More than ever, we can view the genomes of humans and other organisms as drafts—not final and canonical texts, but rough copies to be tweaked and refined. Although scientists have been able to edit genomes for many decades, their tools were often cumbersome to work with, expensive to hire, or sloppy in their efforts. And some were frustratingly artisanal: Tools like zinc finger nucleases and TALENs are specific and powerful, but you effectively need to train a new bespoke editor for every edit you want to make.

By contrast, CRISPR, the youngest technique on the block, is cheaper, more versatile, and more precise than its predecessors. And scientists are racing to improve it even further, developing new versions that are even more efficient, that can subtly change the emphasis of genetic words rather than deleting them outright, and that make fewer mistakes.

CRISPR consists of two components: An enzyme called Cas9—a pair of molecular scissors that grabs DNA and cuts it into two; and a guide RNA—a molecule that tells Cas9 exactly what to cut.

Imagine flipping through a book with a pair of scissors in one hand, and a word—say, “CUSTARD”—scrawled on the other in mirror-image. Whenever you see the matching word on a page, you cut it. It’s an extremely versatile system. The editor, Cas9, remains the same, and you just direct it to the right targets by supplying it with the right guide RNAs.

But Cas9 isn’t perfect. Sometimes, it will cut a target that mostly matches its guide RNA, but not entirely. It’s as if it sees “MUSTARD” and thinks: “Eh, close enough, let’s slice and dice.” These off-target cuts are a nuisance to CRISPR users, who have flocked to the technique precisely because it’s meant to be precise.

They also raise serious questions about proposed (speculative) uses for the technique, such as editing the genomes of embryos to prevent inherited disorders. If you’re going to try that, you’d better make sure that you’re not inadvertently activating a cancer gene or disrupting an essential one. Indeed, when a Chinese team recently (and controversially) used CRISPR to edit a disease gene in (inviable) human embryos, they found surprising and worrying levels of off-target cuts.

That’s why the many position statements about the ethics of CRISPR have universally highlighted the risks of off-target cuts, and the need to identify, understand, and avoid them. And it’s why several groups of scientists have been trying to develop ways of making CRISPR more specific, so that they make fewer off-target cuts.

The simplest involve lowering the levels of active Cas9. But this reduces the frequency of both off-target cuts, and on-target ones—you get a more specific editor, but also a less efficient one.  

You could also change the guide. Keith Joung from the Massachusetts General Hospital has shown that using shorter guide RNAs leads to fewer off-target cuts. When guides are long, the Cas9 enzyme will tolerate several mismatches before it fails; with shorter ones, any single mismatch threatens to derail the editor.  

CRISPR pioneers Feng Zhang and George Church are trying to change the Cas9 editor itself. First, they created half-hearted versions. The usual enzyme checks a DNA sequence against its guide RNA and then cuts both strands, but the mutant versions cut just one strand. So you need two of them, checking their own guide RNAs, to fully sever a stretch of DNA. It’s unlikely that both enzymes will get things wrong, so together, they become more specific.

Today, Zhang has unveiled yet another strategy for engineering Cas9. First, his team searched for mutations that will make Cas9 more discerning in its attacks, so that it only cuts DNA that perfectly matches its guide RNA template. They found five, which make the enzyme more specific but no less efficient. They then tested these mutations in combinations, until they settled on one particularly judicious version of Cas9, which they called esCas9 (“enhanced specificity Cas9”). This upgraded Cas9 cuts its targets just as well as Cas9 Classic, but never veers off target—at least, not that the team could detect. It’s a precision weapon that inflicts no collateral damage.

These methods are complementary, Zhang told me at the International Summit on Human Gene Editing. They could all be fused together—for example, by building a version of esCas9 that cuts just one strand and relies on shorter guides. “Maybe some kind of superposition will lead to the ultimate system,” he says. “And maybe you’ll have a whole toolbox of tools and you’ll pick the right one for the application in mind.”

But at the same meeting, Jin-Soo Kim at Seoul National University asked if off-target effects even matter. No drugs are free of side effects, including some that directly affect DNA. Etoposide is a drug that cuts DNA at random places, but is widely used to treat several cancers. By contrast, the specificity of even the basic Cas9 enzyme is far higher.

Meanwhile, Keith Joung noted that many teams, including his own, have developed ways of measuring off-target cuts. What’s missing is a way of comparing these methods against each other to see which performs best; without that, there’s no yardstick with which to gauge how common these problems are and how well the solutions are actually dealing with it. Partly, that’s because the field is moving very quickly and it would take time for each group to master the others’ methods. “I’m hoping that a company will compare these methods, so they can go to regulators,” he says.