What Is CRISPR and How Does Gene Editing Actually Work?

CRISPR gets described as "molecular scissors" so often that the phrase has stopped meaning anything. Scissors don't search a 3-billion-letter genome for one 20-letter sequence and ignore everything else. The Cas9 protein does. This article skips the discovery story and the medical headlines, both covered elsewhere on this site, and goes down to the level of domains and base pairs to explain how a bacterial protein finds its target, decides it has the right one, and cuts.

Educational content, not medical advice. This article explains the molecular mechanism behind CRISPR-Cas9. It is not guidance for any medical treatment or DIY genetic modification.

What CRISPR-Cas9 is made of

CRISPR is shorthand for a bacterial defense system, and the part that matters for editing is a two-component machine: a protein (Cas9, short for CRISPR-associated protein 9) and a short RNA that tells it where to go. A 2012 paper by Jinek and colleagues showed that in the system from Streptococcus pyogenes, a CRISPR RNA (crRNA) paired with a second RNA called tracrRNA forms a two-RNA structure that directs Cas9 to make double-stranded breaks in DNA. The same paper showed that the two RNAs could be fused into a single chimeric guide that worked too (Jinek et al., Science, 2012). That single guide RNA (sgRNA) is what almost every lab uses today.

Two things follow from this design. First, the protein is the same for every target; only the roughly 20-nucleotide stretch at one end of the guide changes. Second, the protein does not read DNA the way a person reads a page. It reads it by chemistry: the guide RNA either base-pairs with the DNA or it does not.

An analogy helps here. Cas9 is a security guard with a photo ID. The guard does not memorize every face in the building. The guard walks the corridors, glances at each door, and only stops to compare the photo against a person when the door has a particular kind of handle. That handle is the PAM, which comes next.

The search problem Cas9 has to solve

Picture a phone book with three billion entries and no index. You're handed a 20-character string and told to find the one entry that matches. That is roughly the scale of the problem Cas9 solves every time it edits a gene.

The human genome contains about 3.1 billion base pairs, and a guide RNA is typically 20 nucleotides long. A random 20-letter sequence is unlikely to occur twice by chance in a genome of that size, but near-matches are common enough that off-target binding is a real, measurable phenomenon, and a large part of the field exists to minimize it.

Cas9 does not solve this by reading all 3.1 billion letters one at a time. It uses a two-stage filter. A fast, cheap check runs across the whole genome, and a slow, expensive verification step runs only at the sites that pass. That two-stage design is the mechanism worth understanding, and it begins with a three-letter tag called the PAM.

In test-tube experiments with bacteriophage lambda DNA, Cas9 stayed bound to competitor DNA lacking PAMs for a negligible time. The protein rejects most of the genome without ever unzipping it, because a perfect guide match next to no PAM is invisible to it (Sternberg et al., 2014).

How Cas9 finds one sequence in three billion

The PAM: Cas9's first filter

Before Cas9 checks whether a DNA sequence matches its guide RNA, it checks for a short marker called the protospacer-adjacent motif (PAM), which sits right next to the target site. In the most widely used Cas9, from S. pyogenes (SpCas9), that marker is NGG, where N is any base and G is guanine.

The key experiments on this were done by Sternberg and colleagues in 2014 using single-molecule and bulk biochemistry. They reported that both DNA binding and cleavage require recognition of a short trinucleotide PAM, and that a fully complementary target with a single PAM mutation could not compete for cleavage, so it behaved like DNA with no complementarity at all (Sternberg et al., Nature, 2014). Put plainly, a target without a PAM beside it does not exist as far as Cas9 is concerned, no matter how well the guide RNA matches it.

The same study measured how Cas9 gets around the genome. Target sites were found by three-dimensional collisions, with no evidence of the protein sliding along the DNA. Binding to non-target DNA correlated with how many PAMs the DNA contained, and in the bacteriophage lambda DNA used as a test substrate there is one PAM roughly every 8.5 base pairs. Competitor DNA with no PAMs had negligible residence time, so Cas9 leaves PAM-free DNA almost immediately. In effect, Cas9 bounces from PAM to PAM and only lingers when the neighbor is promising.

This is also why PAM availability limits which genome positions SpCas9 can edit, and why researchers have engineered Cas9 variants and discovered entirely different Cas proteins that recognize different PAM sequences, a topic we return to below.

