.2 Component-by-component benchmark

Salts and ionic environment

Question A1: Why do potassium glutamate and magnesium glutamate differ between formulations, and which is more sensitive?

They perform different biochemical roles and are not interchangeable. Potassium glutamate mainly establishes ionic strength and affects the intracellular-like environment of the lysate, whereas magnesium is directly involved in ribosome function, nucleotide binding, and many enzymatic reactions. Consequently, increasing K-glutamate can be tolerated over a much broader range than magnesium. Independent active-learning work has also identified magnesium as the more sensitive variable, with a sharp expression optimum around 10 mM in one IVTT system. Thus, a reasonable design principle is to treat Mg²⁺ as a narrow-range variable and K⁺ as a broader-range variable.

Energy systems

Question A2: PEP+NTP versus ribose+NMP

A PEP+NTP formulation supplies both an energy substrate and the final triphosphate forms required for RNA synthesis. PEP rapidly supports ATP regeneration but generates inorganic phosphate as it is consumed. In contrast, a ribose+NMP formulation supplies less phosphorylated, less expensive nucleotide precursors and relies on the metabolic machinery remaining in the lysate to regenerate NTPs. Olsen et al. showed that RFopt began without exogenous ATP yet maintained approximately 250 µM ATP throughout the reaction, while PANOx-SP generated an early ATP surge followed by depletion and substantially more phosphate accumulation.

This is not simply a substitution of one reagent for another. It changes the metabolic architecture of the reaction.

Nucleotides, ribose and nucleosides

NTPs are among the expensive components of classical CFPS. Olsen et al. found that NMPs could support a much cheaper formulation. Importantly, the replacement strategy is not uniform across nucleotide classes.

• GMP could be replaced by guanine + ribose without loss of protein yield.
• UMP replacement by its corresponding base + ribose reduced yield.
• Complete replacement of NMPs by nucleosides or bases was generally detrimental.
• Smith et al. independently found that partial replacement is beneficial and that AMP can be partially replaced by adenosine.
• Smith et al. also found that attempts to replace CMP and UMP with their nucleosides did not work.

Therefore, our designs retain NMPs rather than eliminating them completely. We use ribose as a metabolic precursor while preserving sufficient direct NMP input.

Buffers

HEPES varies substantially between formulations. Classical PANOx-SP uses approximately 57 mM, whereas RFopt uses 75 mM. Smith et al. observed that higher HEPES could substantially improve titer by preventing pH-driven yield collapse, although the optimum depended on the rest of the formulation.

This is important for our designs because buffer is not treated as a passive background component. We deliberately retain moderate HEPES rather than minimizing it solely for cost.

Polyamines

Spermidine and putrescine are present in many classical formulations. They can affect nucleic-acid interactions and ribosomal function. Smith et al. found a strong association between spermidine addition and increased titer in their experimental search [6].

Our designs therefore retain a modest spermidine concentration rather than eliminating it completely. We deliberately use a lower level than the 1.5 mM commonly seen in PANOx-SP because the objective is to balance performance with reagent cost and reaction volume.

Cofactors

NAD and CoA occur in classical formulations, but Olsen et al. excluded expensive tRNA, CoA and NAD+ from their optimization because previous work showed that these components can be unnecessary in non-dialyzed lysates. Smith et al. later found nicotinamide in all of their top-performing formulations.

We therefore distinguish NAD+ from nicotinamide. Our designs do not add NAD+, while sfGFP and PETase retain a moderate nicotinamide concentration because it repeatedly appeared in the later AI-guided formulations.

Oxygen and redox

Oxygen is not merely a nuisance variable. Olsen et al. showed that RFopt became more sensitive to oxygen transfer than PANOx-SP [1]. Increasing the oxygen feed in a bioreactor from 20% to 100% increased sfGFP production by approximately 45%, reaching 3.7 ± 0.2 g/L.

