IGI researchers discover an ancient progenitor of CRISPR using a combination of AI tools and human ingenuity to solve a genomic mystery.
When scientists first started to decode the language of the genome, the intuitive simplicity made the discovery feel especially profound. The language of our cells works much like our own languages. The code is read consecutively from beginning to end, every three characters like a word in a sentence, every gene a complete story. Biology, however, is never content to be quite so neat.
In two connected papers published today in the journal Science (Paper 1 | Paper 2), researchers from the lab of Jennifer Doudna at the Innovative Genomics Institute (IGI) at UC Berkeley describe the surprising discovery of an ancient progenitor of CRISPR systems that had been overlooked because it uses a never-before-seen coding system to find and bind to DNA.
In nature, CRISPR is an adaptive immune system that bacteria and archaea use to fight viruses. CRISPR systems collect snippets of viral DNA, and they use matching pieces of RNA like mugshots to recognize invaders. The newly discovered system, named “VIPR” (Viral Interference Programmable Repeat), likely predates CRISPR and was instead found in viruses. The new studies rewrite CRISPR’s origin story with a surprising twist: bacteria appear to have stolen a weapon made by viruses to fight other viruses, and they turned it back on their enemy.
The case of the confusing code
As often happens in science, the discovery of VIPR started with an observation that didn’t make sense.
Since the discovery of CRISPR-Cas9’s ability to be harnessed as a gene-editing tool, scientists have searched for other related molecules in nature. CRISPR-Cas9 is what is known as a “Class 2” CRISPR system. In evolutionary terms, these are the latest models, composed of a single large protein. Because of the utility for life science research and gene editing for health and other applications, Class 2 systems have been deeply studied in recent years.
Peter Yoon and Kenneth Loi in the Doudna lab at the IGI were instead digging into the origins of Class 1 CRISPR systems, which are much more common in nature but less studied. Class 1 systems are composed of multiple proteins that come together to form a complex, and are also significantly older than Class 2 systems — in fact, it is thought that the last universal common ancestor of all cellular life on Earth, aka “LUCA,” had a Class 1 CRISPR system.
In an earlier paper, the team developed a process using AI tools to improve on older methods for finding related proteins. With this AI-assisted process, they could uncover molecules with similar structures and functions that were missed by earlier methods that relied on comparing the underlying sequences.
The further you go back in evolutionary history, the more challenging it becomes to find related proteins. For organisms that diverged relatively recently, DNA sequences will be relatively similar. As divergence times get farther and farther back on the timeline, the sequences have had more time to evolve away from each other. The sequences can look unrecognizable while key parts of the protein structure remain comparable.
“If you want to find something truly ancient, you need to look for something with a particular shape, not a particular sequence,” says Doudna.
Using their AI-assisted approach to search through roughly 2.3 million structures for proteins that might be related to early CRISPR systems, Yoon and Loi found a few hundred candidates. One particular protein stood out: it had the same shape as CRISPR proteins but was never found next to CRISPR RNAs. Sequencing this new system revealed that the protein was paired with a mysterious RNA. CRISPR RNA contains an exact match of a section of virus DNA. This new RNA didn’t match anything they could identify.
“We were scratching our heads. It didn’t look like any RNA we had ever seen before. That’s where the project was stuck for months,” says Doudna. “I told Peter and Kenneth, this is either something very interesting or very boring.”
It turned out to be very interesting; it just needed the right combination of human ingenuity and cutting-edge AI tools working together to crack the code.
The mystery of the mismatched RNA
To gain insights into the structure and function of VIPR, the team turned to colleagues in the Doudna and Brohawn labs, Terry Zhang and Trevor Docter.
“Part of the mystery was that we knew that VIPRs look similar to some CRISPR systems, but we didn’t know their biological function. We also didn’t know their biochemical properties, so that made them very hard to study,” says Zhang.
Yoon, Zhang and Docter initially imaged a sample with a low-resolution electron microscope, and confirmed that the VIPR system was assembled from multiple proteins. When they treated it with an enzyme that removes RNA, the complex disappeared, showing that it was likely assembled along its RNA partner.
“After a lot of optimization, we were able to finally get a high-resolution structure. And then everything started coming together quickly. In the first set of structures, we could clearly resolve each protein subunit and the VIPR RNA, which really opened the door for us,” says Docter.