Unzipping the helix and forming the R-loop

Once Cas9 finds a PAM, it needs to check the bases just beside it against the guide RNA. DNA does not hand over its sequence without being unzipped, so Cas9 has to pry the two strands apart at that spot.

This is where the guide RNA takes over the search. It begins pairing with one of the two exposed strands, base by base, starting from the end nearest the PAM. As pairing proceeds it displaces the other strand, forming a three-stranded structure called an R-loop: an RNA-DNA hybrid on one side and a displaced, single-stranded loop of DNA on the other. Sternberg and colleagues found that strand separation starts at the PAM and proceeds directionally toward the far end of the target, and that about 12 base pairs of pairing, roughly one helical turn, are needed before further propagation becomes energetically favorable.

The seed: why the first dozen bases decide everything

The part of the guide closest to the PAM is called the seed, and it is where mismatches hurt most. In Sternberg's experiments, competitor inhibition rose sharply once 12 or more base pairs next to the PAM were complementary, which the authors link to the 8 to 12 nucleotide seed reported in earlier work. Mismatches at positions 1 and 2 beside the PAM abolished binding even when 18 further base pairs matched, whereas mismatches at the far, PAM-distal end still allowed binding. The study also showed that a pre-formed DNA bubble at those first two positions bypassed the block, so the barrier is about starting to unwind the helix, not about the identity of the bases.

If the seed does not match, Cas9 lets go and moves on. Only if the seed pairs well does the R-loop extend toward the PAM-distal end of the 20-base target. A useful analogy is a combination lock whose first few numbers are checked quickly, and only a correct start earns you the patience to dial the rest.

The two-lobed protein that holds it all together

None of this pairing would be stable without the physical architecture of Cas9 itself. The crystal structure of SpCas9 bound to a guide RNA and target DNA, solved at 2.5 angstrom resolution (an angstrom is a ten-billionth of a meter), showed a protein built from two lobes. A recognition lobe grips the guide RNA and the growing RNA-DNA hybrid, and a nuclease lobe houses the two cutting domains and a C-terminal domain that contacts the PAM. The hybrid sits in a positively charged groove at the interface between the lobes (Nishimasu et al., Cell, 2014; the deposited structure includes roughly 1,370 amino acid residues of the protein).

Think of it as a clamshell with a built-in ruler. The recognition lobe measures how far the pairing has progressed, and the nuclease lobe holds the blades, waiting for the signal that the target is confirmed.

The cut: two domains, two jobs

HNH and RuvC

The nuclease lobe contains two different cutting domains. The original 2012 work showed that the HNH domain cuts the DNA strand that is complementary to the guide RNA (the target strand), and the RuvC-like domain cuts the other strand (the non-target strand). A study of a related Cas9 found that the HNH domain cuts at a fixed position relative to the protospacer, while the position of the RuvC cut depends on the distance from the PAM, which the authors interpreted as a ruler-like mechanism (Chen et al., J Biol Chem, 2014; this work used the Cas9 from Streptococcus thermophilus, so treat the ruler detail as a feature of that enzyme).

Domain	Strand it cuts	How its cut position is set	Evidence
HNH	Target strand (paired with the guide RNA)	Fixed position relative to the protospacer	Jinek et al., 2012; Chen et al., 2014
RuvC	Non-target strand (the displaced strand)	Depends on distance from the PAM (a "ruler") in the S. thermophilus enzyme	Jinek et al., 2012; Chen et al., 2014

Who goes first, and who is in charge

A tidy picture would be: the protein cuts one strand, then the other. Reality is more interesting. A 2015 study from the Doudna lab used fluorescence energy transfer to watch the HNH domain and found that cleavage efficiency tracks how often HNH samples an activated conformation. Cleavage of the non-target strand by RuvC correlated with that activated HNH state, and deleting HNH abolished RuvC activity. The authors propose that HNH conformational proofreading acts as a final specificity check after PAM recognition and RNA-DNA pairing, and before a double-strand break forms (Sternberg et al., Nature, 2015).

So HNH is less a blade than a gatekeeper. It has to swing into an active position, and that swing licenses RuvC. This is Cas9's third filter, and it is the reason the protein tolerates some distal mismatches (the study found that substrates with one to three distal mismatches were still cleaved at near wild-type rates) but cleaves slowly or not at all when four or more mismatches are present.