Because the present reaction platform does not provide an adjustable oxygen-feed parameter, we use available oxygen-related chemistry conservatively. For sfGFP we include catalase as a candidate component because it appeared in the experimentally successful Smith formulation. This should be interpreted as a hypothesis about oxidative stress management, not as equivalent to increasing dissolved oxygen.

Disulfide components

The three proteins deliberately span increasing folding difficulty:

• sfGFP: no structural disulfide bond.
• TfCut2/PETase: one disulfide bond.
• Reteplase: nine disulfide bonds.

Therefore, applying one oxidative formulation to all three targets would not be biologically justified.

Olsen et al. demonstrated that oxidative RFopt containing glutathione and DsbC can support disulfide-containing proteins, but also showed that indiscriminate addition of IAM/GSSG/GSH can dramatically reduce sfGFP production. Importantly, increasing HEPES pH to 7.5 recovered yield by preventing the pH decrease caused by glutathione.

Our design therefore uses a gradient of oxidative pressure rather than a binary reducing/oxidizing switch.

4. Additional literature contrasts

4.1 Magnesium is a sensitive variable

Jewett and Swartz reported very different Mg-glutamate concentrations across PANOx-SP and Cytomim, while later active-learning experiments showed that magnesium can have a sharp optimum. This explains why transferring a magnesium concentration directly from one formulation to another is unsafe.

Our designs therefore keep magnesium close to the experimentally recurrent ~8 mM region instead of making a large protein-specific change.

4.2 PEP versus pyruvate/glucose

Jewett and Swartz showed that PEP-based systems accumulate substantial phosphate, whereas Cytomim using pyruvate accumulated much less phosphate. Calhoun and Swartz further demonstrated that glucose can support high protein production when pH and phosphate limitation are controlled.

These studies provide the mechanistic basis for replacing expensive phosphorylated energy substrates with inexpensive metabolic substrates.

4.3 Modern optimization confirms that interactions matter

Caschera et al. used evolutionary design of experiments to optimize CFPS and obtained conditions substantially better than the starting standard. Borkowski et al. later used active learning to explore hundreds of reaction conditions. Olsen et al. combined definitive screening designs, statistical models, active learning and domain expertise across 1,231 formulations.

This supports our methodological choice: we do not claim that a mathematical formula can calculate the true optimum concentration without experimental data. Instead, we use literature-supported concentration windows and formulate a testable candidate.

5. Target-specific candidate formulations

5.1 sfGFP — oxygen-aware low-cost NMP/ribose formulation

Hypothesis

sfGFP is a relatively robust protein and therefore provides the cleanest target for maximizing the benefit of the low-cost NMP/ribose strategy. We start from the experimentally successful region defined by RFopt and Smith et al., retain moderate HEPES, K-glutamate, Mg-glutamate, phosphate, ribose and NMPs, and add catalase as an oxygen/redox-related variable.

The formulation is intentionally not a copy of RFopt or 1793122_35. We reduce the total NMP burden to reduce cost and reaction volume, retain partial AMP substitution with adenosine, and keep catalase at the concentration used in the Smith top-performing formulation.

Candidate final concentrations

Why these changes?

• Ribose + NMPs: based directly on Olsen and Smith.
• AMP + adenosine: Smith found evidence for partial AMP replacement by adenosine.
• CMP/UMP retained: Smith found that replacing these pyrimidines was detrimental.
• Phosphate ~7.5 mM: within the strong-performing 5–10 mM region identified in the autonomous-lab experiments.
• Moderate HEPES: supports pH stability.
• Catalase: experimentally present in the Smith top-performing composition and therefore preferable to inventing an arbitrary oxygen-related reagent concentration.
• No PEP: avoids the major cost and phosphate burden of the classical system.

5.2 PETase/TfCut2 — mild oxidative folding formulation

Hypothesis

TfCut2 contains one disulfide bond. It therefore occupies an intermediate position between sfGFP and Reteplase. We propose keeping the low-cost NMP/ribose backbone while introducing a mild oxidative folding module consisting of GSSG, GSH and DsbC.

