The Soup That Cracked the Genetic Code — and the AI That Tried to Make It Cheaper

Rebioin AI Blog · Biotechnology × AI

In short

1. Cell-free protein synthesis (CFPS) makes proteins in a test-tube "soup" of cell machinery, with no living cells. The same idea helped crack the genetic code in 1961.
2. In February 2026, OpenAI and Ginkgo Bioworks connected GPT-5 to a robotic cloud laboratory. Over six months, the system designed, ran and learned from tens of thousands of CFPS recipes.
3. The best recipe made a test protein at about $422 per gram of reaction-component cost, versus $698 for the previous state of the art (a 40% reduction), and the yield rose by 27% in tubes.
4. It is a promising preprint, not a finished revolution. It was shown on one protein in one system, and the authors work for the companies that funded it.

1961: a soup, a synthetic RNA and the first word of the genetic code

In 1961, Marshall Nirenberg and Heinrich Matthaei did something that sounds almost too simple. They broke open E. coli cells and kept the cellular "juice": ribosomes, enzymes and the other machinery needed to build proteins. Then they added a synthetic RNA made of nothing but uracil (poly-U). The soup produced a protein chain made entirely of one amino acid, phenylalanine [1][2]. That single result showed that UUU codes for phenylalanine, and it proved that messenger RNA directs protein synthesis. It opened the way to decoding the rest of the genetic code, and Nirenberg shared the 1968 Nobel Prize in Physiology or Medicine with Robert Holley and Har Gobind Khorana [2][3]. No living cell was needed, only a carefully assembled mixture in a tube. That is the idea behind cell-free protein synthesis, and 65 years later it is the stage for one of the most interesting experiments in AI-driven biology.

What is cell-free protein synthesis (CFPS)?

Normally, to make a protein you put a gene into living cells (bacteria, yeast or mammalian cells), keep them alive and fed, and then purify your protein from them. CFPS skips the living cell. You take a lysate (the cell's protein-making machinery, released by breaking the cells open) and add:

1. a DNA template that carries the gene,
2. amino acids, the building blocks,
3. an energy source and cofactors, to power translation,
4. salts and buffers, to keep the chemistry stable.

Because there is no cell to keep alive, the reaction is open: you can tweak any ingredient and see what happens. Researchers use CFPS for rapid prototyping of proteins, antibodies, enzymes and peptides, for proteins carrying unusual amino acids, and even in the commercial production of an antibody-drug conjugate (luveltamab tazevibulin) [4].

The catch is that a CFPS reaction is like a soup with dozens of ingredients that interact with each other. Change the phosphate and the pH shifts. Change the pH and the magnesium behaves differently. The number of possible recipes is enormous, and every test costs reagents, lysate and DNA. So optimizing CFPS is slow and expensive for humans doing it by hand.

Two numbers that matter: titer and specific cost

There are two figures of merit to keep in mind [4]:

1. Titer is how much protein you get, in grams per litre of reaction (g/L). Higher is better.
2. Specific cost is the total reaction cost divided by the grams of protein made ($/g). Lower is better.

A recipe that is cheap but gives almost no protein is useless, so the real goal is to push both in the right direction at once.

The benchmark to beat was a 2025 preprint from Olsen and colleagues (reported as coming from Northwestern University researchers). They designed low-cost reagent formulations for CFPS and produced the standard test protein, superfolder GFP (sfGFP), at about $698 per gram and a titer of about 2.4 g/L [4][5][6]. After their reagent savings, the cell lysate and DNA template made up more than 90% of the remaining cost, so cutting cost further meant improving the whole system, not just the cheap chemicals [4].

The experiment: GPT-5 plus a robotic lab

In February 2026, researchers from OpenAI and Ginkgo Bioworks posted a preprint describing a closed-loop "autonomous lab" [4][7]. The idea is simple to describe:

1. Design: GPT-5 proposes a batch of CFPS reaction recipes.
2. Check: A validation layer (a Python tool called Pydantic) acts like a strict form. Every proposed experiment must fill it in correctly, with a valid plate layout, controls, replicates and allowed reagent volumes, before any robot touches it.
3. Run: Ginkgo's cloud lab in Boston, built from modular robotic carts and orchestrated by its Catalyst software, mixes the reactions in 384-well plates and measures the protein.
4. Learn: The results go back to GPT-5, which analyzes them, writes notebook-style notes on what it saw, and designs the next round.

Humans mostly prepared and loaded reagents and consumables, and supervised [4]. The work ran in six steps over six months. In the first step, GPT-5 designed reactions "zero-shot", meaning with no example data, using only what it had learned in training. From step 3 onward, the team gave it internet access, a computer for data analysis, richer experimental metadata and the Olsen preprint. Across the project the lab ran more than 580 plates, about 36,000 reaction compositions and almost 150,000 data points [4][8][9].