The order of the two cuts is not fixed. A 2021 kinetic study with supercoiled plasmids described two parallel routes to a double-strand break, one in which HNH nicks the target strand first and one in which RuvC cuts the non-target strand first. For SpCas9 on perfectly matched DNA, the target-strand-first route was favored, at about 72 percent versus 28 percent (Babu et al., Biochemistry, 2021). The picture of strictly sequential cuts that you may see in simplified diagrams therefore overstates the case.

What disabling one blade does

The two-domain design is why disabling either domain gives researchers a different tool. A Cas9 with one nuclease domain inactivated becomes a "nickase" that cuts a single strand. Inactivate both and you get "dead" Cas9 (dCas9), which still finds and holds its target but never cuts. Dead Cas9 is used to block a gene physically or to carry other proteins to a specific genomic address.

After the cut: how cells repair a double-strand break

Cas9 only makes the break. What ends up in the genome depends on the cell's repair machinery, which treats a broken chromosome as an emergency.

• Non-homologous end joining (NHEJ). The cell glues the broken ends back together without a template. It is fast and common, and it often introduces small insertions or deletions (indels). If the cut falls in the coding sequence of a gene, those indels commonly disrupt the reading frame and knock the gene out.
• Homology-directed repair (HDR). If a DNA template with the desired sequence is supplied alongside Cas9, the cell can copy it into the break and make a precise change. HDR is less efficient than end joining in mammalian cells, which is a major reason that precise edits are harder than disruptions.

The practical consequence is that "cutting" and "editing" are different things. A cut is an invitation to the repair machinery, and the outcome is a distribution of results, not a single guaranteed sequence. Komor and colleagues put the problem bluntly in the paper that introduced base editing: current approaches to point-mutation correction are inefficient and typically induce an abundance of random indels at the target locus (Komor et al., Nature, 2016).

Why the molecular details create real limitations

Off-target cutting is a probability problem

Every step above is a statistical filter, not a guarantee. A PAM match, plus a seed match, plus enough pairing to activate HNH, is a high bar, but it is not an impossible one. A near-match elsewhere in the genome can occasionally clear all three. That is the molecular root of the off-target editing problem. It is not that Cas9 is sloppy; a multi-step probabilistic search has a nonzero false-positive rate by construction.

Sternberg's 2014 study drew a useful conclusion from the PAM dependence: off-target assessment need only consider sites adjacent to a PAM, which the authors present as an inference from in vitro data and not a direct test in cells. For an NGG-recognizing enzyme that still means a very large number of candidate sites, because NGG occurs so frequently.

Engineering stricter checkpoints

Researchers address off-targets from both ends. One approach is a protein-engineering one. In 2016, Kleinstiver and colleagues built SpCas9-HF1, a variant with alterations designed to reduce non-specific DNA contacts. It kept on-target activity comparable to wild-type SpCas9 with more than 85 percent of the guides tested in human cells, and for standard targets it made all or nearly all off-target events undetectable by the genome-wide methods the team used (Kleinstiver et al., Nature, 2016). The abstract's wording matters: "all or nearly all" and "undetectable by those methods" is not the same as "none."

The other approach is computational: screening guide RNAs against the genome to avoid ones with many near-matches before they are ever synthesized.

The PAM requirement is a hard constraint

Because SpCas9 will not engage a target without an adjacent PAM, stretches of the genome are off-limits to it regardless of how well the guide is designed. In an NGG-recognizing system, that is a practical bottleneck in therapy design, and it is part of why the toolbox now includes other Cas proteins and engineered variants with different PAM requirements.

Not every Cas protein reads the same password

SaCas9: same job, smaller lock

SpCas9's biggest practical limitation is its size. At 1,368 amino acids, its gene is large for the adeno-associated virus (AAV) vectors commonly used to deliver gene editors in living animals, which have limited carrying capacity. The Cas9 from Staphylococcus aureus (SaCas9) is 1,053 residues, about 315 residues smaller. A 2015 Nature study tested six smaller Cas9 orthologs and found SaCas9 edits genomes about as efficiently as SpCas9 while being over 1 kilobase shorter in coding sequence. The team packaged SaCas9 and its guide into a single AAV and targeted the Pcsk9 gene in mouse liver, modifying more than 40 percent of the gene within a week, with significant drops in serum Pcsk9 and total cholesterol (Ran et al., Nature, 2015).