The concentrations are deliberately lower than the Reteplase oxidative system. This is a mechanistic hypothesis: one disulfide bond should not require the same oxidative burden as nine disulfide bonds.

Candidate final concentrations

Rationale

The GSSG/GSH pair provides a controlled redox environment, while DsbC provides disulfide-bond isomerization capacity. The important point is not simply that PETase receives “more reagents”: we are adding a folding module that is biologically relevant to its one disulfide bond [1].

We deliberately keep this module milder than the Reteplase condition. Olsen et al. showed that oxidative supplementation can be strongly protein-dependent and that excessive oxidative conditions can reduce sfGFP expression, even though oxidative RFopt can support disulfide-containing proteins [1]. Thus, the difference between PETase and Reteplase is a hypothesis based on folding complexity, not an experimentally established optimum.

This is a protein-specific adaptation, not a claim that 1 mM GSSG/0.5 mM GSH is an experimentally established TfCut2 optimum.

5.3 Reteplase — controlled oxidative folding formulation

Hypothesis

Reteplase contains nine disulfide bonds and therefore requires a fundamentally different folding environment. We retain the cost-saving NMP/ribose energy architecture but increase oxidative support and DsbC relative to PETase.

Candidate final concentrations

Rationale

The main change relative to PETase is the strength of the oxidative-folding module. Reteplase has nine disulfide bonds, so we hypothesize that it needs substantially more redox support than a one-disulfide protein. We therefore increase GSSG/GSH and DsbC rather than simply copying the PETase formulation.

The GSSG/GSH concentrations are anchored to the oxidative RFopt literature, and the 5 µM DsbC level follows the oxidizing disulfide-protein condition described by Olsen et al. [1]. This is the strongest protein-specific change in our design.

We also keep the ribose/NMP energy architecture rather than returning to a PEP/NTP system. The reason is economic: the benchmark shows that non-phosphorylated energy systems can maintain production at much lower reagent cost [1]. In other words, we are changing the folding environment because Reteplase needs it, but we are not giving up the cost-saving energy strategy.

A separate literature on recombinant Reteplase refolding supports the importance of a controlled GSH/GSSG environment. Importantly, that work concerns post-expression refolding, not CFPS itself, so we use it as mechanistic support rather than as a direct recipe [8].

5.4 What changes across the three formulations and why?

The central idea is therefore not to optimize every component independently. We preserve the parts of the formulation that are supported by the cost/yield benchmark and change only the components for which the protein creates a specific biochemical reason to do so.

6. How the numerical concentrations were selected

We did not calculate a theoretical optimum from first principles. Instead, each concentration follows this hierarchy:

1. Published experimentally successful concentration when available.
2. Range observed across multiple published formulations when no single optimum exists.
3. Protein-specific mechanistic adjustment when the target has a distinctive folding requirement.
4. Volume and inventory constraints imposed by the 20-µL reaction.
5. 25-nL dispensing resolution.

For example:

Mg-glutamate

Literature repeatedly places Mg-glutamate in the low-millimolar range, and independent active-learning experiments demonstrate strong sensitivity to magnesium. We therefore select approximately 8.2 mM rather than treating magnesium as a freely tunable high-range variable.

Ribose

RFopt uses 50 mM, while the Smith top-performing formulations span approximately 40–77 mM. We select 40 mM for sfGFP and PETase and 30 mM for Reteplase to leave volume for the larger oxidative-folding module.

GSSG/GSH

PETase receives a mild redox condition as a hypothesis, whereas Reteplase receives the stronger oxidized RFopt-like condition. This follows the number of structural disulfides rather than copying a single formulation across all proteins.

7. Volume calculations

For a 20-µL reaction:

[ V_i = \frac{(C_{target}-C_{base})V_{reaction}}{C_{stock}} ]

where:

• (C_{target}) = desired final concentration
• (C_{base}) = contribution from the fixed 2-µL 10X base buffer
• (V_{reaction}) = 20,000 nL
• (C_{stock}) = reagent stock concentration

Volumes are then rounded to the nearest 25 nL.

