AI protein design has moved from predicting existing biology towards building parts that did not previously exist in nature. In a peer-reviewed Nature Communications paper published on 22 September 2026, researchers used generative protein models to create new thiolation domains, inserted them into enzyme systems and tested hundreds of engineered variants in living cells.
The team produced 76 de novo domains and evaluated 578 recombinant non-ribosomal peptide synthetase variants. Several of the AI-designed parts remained functional when substituted into different enzyme architectures, and some variants increased product yield by up to about threefold compared with versions carrying the native domain. That is a laboratory result, not a claim that AI can design arbitrary organisms or biological systems on demand.
What AI protein design produced in the lab
The researchers used generative models including ESM3, ProteinMPNN and EvoDiff to design protein sequences for thiolation domains. These are small components inside large enzyme assembly lines known as non-ribosomal peptide synthetases, or NRPSs. Such enzymes help microbes manufacture many chemically complex molecules.
A generated sequence is only a proposal. The important step was making the proteins and placing them into working biological systems. The team built hundreds of engineered NRPS variants and tested whether the new domains could support product formation.
Some did. The paper reports functional de novo domains across several architectures, showing that the generated parts were not confined to one carefully selected enzyme. A representative design also showed higher thermal stability than its natural comparison, with a melting temperature 12 degrees Celsius higher.
That combination of generation and physical validation is what makes the result notable. AI can produce enormous numbers of plausible protein sequences. Biology decides whether any of them actually fold, interact and perform the intended job.
Thiolation domains are small parts with a difficult job
NRPS enzymes behave loosely like molecular assembly lines. Different domains perform different operations as a product is constructed. Thiolation domains carry intermediate molecules between catalytic steps, so they have to interact correctly with neighbouring components.
That makes them attractive targets for engineering and difficult ones. A replacement can look structurally plausible yet fail because it does not communicate properly with the rest of the enzyme. Natural evolution has already tuned these interactions over long periods.
Generative design offers another route. Instead of finding a natural domain that is close enough and modifying it, a model can propose a sequence built for the desired structural and functional constraints. The September paper tests whether that approach can survive contact with a real biosynthetic system.
LiveAIWire has covered AI-designed bacteriophages and biological sequences, another example of generated designs moving into physical experiments. The protein work is distinct, but the common lesson is the same: a computational proposal becomes scientifically meaningful only after it is built and tested.
The threefold result needs context
“Tripled output” is the most dramatic number in the paper, but it should not be read as a universal performance gain from AI-designed proteins. The result applies to particular engineered variants and products in the experimental systems studied.
Different NRPS architectures responded differently to substitution. Some generated domains worked better than others, and the study does not imply that replacing any natural domain with an AI-designed one will increase yield. Protein engineering remains highly dependent on context.
The comparison also concerns product formation within engineered biological systems, not factory-scale manufacturing. A higher yield in a laboratory experiment can be promising without guaranteeing that the same advantage survives scale-up, purification, process optimisation and commercial production.
Keeping those boundaries clear makes the result more useful. The achievement is not a magical threefold button. It is evidence that de novo domains can sometimes function inside complex enzymes well enough to improve a measurable output.
From generated sequence to working biology
The researchers also provide an official code and workflow repository for the project, including generation workflows using ESM3, EvoDiff and ProteinMPNN. That technical record helps show what sits behind the headline result: a reproducible design pipeline rather than a single chatbot prompt.
The pipeline illustrates how AI-assisted biology differs from ordinary content generation. Producing more candidates is only the first stage. Researchers still have to choose which sequences to synthesise, build DNA constructs, express proteins, assemble enzymes and measure products.
Those steps are expensive enough that model quality matters economically. If generative methods can raise the proportion of candidates that work when built, laboratories can explore a larger design space without synthesising every possibility.
Protein design could expand the chemical search space
NRPS systems are interesting because they produce structurally diverse molecules. Re-engineering their component parts can potentially change what those molecular assembly lines make or how efficiently they make it.
That connects protein design with the wider promise of AI-assisted drug and materials discovery. LiveAIWire’s examination of AI in the drug-discovery timeline emphasises that computational acceleration does not remove the later experimental stages. De novo protein design follows the same logic. Better candidates can improve the front of the funnel without eliminating the rest of the pipeline.
The most important long-term question is therefore not how many protein sequences a model can generate. It is how often those sequences produce useful behaviour when inserted into real biological machinery, and whether researchers can predict that success before paying to build them.
This is designed biology, not autonomous biology
Headlines about AI creating biological parts can suggest a system operating independently of scientists. The paper describes something much more controlled. Researchers selected the target, chose and applied models, built the designs, created the engineered enzyme variants and performed the experiments.
The publication also includes researchers with academic and industry affiliations, so its claims should be judged through the reported methods and results rather than treated as a neutral product benchmark. Peer review increases confidence in the scientific record, but it does not remove the need for replication across other systems.
For now, the strongest conclusion is specific and significant. Generative protein models produced new thiolation domains that were physically built, inserted into complex enzyme systems and shown to function. In some cases they did more than function: they increased the amount of product the engineered system made.
Nature did not supply those exact sequences. AI proposed them, and the laboratory decided whether they deserved to exist.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.