The trade-off is the password. SaCas9 recognizes an NNGRRT PAM, where R stands for either A or G (Friedland et al., Genome Biology, 2015). A longer, more specific PAM is rarer in the genome, which shrinks the number of eligible sites. It is the same trade a six-pin lock makes against a three-pin lock: harder to pick, but fewer keys fit.

Cas12a: a different family, a different rulebook

Cas12a (originally called Cpf1) is not a smaller Cas9. It is a protein from a different family, described by Zetsche and colleagues in 2015, that works with a single RNA and no tracrRNA, recognizes a T-rich PAM (the reverse of Cas9's G-rich one), and makes a staggered double-strand cut rather than a blunt one. In that study, two of 16 Cpf1-family proteins tested, from Acidaminococcus and Lachnospiraceae, edited efficiently in human cells (Zetsche et al., Cell, 2015).

The T-rich PAM lets Cas12a reach AT-rich stretches of the genome that an NGG-reading enzyme skips. And the staggered cut leaves short single-stranded overhangs, similar to some restriction enzymes, which can be useful when an inserted sequence needs to go in a defined orientation.

Nuclease	Source organism	PAM	PAM position relative to target	Cut	Protein size
SpCas9	Streptococcus pyogenes	NGG	Downstream (3')	Typically blunt	1,368 amino acids
SaCas9	Staphylococcus aureus	NNGRRT	Downstream (3')	Typically blunt	1,053 amino acids
AsCas12a / LbCas12a	Acidaminococcus / Lachnospiraceae	T-rich (often written TTTV)	Upstream (5')	Staggered	Not compared here

Sources: Ran et al., 2015; Friedland et al., 2015; Zetsche et al., 2015; Sternberg et al., 2014. The 3' versus 5' positions follow the usual convention for these enzymes.

How researchers catch off-target cuts

Knowing that off-target cutting is possible is not the same as knowing where it happens in a particular genome. Two complementary experimental approaches dominate.

CIRCLE-seq: testing naked DNA outside a cell

CIRCLE-seq is an in vitro biochemical screen. Genomic DNA is fragmented and circularized, and Cas9 is added to the library of circles; each cut linearizes a circle, and sequencing then identifies the cut site. The authors, Tsai and colleagues, describe it as a highly sensitive and sequencing-efficient strategy that does not need living cells or a reference genome. In their test with a guide for the HBB gene on K562 cell DNA, CIRCLE-seq found 26 of the 29 off-target sites previously identified by GUIDE-seq and 156 new ones (Tsai et al., Nature Methods, 2017).

Because bare DNA is more physically accessible than DNA wound around proteins in a nucleus, CIRCLE-seq tends to flag sites a cell would never actually cut. That is a feature for characterizing a guide's biochemical preferences, and a caveat for interpreting them.

GUIDE-seq: catching cuts in living cells

GUIDE-seq works in living, cultured cells. A short double-stranded DNA tag is delivered with the editing machinery and gets captured at double-strand breaks, wherever they occur. Sequencing then maps every site where the tag was captured, which gives a direct record of breaks made and repaired inside a real cell. In the original 2015 study, the authors applied the method to 13 RNA-guided nucleases in two human cell lines and found that off-target activity varied widely between nucleases and that most sites identified were missed by computational prediction (Tsai et al., Nature Biotechnology, 2015).

Method	Where it runs	What it measures	Typical strength	Typical limitation
CIRCLE-seq	Test tube, purified genomic DNA	Where Cas9 can cut naked DNA	Very sensitive	May flag sites a cell would not cut
GUIDE-seq	Living cultured cells	Where breaks actually occurred and were repaired	Reflects chromatin and cell biology	Needs efficient tag delivery; cell type matters

No single assay sees the whole picture, which is why development of a therapeutic editor usually layers several methods.

Beyond cutting: base editing and prime editing

Because double-strand breaks lead to unpredictable repair, a generation of tools rebuilt Cas9 so that it edits without making that break.