What happened?

The best recipes beat the benchmark on both cost and yield [4]:

The winning recipe also used less lysate (25% instead of 30% of the reaction volume) and less DNA template (5 nM instead of about 8.7 nM), which are exactly the expensive parts [4]. In the 384-well plates used for high-throughput screening, a well-balanced recipe reached 0.72 g/L versus an estimated 0.25 g/L for the benchmark. Titers are generally higher in tubes than in plates, which the authors attribute to differences in oxygen supply [4].

What did the AI "discover"?

This part is a gift for biochemistry students, because the lessons are classic principles. According to the study, GPT-5 pointed to these patterns in the data. The authors describe them as plausible but still needing further experimental validation [4]:

1. Buffering matters a lot. In one screening round, reactions with extra HEPES buffer had a median titer of about 0.18 g/L, compared with about 0.06 g/L without it. A cheap ingredient that holds the pH steady protects the yield.
2. Phosphate has a sweet spot. Too little or too much (above about 25 mM) crashed the titer, while 5–10 mM worked well.
3. NMPs beat NTPs on cost. Nucleoside triphosphates (NTPs, the "ready-to-use" building blocks and energy carriers) are expensive. Nucleoside monophosphates (NMPs) are cheaper, and the lysate's own metabolism can convert them. Partly replacing some NMPs with the base or nucleoside plus ribose worked for the purines (guanine and adenine) but failed for the pyrimidines (C and U).
4. Spermidine helps. Reactions with this polyamine had roughly double the median titer in one round (about 0.25 versus 0.13 g/L). GPT-5 suggested that polyamines stabilize nucleic acids and ribosome function, which is a hypothesis rather than a proven mechanism.

One detail stood out. Before it was given the Olsen preprint, GPT-5 had already suggested reagents such as NMPs, potassium phosphate and ribose, which were central to the benchmark's cost savings [4]. It also tried nicotinamide on its own, and nicotinamide ended up in all of the top recipes.

Where the AI stumbled

As a reality check, not everything went smoothly. Of the 480 plates GPT-5 designed, two had fundamental flaws (under 1%) [4]. In one, the model tried to override a fixed 2 µL buffer volume so that it could add more reagents, pushing the reaction above the 20 µL limit. In another, a small unit-conversion bug (nanolitres to microlitres) left every well with only glucose and ribose. With no nucleotide building blocks, no protein formed.

The validation layer caught most mistakes before the robots ran them, and the team patched it after each error. The lesson is that LLMs make mistakes in the lab just as they do elsewhere, so guardrails and human review are part of the system, not an afterthought.

A fair reading: what this does and does not show

It is easy to over-read a headline like "AI cuts protein cost by 40%." Here is what is solid and what is not:

1. It is a preprint. As far as I could find, it has not been peer reviewed, and all authors are current or former employees of Ginkgo Bioworks and OpenAI, who funded the work [4].
2. The credit is shared. The team also improved the DNA template and the lysate preparation after step 2, and gave GPT-5 tools and the benchmark paper from step 3, and the authors note that this contributed to a jump in performance [4].
3. "Cost" means reaction components. The figure covers reagents, lysate and DNA at the same unit prices used by the benchmark study. It is not a market price or a full manufacturing cost, because it excludes labor, equipment and purification [4].
4. One protein, one system. The results come from sfGFP in an E. coli CFPS system, and the sources stress that generalization still needs to be shown [4][7][9]. When one of the best recipes was tested on twelve other proteins, six were visible by SDS-PAGE and nine were detectable by a more sensitive tag-based assay, so recipes will likely need tuning for each protein [4].
5. It is already reaching the market. Ginkgo announced that it is selling the AI-improved reaction mix, a sign of commercial interest [8][9].

Why this matters for the next generation of scientists

The authors frame research as having two bottlenecks: the hands-on work of experiments and the thinking needed to plan them and interpret the data [4]. Robotic labs ease the first, and reasoning models may help with the second. Putting them together creates a loop that can explore a huge recipe space faster than a person could. The authors predict that this style of work will become common across biology and chemistry, though that remains a prediction.

For a student, the takeaway is not "AI replaces the biochemist." Every finding above, from buffers to polyamines, still had to be judged by people who understand the chemistry. A 2025 review in Trends in Biotechnology on LLMs in biomanufacturing makes a similar point. It covers how LLMs can extract knowledge from literature, support knowledge graphs and retrieval-augmented generation, and help enable self-driving labs, while emphasizing trustworthy benchmarks, biosecurity safeguards and collaboration between AI scientists, biologists and engineers [10].

