A Memory Of Every Virus

Each spacer is a captured piece of something that tried to kill an ancestor. The name is a description of the architecture: Clustered Regularly Interspaced Short Palindromic Repeats. In the genome there is an array where short repeated sequences alternate with spacers, and each spacer is a fragment of a virus or plasmid that once attacked the cell line. Beside the array sit the cas genes, which encode the proteins that do the capturing and the cutting. Because the array is part of the chromosome, it is inherited. A bacterium is born carrying its lineage's list of old enemies, and it uses that list to recognise them again. This is an adaptive immune system, with a memory written down, in an organism that has no room for anything it does not need.

A Spanish microbiologist found the repeats in 1993 and worked out what they were for in 2005. Francisco Mojica, at the University of Alicante, first noticed the unusual repeated sequences in Haloferax mediterranei, an archaeon from a salt marsh. Nobody knew what they were. Over the following decade he kept at them, and by 2005 he had recognised the thing that makes the whole story: the spacers between the repeats matched foreign viral and plasmid DNA. If the sequences in the ledger were copies of viruses, the ledger was an immune record. Alexander Bolotin and Gilles Vergnaud reached the same recognition independently the same year. It is worth pausing on the shape of that: the central insight came from a curiosity-driven study of a microbe in a Spanish salt marsh, and it took twelve years to interpret.

The experimental proof came out of an industrial problem with cheese cultures. Rodolphe Barrangou and Philippe Horvath were working at Danisco, a food-ingredients company, on a thoroughly commercial difficulty: bacteriophages destroy the starter cultures used to make yoghurt and cheese, and that costs money. In 2007 they published in Science the demonstration that CRISPR provides acquired resistance against phages in Streptococcus thermophilus. They challenged the bacteria with phage, watched the survivors acquire new spacers matching that phage, and showed the survivors were now resistant. Add a spacer, gain immunity; remove it, lose immunity. The first proof that this was an immune system at all came from a dairy laboratory.

The steps are short and each one is somebody's whole career. 2010: the Moineau group shows that Cas9 cuts the target DNA within the protospacer, so Cas9 is the blade. 2011: Charpentier's group identifies the tracrRNA, a second RNA needed to mature the guide, which is the missing component that made the system reproducible outside a bacterium. 2012: Jinek and colleagues, in the Doudna and Charpentier collaboration, fuse the two RNAs into a single synthetic guide and show that Cas9 can then be programmed to cut purified DNA at any chosen target. 2013: Feng Zhang, George Church and Doudna's groups independently get it working inside mammalian cells. Nineteen years from an unexplained repeat in a salt-marsh archaeon to editing a human genome.

Cas9 does not read the genome. It looks for a three-letter motif and only then checks the address. SpCas9 is a protein of 1,368 amino acids with two nuclease domains that cut one strand each: RuvC takes the non-target strand and HNH takes the target strand, and between them they make a blunt double-strand break. The search works like this. The enzyme is inert until it binds its guide RNA, at which point it undergoes a large conformational change that opens a channel for scanning DNA. It then hunts for a PAM, a protospacer adjacent motif, which for SpCas9 is the three letters 5'-NGG-3'. Finding one, the PAM-interacting domain melts the helix locally, and the guide RNA tries to base-pair with the exposed strand. Only on full twenty-nucleotide complementarity do the nucleases fire, cutting three base pairs upstream of the PAM.

The twenty letters of the guide are not equally strict, and the slack end is the whole problem. Specificity is dominated by roughly twelve seed nucleotides nearest the PAM. A mismatch in that region strongly reduces cutting. A mismatch at the far end may be tolerated. So the guide is really a short strict password followed by a longer, laxer one, and the laxity at the tail is exactly where off-target cuts come from. This is not a flaw someone introduced. It is a property of a system that evolved to catch viruses which are themselves mutating, where a little tolerance for mismatch is an advantage. What is a virtue in a bacterium under attack is a liability in a tool meant to hit one site and no other.

Everything called gene editing is done by the cell's own repair machinery. Cas9 breaks the DNA and stops. What happens next is the cell's response to a double-strand break, and there are two routes. Non-homologous end joining simply shoves the ends back together; it is error-prone and typically leaves small insertions or deletions, which disrupt the gene. That is how a gene gets switched off. Homology-directed repair copies from a template, so if a donor sequence is supplied the cell will write that in, which is how a sequence gets corrected or inserted. Which route the cell takes depends on the cell type and where it is in its cycle, and that is why the same edit works beautifully in one tissue and poorly in another.

If the break is the dangerous part, remove the break. Base editors do exactly that. Komor and colleagues in David Liu's laboratory fused a catalytically dead Cas9, which still finds and holds the target but cannot cut, to a cytidine deaminase, converting C to T chemically, in place, in 2016. Gaudelli and colleagues in the same laboratory then evolved a modified transfer-RNA adenosine deaminase to convert A to G, in 2017. Neither makes a double-strand break at all. The enzyme becomes a delivery vehicle: it stops being scissors and starts being a way to carry a small chemical modification to a chosen letter.

Prime editing lets the guide RNA specify not just where, but what. Anzalone and colleagues, again in Liu's laboratory, published it in 2019. A Cas9 nickase, which cuts only one strand, is fused to a reverse transcriptase, and the guide is extended into a pegRNA that carries both the address and the replacement text. The reverse transcriptase writes the new sequence directly into the nicked strand. It can make any of the twelve possible point mutations, insertions up to about 44 base pairs and deletions up to about 80, with no double-strand break and no donor template. The name is precise: search, and replace.