Base editors. In 2016, Komor and colleagues fused a Cas9 variant to a cytidine deaminase enzyme. The fusion is still guided by RNA but does not induce double-strand breaks, and it converts cytidine to uridine within a window of about five nucleotides, producing a C-to-T change (or G-to-A on the other strand). In four human and mouse cell lines, the second- and third-generation editors corrected roughly 15 to 75 percent of cellular DNA, with indels typically at or below 1 percent (Komor et al., 2016).

Prime editors. In 2019, Anzalone and colleagues described "search-and-replace" editing using a Cas9 that cuts only one strand, fused to a reverse transcriptase (an enzyme that writes DNA from an RNA template). The guide RNA both locates the target and carries the intended edit. The authors reported more than 175 edits in human cells, including insertions, deletions, and all 12 types of point mutation, without double-strand breaks or donor DNA templates, and lower off-target editing than Cas9 nuclease at known off-target sites (Anzalone et al., Nature, 2019).

A simple analogy: standard CRISPR cutting is like tearing a page out and trusting someone to tape it back in the right order. Base editing is like using a correction pen on one letter. Prime editing is closer to a find-and-replace function. None of the three is risk-free, and each has its own limits on which changes it can make.

One real-world example

The first CRISPR-based therapy cleared by the US Food and Drug Administration was Casgevy, approved on December 8, 2023 for sickle cell disease in patients 12 and older with recurrent vaso-occlusive crises. According to the FDA, the treatment modifies a patient's own blood stem cells using CRISPR-Cas9, and the edited cells are transplanted back after high-dose chemotherapy, where they increase production of fetal hemoglobin, which prevents red blood cells from sickling (FDA announcement). The status of other therapies and the hype around them is covered in CRISPR gene editing in 2026: what's been cured and what's hype.

Common misunderstandings

"CRISPR is a pair of scissors." The cut is the last of several steps. The PAM check, seed pairing, and HNH activation all come first, and the cell's repair machinery decides the final result.

"Cas9 reads the genome." Cas9 does not scan sequences letter by letter. It samples PAM sites by collision and tests each with its guide RNA.

"CRISPR is perfectly precise." The system is highly selective but probabilistic. High-fidelity variants and careful guide design lower off-target editing, and sensitive assays like GUIDE-seq and CIRCLE-seq exist because the risk is real.

"HNH cuts one strand, then RuvC cuts the other, always in that order." The domains are coupled, and either order can occur. HNH's activation licenses RuvC, and the favored order on matched DNA is HNH first in plasmid experiments.

"Gene editing and gene therapy are the same thing." Editing changes a sequence that already exists; gene therapy more broadly adds or modifies genetic material to treat disease. Editing is one way to do gene therapy, and our explainer on gene therapy covers the broader field.

Where to go next

If you want the mechanism of DNA itself, start with how DNA works, then DNA replication and genetics, which explains why a double-strand break is such a serious event for a cell. For the historical thread, see CRISPR gene editing: how it works in 2026, the CRISPR breakthrough as a landmark experiment, and the invention of CRISPR gene editing. For how machine learning is used to design guides and predict edits, read AI meets CRISPR. And for the gene-level background, how genes actually work covers what a "target" is in the first place.

What this means

Cas9 is not a pair of scissors. It is a multi-stage verification system: a cheap PAM scan across the genome, a fast seed check at PAM sites, a slower extension of the R-loop at strong matches, and a gatekeeper domain, HNH, that has to activate before the cut is licensed. Every headline about a CRISPR therapy or a CRISPR safety concern traces back to this chain. PAM availability decides what is even editable, seed stringency decides how selective a guide needs to be, and the repair pathway the cell picks decides what the edit finally looks like. Understanding the chain is what separates "gene editing is precise" from knowing why, and where, it is not.

Frequently Asked Questions

Q: What exactly is a PAM sequence, and why does Cas9 need one? A: The protospacer-adjacent motif is a short DNA sequence next to the target site, NGG for the common SpCas9. Sternberg and colleagues showed that Cas9 binding and cleavage both require PAM recognition, and that a perfectly matching target with a mutated PAM behaved like DNA with no match at all.