The picture was coming into focus, but the team was stuck on the mysterious VIPR RNA that matched no known sequence. What could it possibly do? And why did its sequence make no sense?
This time, with a different AI language model, the team spotted a unique pattern.
“I reasoned that if a genomic language model has been trained on enough of these sequences, it should have an idea of how to generate one on its own,” says Loi.
Loi’s intuition turned out to be correct: the AI model picked up patterns in the training sequences and was able to generate sequences that were variations on what the team had found in nature. A language model you might use to write an email is good at predicting what word should follow another in a sentence based on probabilities. Similarly, the genomic language model scores and predicts which base should come next in a string of code.
In the VIPR RNA, three bases in a row were easy for the model to predict: they were almost always GGT with very high probability. This would repeat consistently in the sequence like beads on a string, but in between these repeats there were two bases that were almost impossible to predict.
“I looked at this for a very long time. I never picked it up by eye. It was only these language models that were able to pull out a pattern and make it visible to us,” says Yoon.
Instead of a standard consecutive string of bases like CRISPR RNA, the VIPR RNA appeared to be using what is called a “skip cipher” in cryptography. In between the repeating GGTs was a hidden code. Skip 3, read 2, skip 3, read 2, and repeat.
The team pulled out this hidden code in a consecutive string to see if it matched any known sequence. No match. If the VIPR RNA has two letters of code and three gaps, perhaps that same pattern would be found in the target sequence? Again no match. They tried a gap of two bases in the target sequence, and then a single-base gap, and finally they found matches in viral DNA.
“This was really bizarre,” says Yoon. “We were able to observe that if you follow this one predictive pattern from the structure, that we start to find clear matches. But if we follow any other pattern — just nothing.”
As the team looked more closely at how the VIPR RNA matched with virus genes, their opinion of this system quickly changed. What looked bizarre at first glance was actually an ingenious solution to a problem: virus DNA mutates rapidly, so if you want to attack a virus’ genome, your tool needs to be adaptable to change.
Finding the signal, ignoring the noise
At some point you might have come across a viral message (in the internet meaning of the term) that looked something like this:
” Aoccdrnig to rscheearch at Cmabrigde Uinervtisy, it deosn’t mttaer in waht oredr the ltteers in a wrod are, the olny iprmoetnt tihng is taht the frist and lsat ltteer be in the rghit pclae.”
This strange-looking sentence is a demonstration of the phenomenon humorously called “typoglycemia,” where most people can understand a sentence made of jumbled words as long as the first and last letters of each word remain unchanged. It might slow us down a little, but our brains can deal with some noise and still recognize the words and understand the sentence.
What the VIPR RNA is doing is comparable, but it follows a different pattern. Instead of relying on the first and last letters of a word to remain stable while the other letters change, VIPR looks at the genetic code in sets of three letters, relying on the first two to be stable, while the third can change in any way it likes:
Using VIPR’s rules, 
would be the same as 
.
With CRISPR’s rules — and the standard way genetic strands are known to pair — these two sentences would have to be a perfect, letter-for-letter match.
To say that VIPR’s gapped code is not the way scientists have come to expect the genetic code to work is an understatement. There are sometimes large gaps in coding regions of DNA, but the code itself is supposed to be consecutive, read letter-by-letter, just like the letters in this sentence.
For a quick refresher on basic genetics, the genetic code in DNA is made of four nucleotide bases — A, C, G, and T, for short. Within the coding regions of the genome, sets of three of these bases in a row make up a codon, each one corresponding to an amino acid building block for proteins. For example, TCA translates to the amino acid serine, but there is flexibility built into the system: TCC, TCG, and TCT also translate to serine. The third letter is sometimes called the “wobble position” because in many cases changing it has no impact.
The most likely place to find a mutation in a gene is in these wobble positions — and those are exactly what the VIPR’s unique system ignores. If the targeted virus mutates, the VIPR RNA will still find and pair to the more stable parts of the code.
“No known system does this, and this provides a very clear evolutionary rationale,” says Yoon. “This is a mechanism that intentionally masks out the parts of the code that are likely to change. It’s a de-noising algorithm. It makes it very difficult for the targeted virus to evolve out of danger.”