The amount of machinery built purely to detect mistakes is itself the answer about precision. Our file names four separate methods developed to map off-target cleavage across a whole genome: GUIDE-seq, CIRCLE-seq, Digenome-seq and DISCOVER-seq. Alongside them sit engineered high-fidelity variants, eSpCas9, SpCas9-HF1 and HiFi Cas9, which reduce off-target activity by altering the protein-DNA interface. For any therapeutic use, edited cells get whole-genome sequenced to check what else happened. Nobody builds four detection platforms for a problem that does not exist.

It is not molecular scissors and it is not perfect. Our file marks the early media framing debunked, and the list of what has actually been documented at cut sites is specific: off-target effects, mosaicism, large-scale chromosomal rearrangements, and chromothripsis, which is the shattering of a chromosome and its faulty reassembly. The file's own verdict is the one to keep: CRISPR is powerful but not infallible. And note where that refusal comes from. Every item on that list is a finding published by the same field that built the tool, which is roughly the opposite of a technology being oversold.

A gene drive is designed to spread through a wild population faster than Mendel allows. Ordinarily an engineered allele passes to half the offspring. A CRISPR gene drive copies itself onto the partner chromosome, so nearly all the offspring inherit it, and it can sweep through a population. Kyrou and colleagues showed in 2018 that a drive targeting the gene doublesex produced complete population suppression in caged Anopheles gambiae, the mosquito that carries malaria. Andrea Crisanti and Austin Burt at Imperial College London led the work. Our file states the position without decoration, and it has not changed in the file since: field release has not occurred, because of ecological and ethical concerns.

The hard part is no longer cutting. It is getting there. Our file names the constraint precisely: adeno-associated virus vectors, the workhorse of gene delivery, have a packaging limit of about 4.7 kilobases, and SpCas9 plus its guide RNA does not fit. Lipid nanoparticles do work, but currently deliver efficiently mainly to the liver. So a tool that can in principle address any sequence in any genome is limited in practice by a cargo-space problem and an organ preference. That constraint, incidentally, is why smaller Cas9 proteins from other bacterial species are of interest at all, and why the enzyme in this article's opening image is not the famous one.

One sentence on that, because it is a different story from this one. Casgevy, approved by the FDA on 8 December 2023 for sickle cell disease and transfusion-dependent beta-thalassemia, was the first approved CRISPR therapy anywhere, and it works by an elegant sidestep: rather than repair the faulty adult haemoglobin gene, it edits a patient's own blood stem cells to disrupt an enhancer of BCL11A and switch the fetal haemoglobin gene back on. The clinical account, and the ethics of editing embryos, belong to our companion articles on gene therapy and on designed humans. This one is about the bacterium.

Twenty-six research files into this project, this is the first with no wrong pointer in it. Five of the fourteen entries carry an identifier. All five resolve exactly, on every field: journal, volume, issue, pages, year, authors. Jinek 2012 in Science, Cong 2013 in Science, Barrangou 2007 in Science, Mojica 2005 in the Journal of Molecular Evolution, Nishimasu 2014 in Cell. And not one of them is a review, which is what the pattern we have been testing across this series requires of a bibliography containing no books. We wrote that prediction down before checking, and also wrote down that it was a weak test, because these are famous papers whose identifiers look right on their face. Nothing failed, so nothing could fail in an informative way.

That is the real finding, and all nine of them should have had one. The other nine entries are papers in Nature, Science, the New England Journal of Medicine, Nature Biotechnology and the Annual Review of Biochemistry. Journals of that kind register an identifier for everything they publish; there is no such thing as one of these papers not having one. We found all nine and verified each against our file's own claimed volume, issue, pages, year and first author. Every one matched. That is a sixty-four per cent omission rate in the easiest case there is, and it is the largest single recovery we have made from one document.

An empty identifier field is not automatically wrong. It is not automatically right either. An earlier article in this series worked through a bibliography of old German monographs where five identifiers were all wrong, and established something counterintuitive: the correct repair there was to leave the field blank, because those books have no registered identifier anywhere in the world and filling the gap could only mean filling it with something false. This file is the mirror image. Here nine blanks are all defects. Put the two side by side and you get the rule: whether a blank is honest depends entirely on whether that kind of work gets registered at all. Anyone repairing a bibliography needs both halves. With only the first, they will leave modern papers bare; with only the second, they will invent identifiers for books that never had any.

Searching for the missing nine kept turning up other records wearing the paper's own title. A Faculty Opinions recommendation of the prime-editing paper. A Commentary on the adenine base-editor paper, in a different journal. A Publisher Correction to that same paper. And for the sickle cell trial, a piece of correspondence in the same journal whose title is nearly identical to the article's. Earlier in this series we described this decoy problem as something that happens to books, on the reasoning that a book has no neighbouring article and the only other record carrying its title is a review of it. That reasoning was too narrow. Articles accrete corrections, correspondence, commentaries and recommendation records. The decoy is a property of being worth citing, not of being a book.

The New England Journal of Medicine writes the answer into the string. Its identifiers carry a stem: nejmoa marks an original article, nejmc marks correspondence. The two records for the sickle cell trial differ by exactly that, and knowing the convention settles in one glance which is the paper and which is a letter about it. Almost no other publisher offers that help, and the reason it matters is the mechanism underneath all of this: when a bibliography is assembled by matching titles, the satellite and the paper look identical, and the identifier is the only thing that can tell them apart. Which is precisely the thing that keeps going missing.

The entry for the man who started all of this has his name broken in half. Our file's author column reads Mojica, Francisco J, with M., et al stranded out in a column of its own. The registry gives his name as Francisco J. M. Mojica. The orphaned M. is part of the man's name, cut loose and parked in the wrong field. We have now found this same mechanism shattering author lists in four different sections of our library, and it always breaks at an initial. It is a small thing next to a wrong citation, and it is worth recording because it is mechanical, repeatable, and therefore fixable everywhere at once.