Q: What is an R-loop and why does it form during CRISPR editing? A: An R-loop is a three-stranded structure formed when the guide RNA pairs with one DNA strand and displaces the other as a single-stranded loop. It is the physical evidence that the guide matches the target. Experiments show roughly 12 base pairs of pairing are needed before further unwinding becomes energetically favorable.

Q: Do HNH and RuvC cut DNA at the same time? A: Not exactly. HNH cuts the target strand and RuvC cuts the non-target strand, but the two are coupled. HNH has to reach an activated conformation before RuvC works, and kinetic studies show either cut can come first, with the target-strand-first route favored on perfectly matched DNA in plasmid experiments.

Q: Why doesn't Cas9 cut every near-match sequence in the genome? A: Several filters stack. The sequence needs a PAM, the seed region next to it must pair well, and HNH must reach its active state. A mismatch in the first two positions beside the PAM abolished binding in one study. Near-matches that clear all three filters can still be cut, which is the source of off-target edits.

Q: How do researchers find off-target edits? A: By combining methods. CIRCLE-seq cuts purified genomic DNA in a test tube and is very sensitive, and GUIDE-seq tags breaks in living cells and shows where cuts actually occurred. In one comparison CIRCLE-seq found 26 of 29 sites previously identified by GUIDE-seq and 156 more, so the assays do not simply agree.

Q: What is the difference between a cut, a base edit, and a prime edit? A: A Cas9 cut makes a double-strand break and leaves the outcome to cellular repair. A base editor chemically converts one base without that break, within a window of about five nucleotides. A prime editor uses a one-strand-cutting Cas9 fused to a reverse transcriptase to write a new sequence directly, again without a double-strand break.

Q: Why do scientists use smaller Cas proteins like SaCas9? A: Size limits delivery. SpCas9 is 1,368 amino acids and SaCas9 is 1,053, so SaCas9 plus its guide fits more easily into a single adeno-associated virus. The cost is a stricter PAM, NNGRRT, which leaves fewer eligible target sites.

Sources

• Jinek M, et al. A programmable dual-RNA-guided DNA endonuclease in adaptive bacterial immunity. Science 2012;337:816-821.
• Sternberg SH, et al. DNA interrogation by the CRISPR RNA-guided endonuclease Cas9. Nature 2014;507:62-67.
• Nishimasu H, et al. Crystal structure of Cas9 in complex with guide RNA and target DNA. Cell 2014;156:935-949. Structure entry: RCSB PDB 4OO8.
• Chen H, Choi J, Bailey S. Cut site selection by the two nuclease domains of the Cas9 RNA-guided endonuclease. J Biol Chem 2014;289:13284-13294.
• Sternberg SH, et al. Conformational control of DNA target cleavage by CRISPR-Cas9. Nature 2015;527:110-113.
• Babu K, et al. Coordinated actions of Cas9 HNH and RuvC nuclease domains. Biochemistry 2021.
• Ran FA, et al. In vivo genome editing using Staphylococcus aureus Cas9. Nature 2015;520:186-191.
• Friedland AE, et al. Characterization of Staphylococcus aureus Cas9: a smaller Cas9 for all-in-one adeno-associated virus delivery and paired nickase applications. Genome Biology 2015.
• Zetsche B, et al. Cpf1 is a single RNA-guided endonuclease of a class 2 CRISPR-Cas system. Cell 2015;163:759-771.
• Kleinstiver BP, et al. High-fidelity CRISPR-Cas9 nucleases with no detectable genome-wide off-target effects. Nature 2016;529:490-495.
• Tsai SQ, et al. GUIDE-seq enables genome-wide profiling of off-target cleavage by CRISPR-Cas nucleases. Nat Biotechnol 2015;33:187-197.
• Tsai SQ, et al. CIRCLE-seq. Nat Methods 2017;14:607-614.
• Komor AC, et al. Programmable editing of a target base in genomic DNA without double-stranded DNA cleavage. Nature 2016;533:420-424.
• Anzalone AV, et al. Search-and-replace genome editing without double-strand breaks or donor DNA. Nature 2019;576:149-157.
• US Food and Drug Administration. FDA approves first gene therapies to treat patients with sickle cell disease, December 8, 2023.

---

Read next:

• What Is Gene Therapy — and How Is It Curing Genetic Diseases?
• How Do Genes Actually Work?
• How Does DNA Work? The Molecular Blueprint of Life Explained