The skills that become more valuable are asking good scientific questions, designing sound controls, spotting artifacts and knowing when a result is too good to be true.

Think about it

1. Why might swapping NTPs for NMPs lower the cost, and what does it require of the lysate's metabolism?
2. Why might protein yield be lower in a 384-well plate than in a 2-mL tube?
3. The recipe was optimized for sfGFP. What experiments would you run to test whether it works for a therapeutic enzyme or an antibody fragment?

Conclusion: the same soup, a new kind of cook

In 1961, a tube of ground-up bacteria taught us how genes become proteins. In 2026, a similar soup became the testing ground for an AI that plans experiments, a robot lab that runs them, and a loop that learns from every result. The headline numbers are promising, but the more lasting story is how scientists are learning to pair language models with real experiments, with guardrails, honest caveats and human judgment.

At Rebioin AI, we believe understanding the connection between artificial intelligence and life sciences is an important step towards the next generation of biotechnology research.

Quick glossary

1. CFPS (cell-free protein synthesis): making proteins using cell extracts in a tube instead of living cells.
2. Lysate: the cell contents (ribosomes, enzymes, etc.) left after breaking cells open.
3. Titer: protein yield, in grams per litre of reaction.
4. Specific cost: reaction cost per gram of protein produced ($/g).
5. sfGFP: superfolder green fluorescent protein, a standard test protein that is easy to measure by fluorescence.
6. NTP / NMP: nucleoside tri- or monophosphates, the RNA building blocks (and energy carriers, in the case of NTPs).
7. Autonomous lab: a system in which software plans experiments and robots run them in a closed loop.
8. Preprint: a research paper shared before peer review.

References

1. Nirenberg MW, Matthaei JH. The dependence of cell-free protein synthesis in E. coli upon naturally occurring or synthetic polyribonucleotides. Proc Natl Acad Sci USA 1961;47(10):1588–1602. doi:10.1073/pnas.47.10.1588
2. U.S. National Library of Medicine, Profiles in Science. Synthetic RNA and the Poly-U Experiments, 1959–1962. https://profiles.nlm.nih.gov/spotlight/jj/feature/syntheticrna
3. Nirenberg M. Oral history, Chapter 3: RNA and Poly-U (NIH History Office archive). https://webarchive.library.unt.edu/web/20130214163154/http://history.nih.gov/archives/downloads/Nirenberg%20oral%20history%20Chap%203-%20%20RNA%20and%20PolyU.pdf
4. Smith AA, Wong EL, Donovan RC, et al. Using a GPT-5-driven autonomous lab to optimize the cost and titer of cell-free protein synthesis. Preprint, OpenAI and Ginkgo Bioworks, 2026. https://cdn.openai.com/pdf/5a12a3bc-96b7-4e07-9386-db6ee5bb2ed9/using-a-gpt-5-driven-autonomous-lab-to-optimize-the-cost-and-titer-of-cell-free-protein-synthesis.pdf
5. Olsen ML, et al. Design-driven optimization of low-cost reagent formulations for reproducible and high-yielding cell-free gene expression. bioRxiv 2025. doi:10.1101/2025.08.01.668204
6. The Decoder. OpenAI and Ginkgo Bioworks build an autonomous lab where GPT-5 calls the shots. https://the-decoder.com/openai-and-ginkgo-bioworks-build-an-autonomous-lab-where-gpt-5-calls-the-shots/
7. OpenAI. GPT-5 lowers the cost of cell-free protein synthesis. 5 February 2026. https://openai.com/index/gpt-5-lowers-protein-synthesis-cost
8. Ginkgo Bioworks. Autonomous laboratory driven by OpenAI's GPT-5 achieves 40% improvement over state-of-the-art scientific benchmark (press release, 5 February 2026). https://www.barchart.com/story/news/54531/ginkgo-bioworks-autonomous-laboratory-driven-by-openai-s-gpt-5-achieves-40-improvement-over-state-of-the-art-scientific-benchmark
9. Buntz B. OpenAI's GPT-5 autonomously ran 36,000 protein synthesis experiments in Ginkgo Bioworks' cloud lab. R&D World, 5 February 2026. https://www.rdworldonline.com/openais-gpt-5-autonomously-ran-36000-protein-synthesis-experiments/
10. Li W, Mao Z, Xiao Z, Liao X, et al. Large language model for knowledge synthesis and AI-enhanced biomanufacturing. Trends Biotechnol 2025. doi:10.1016/j.tibtech.2025.02.008