Final reaction volumes

8. R script used to calculate reaction volumes

# ============================================================
# HTBAA Week 3 — Literature-guided CFPS formulation design
# All reaction volumes are in nL.
# Final reaction volume = 20,000 nL.
# ============================================================

reaction_volume <- 20000
fixed <- list(
  lysate = 5000,
  template = 2000,
  base_buffer = 2000
)

variable_capacity <- reaction_volume - sum(unlist(fixed))

# Stock concentrations
# mM unless otherwise noted
stock <- c(
  hepes_koh = 1000,
  potassium_glutamate = 1550,
  magnesium_glutamate = 500,
  amino_acid_mix_17 = 50,
  tyrosine = 50,
  cysteine = 200,
  glucose = 1000,
  potassium_phosphate = 500,
  ribose = 666.6667,
  amp = 20,
  cmp = 20,
  gmp = 20,
  ump = 20,
  adenosine = 20,
  nicotinamide = 100,
  spermidine = 20,
  gssg = 100,
  gsh = 100,
  dsbc_ecoli = 100
)

# Contribution from the fixed 2-uL 10X base buffer
base_contribution <- c(
  hepes_koh = 30,
  potassium_glutamate = 150,
  magnesium_glutamate = 2.6,
  amino_acid_mix_17 = 1,
  tyrosine = 1,
  cysteine = 1
)

# Desired final concentrations
targets <- list(

  sfGFP = c(
    hepes_koh = 45,
    potassium_glutamate = 280,
    magnesium_glutamate = 8.2,
    amino_acid_mix_17 = 4.8,
    tyrosine = 1.2,
    cysteine = 5,
    glucose = 8.3,
    potassium_phosphate = 7.5,
    ribose = 40,
    amp = 0.5,
    cmp = 0.75,
    gmp = 1.0,
    ump = 0.75,
    adenosine = 0.4,
    nicotinamide = 4,
    spermidine = 0.6,
    catalase = 187.5
  ),

  PETase = c(
    hepes_koh = 45,
    potassium_glutamate = 280,
    magnesium_glutamate = 8.2,
    amino_acid_mix_17 = 4.8,
    tyrosine = 1.2,
    cysteine = 5,
    glucose = 8.3,
    potassium_phosphate = 7.5,
    ribose = 40,
    amp = 0.5,
    cmp = 0.75,
    gmp = 1.0,
    ump = 0.75,
    nicotinamide = 3,
    spermidine = 0.6,
    gssg = 1,
    gsh = 0.5,
    dsbc_ecoli = 2.5
  ),

  Reteplase = c(
    hepes_koh = 57.5,
    potassium_glutamate = 280,
    magnesium_glutamate = 8.2,
    amino_acid_mix_17 = 4.0,
    tyrosine = 1.0,
    cysteine = 4.0,
    glucose = 8.3,
    potassium_phosphate = 7.5,
    ribose = 30,
    amp = 0.5,
    cmp = 0.5,
    gmp = 0.75,
    ump = 0.5,
    spermidine = 0.6,
    gssg = 4,
    gsh = 1,
    dsbc_ecoli = 5
  )
)

# Catalase stock is U/mL rather than mM
stock["catalase"] <- 50000

calc_volume <- function(reagent, target) {

  base <- ifelse(
    reagent %in% names(base_contribution),
    base_contribution[reagent],
    0
  )

  # Catalase is expressed in U/mL.
  if (reagent == "catalase") {
    volume <- target / stock[reagent] * reaction_volume
  } else {
    volume <- (target - base) / stock[reagent] * reaction_volume
  }