The VIPR attacks
With the remarkable gapped coding system cracked, the research team could move on to figuring out how VIPR worked in nature, and what they found were signs of an ancient war between viruses.
“The viruses that are encoding these VIPR systems are targeting related viruses, often viruses that infect the same host, so it’s part of viral competition,” says Yoon.
Once they understood how to search for them, they started finding VIPR systems everywhere.
“In metagenomic data, where researchers sequence environmental samples from soils and other places, we could find hundreds of thousands easily,” says Yoon. “They are widespread enough that we found them in bacteriophage viruses that were literally sitting in our fridge.”
Nature has found various ways to deploy and repurpose VIPR systems. In some cases, the viruses being targeted by VIPR carry their own VIPR RNAs that serve to lead the other VIPR system astray. In other cases, VIPR systems target other VIPR systems, in a never-ending VIPR-on-VIPR war.
Unlike CRISPR, which pushes the two strands of DNA apart to find and cut the targeted DNA, VIPR takes a different approach that inspired its name: it twists itself around the DNA double helix like a snake.
“CRISPR systems like Cas9 have to overcome the challenge of double-stranded DNA, which is very stable and well-protected. They use their guide RNA to invade the DNA duplex and force it to open up so that the pairing can happen, and they use a continuous string of bases that creates a stable pairing,” says Zhang.
VIPR’s RNA, on the other hand, is full of gaps, and it wouldn’t be stable in the same scenario.
“From our observations, the DNA strand is not being replaced by the RNA strand, but rather the VIPR filament is destabilizing the DNA in a way that the DNA gives up one of its pairing strands to interact with the VIPR RNA guide and forms a triplex,” says Zhang.
This serpentine triplex structure is key to VIPR’s attack. It doesn’t use a protein to cut the DNA like CRISPR, instead it wraps around the targeted DNA, effectively turning off the gene.
“One of the things we demonstrate in these papers is that we can reprogram VIPR to silence a gene by binding upstream to the gene’s promoter region,” says Loi.
CRISPR has been used to silence genes as well, but VIPR has several advantages as a potential tool. First, it’s incredibly small — the smallest RNA-guided system that has been found. This makes it significantly easier to deliver into cells, often a challenge for the large CRISPR enzymes. VIPR also has no limitations on where it can target in a genome. CRISPR-Cas9 requires a short sequence called a “PAM” near where it binds and cuts DNA; VIPR can be reprogrammed to target anywhere in a genome.
“There’s no system that wraps around DNA like this. We suspect that this unusual geometry should allow for different kinds of applications people couldn’t do before,” says Yoon.
Intriguingly, the team found cases where fully functional VIPR systems were transferred to bacteria. The bacteria picked up a weapon used by viruses and turned it back on viruses, and this act eventually led to the evolution of Class 1 CRISPR systems before LUCA, the last universal common ancestor of all cellular life.
“We’ve been calling it the ultimate anime betrayal. These viruses are fighting each other, and the host just goes, ‘Oh, boop, let me take that gun from you guys’,” says Loi. “It’s crazy to think that this one act could be the origins of adaptive immune systems broadly for life on planet Earth.”
Projects like this keep Doudna endlessly curious about the next discovery down the road.
“This work is a great reminder of how complex biology is, and how little we still know,” says Doudna. “Fundamental concepts are still out there to be discovered. I think that’s wonderful.”
Read more:
- Yoon PH, Loi K, Zhang ZT, Docter TA, Lopez SC, Lengeberg CJ, Ur-Rehman MM, Vohra K, Zhow Z, Esain-Garcia I, Trinidad MI, Shi H, Boger R, Wang PY, Adler BA, Brohawn SG, and Doudna JA. 2026. A Noncontiguous Code for RNA-Guided DNA Recognition at the Origin of CRISPR-Cas. Science.
- Yoon PH, Docter TA, Zhang ZT, Loi K, Lopez SC, Valentin-Alvarado LE, Tuck O, Brohawn SG, and Doudna JA. VIPR RNA-guided DNA recognition by noncontiguous geometric triplex formation. 2026. Science.
- Science Perspective: A viral origin for RNA-guided immunity
- Media contact: Andy Murdock andymurdock@berkeley.edu
- Read this story on the IGI website.