Designing a protein that actually does something is hard. Designing a floppy, shapeless one is harder still. So it is a genuinely big deal that a team at UC Berkeley built AI-designed gene switches from scratch and found that more than 90% of them worked in living yeast cells.

The results were posted to bioRxiv on November 2, 2026. This is a preprint, which means it has not yet been peer reviewed. The data and code are open, which makes it easy for other labs to check.

What a gene switch is, in plain terms

Cells turn genes on using proteins called transcription factors. Part of each transcription factor is an activation domain — the piece that recruits the machinery that switches a gene on.

Activation domains are “intrinsically disordered.” They have no fixed shape. That makes them very difficult to design, because the usual AI protein-design playbook depends on a tight link between structure and function.

Random peptides function as activation domains only 1–2% of the time. Earlier, smarter attempts pushed that to around two-thirds. Clearing 90% is a real jump.

How the team designed them

The pipeline is refreshingly straightforward.

Step 1: Generate

The researchers trained Markov chain models on roughly 6,200 known strong plant activation domains. They used state sizes from 1 to 5, which controls how much local sequence “grammar” each model carries over.

Each model generated 100,000 sequences of 40 amino acids each — 500,000 candidates in total. Sequences that overlapped the training data too closely were thrown out.

Step 2: Filter

Every candidate was scored by four existing activation-domain predictors: Comp, PADDLE, TADA, and ADhunterLite. Three of those are neural networks.

Step 3: Test

The team picked 18,750 diverse sequences and measured them in yeast using a high-throughput sorting assay. The screen returned clean measurements for 16,770 of them.

The result that makes these AI-designed gene switches notable

92.1% of the well-measured sequences had detectable activity.

Strength mattered too, not just a yes-or-no signal:

• 72.6% were more active than Gcn4, a moderate natural activation domain
• 11% beat the team’s strongest yeast positive control
• 2.9% beat CITED2, a very strong human activation domain

The designs also taught the researchers something. The best sequences balance net negative charge against key hydrophobic residues (tryptophan, phenylalanine, tyrosine, leucine, methionine). Push either one alone and activity rises then falls. Raise both together and strength climbs steadily. Local sequence grammar adds a further boost, though composition does most of the work.

They work beyond yeast

The team picked 14 designs — ten strong, four medium — and tried them elsewhere.

In plants: all fourteen significantly outperformed a negative control in Nicotiana benthamiana leaves. One 40-residue design matched the VP-16 positive control, which is 78 residues long.

In human cells: three of the fourteen clearly activated two endogenous genes (EpCAM and ICAM1) in a CRISPR activation setup in HEK293T cells. The best, Synth #4, outperformed dCas9-VP16 but was weaker than the much larger VPR activator.

So at least three designs function across fungi, plants, and animal cells.

A faster predictor came out of it

Because agreement between predictors flagged the strongest designs, the team trained a model called MADRIGAL to mimic the consensus of six predictors at once.

It reaches a Pearson correlation of 0.92 on held-out sequences, runs far faster than the models it imitates, handles sequences of any length, and is available as a website and a Python package.

What this does not claim

Honesty check, straight from the authors’ own limitations section:

• It is a preprint, not yet peer reviewed.
• The designs only cover acidic activation domains. Other classes — glutamine-rich, proline-rich, serine-rich — remain out of reach, partly because today’s predictors are biased toward acidic ones.
• Nobody knows which coactivator proteins these designs recruit. The team suspects Mediator, based on prior work.
• Performance drops off as you move away from yeast. Only 3 of 14 tested designs clearly worked on endogenous human genes.
• A patent is pending on 14 of the sequences.
• The benefit here is mainly for the toolbox, not for patients. These are research and biotech parts, not therapies.

Why it still matters

Useful tools are how biology accelerates.

Better gene switches feed directly into CRISPR activation, Perturb-seq screens, epigenetic editing, synthetic biology circuits, and crop engineering. Researchers currently reuse a handful of the same activation domains, often borrowed from viruses and often bulky. Compact, orthogonal, tunable alternatives are genuinely useful.

And the raw sequencing data is public (GEO accession GSE352640), with analysis code on GitHub. Anyone can check the work or build on it.

The bigger picture

The headline is not only that AI-designed gene switches worked. It is that a simple recipe — generate with biased composition and local grammar, filter with several neural networks, then test at scale — turned a famously hard design problem into something close to routine.

As the authors put it, for acidic activation domains in yeast, they have achieved predictable design at scale.

That is a sentence protein designers have been working toward for forty years.

Read the preprint: Machine learning guided design of strong transcriptional activation domains