  # All dispensing volumes must be multiples of 25 nL.
  round(volume / 25) * 25
}

make_recipe <- function(targets_for_protein) {

  volumes <- sapply(
    names(targets_for_protein),
    function(x) calc_volume(x, targets_for_protein[[x]])
  )

  total_variable <- sum(volumes)
  water <- variable_capacity - total_variable

  if (water < 0) {
    stop("Variable reagent volume exceeds the 11,000-nL capacity.")
  }

  recipe <- c(
    fixed,
    as.list(volumes),
    nuclease_free_water = water
  )

  # Check total volume
  stopifnot(sum(unlist(recipe)) == reaction_volume)

  recipe
}

recipes <- lapply(targets, make_recipe)

recipes

9. Interpretation and limitations

These formulations are candidate designs, not experimentally established optima. The literature demonstrates that CFPS is strongly nonlinear and that reagent interactions matter. Olsen et al. required 1,231 experimental formulations and combined Design of Experiments, statistical modeling, active learning and domain expertise to obtain RFopt. Caschera et al. and Borkowski et al. similarly demonstrated that experimental feedback is essential for navigating the high-dimensional formulation space.

Therefore, the appropriate claim is not:

“These are the theoretically optimal concentrations.”

Instead:

“These are literature-constrained, protein-specific candidate formulations designed to occupy experimentally supported regions of the CFPS design space while minimizing reagent cost and respecting the reaction-volume constraints.”

A true optimization stage would experimentally measure titer and functional activity across candidate formulations and then use DOE/response-surface or active-learning methods to refine the concentrations and interactions.

For Reteplase in particular, functional activity rather than total protein yield should be the relevant endpoint because the main challenge is correct disulfide pairing and folding.

10. Overall design logic

Published formulations
        ↓
Component-by-component benchmark
        ↓
Cost + biochemical function
        ↓
Identify robust concentration ranges
        ↓
Literature-supported design rules
        ↓
Protein-specific modifications
        ↓
Volume-constrained candidate formulation
        ↓
Experimental validation / future DOE

The central design principle is therefore:

We are not replacing the literature benchmark with our own ideas. We are using the benchmark as the foundation and adding experimentally motivated, protein-specific modifications on top of it.

References

1. Jewett, M. C. & Swartz, J. R. Mimicking the Escherichia coli cytoplasmic environment activates long-lived and efficient cell-free protein synthesis. Biotechnology and Bioengineering (2004). DOI: 10.1002/bit.20025.

2. Calhoun, K. A. & Swartz, J. R. Energizing cell-free protein synthesis with glucose metabolism. Biotechnology and Bioengineering (2005). DOI: 10.1002/bit.20449.

3. Caschera, F. et al. Coping with complexity: machine learning optimization of cell-free protein synthesis. Biotechnology and Bioengineering (2011). DOI: 10.1002/bit.23178.

4. Borkowski, O. et al. Large scale active-learning-guided exploration for in vitro protein production optimization. Nature Communications 11, 1872 (2020). DOI: 10.1038/s41467-020-15798-5.

5. Olsen, M. L. et al. Design-driven optimization of low-cost reagent formulations for reproducible and high-yielding cell-free gene expression. Nature Communications 17, 3478 (2026). DOI: 10.1038/s41467-026-69605-8.

6. Smith, A. A. et al. Using a GPT-5-driven autonomous lab to optimize the cost and titer of cell-free protein synthesis. bioRxiv (2026). DOI: 10.64898/2026.02.05.703998.

7. Silverman, A. D. et al. Cell-free gene expression: an expanded toolkit for synthetic biology. Nature Reviews Genetics (2020).

8. Fathi, M. et al. Efficient refolding of recombinant reteplase expressed in Escherichia coli strains using response surface methodology. (2020). PMID: 32698065.

9. Fu, C. et al. A cell-free platform for the high-efficiency synthesis and screening of PET hydrolase. Journal of Biotechnology 417, 101–107 (2026). DOI: 10.1016/j.jbiotec.2026.05.012.

Final Project

Week-02-Hw-Dna-Designs