khazana — your personal signal terminal khazana — your personal signal terminal
No. 019 primer

The Molecular Scissors: How CRISPR-Cas9 Finds One Target in Three Billion

How a bacterial protein finds one 20-nucleotide sequence in three billion bases — the cascade of molecular events behind CRISPR-Cas9, from a 1987 accident in an E. coli lab to the first approved gene therapy in 2023.

science

16 min read 17 sources

Here is the question worth sitting with before any molecular machinery appears: how does a protein find one specific 20-letter sequence in a three-billion-letter genome — not approximately, not in the right neighborhood, but the exact target, at nucleotide resolution — without a map, without a GPS signal, without any prior knowledge of where in three billion bases to look?

Not only does it find it. It finds it with enough fidelity to be used as a medicine. Get one letter wrong in the wrong place and the protein falls off without cutting. Get every letter right and it holds on for hours, triggering a double-strand break that, when repaired, permanently edits the genome. This precision — at the level of a single nucleotide, across three billion of them — is what makes CRISPR-Cas9 not merely a useful tool but a conceptually startling one.

The answer is a chain of molecular events: a bacterial protein that interrogates candidate sites one by one, runs a rapid specificity test at each, and commits to cutting only when every step of that test passes. The mechanism was worked out in a biochemistry paper published in June 2012. It was awarded the Nobel Prize in Chemistry eight years later. And in December 2023 it was deployed inside the blood cells of patients with sickle cell disease — the first gene-editing therapy approved in the United States.

This primer builds the mechanism from the ground up. The biology can wait for a moment; the underlying question comes first.

GENOME BASEShaploid human reference
GUIDE RNA LENGTHnucleotides specify the target
YEARSIshino 1987 to Casgevy 2023
The numbers that frame the target-finding problem.

The misconception that blocks the whole story

Most people arrive at CRISPR with a GPS model in their heads: you program a target sequence, the protein navigates to it, and the cut happens. In this model, the protein has an address. The genome is the city. You type in the coordinates and the scissors go there.

This model is wrong in a way that matters. Cas9 has no address book. It has no prior knowledge of where the target sits. What it has is a guide RNA — a 20-nucleotide strand that can base-pair with the target — and a rule about what kind of DNA sequence to interrogate first. It then does something far more interesting than navigating: it physically samples the genome, testing candidate sites one at a time, rejecting the wrong ones within milliseconds, and committing only when the molecular test passes.

The first step in that test is recognizing a three-letter DNA motif called a Protospacer Adjacent Motif — a short DNA sequence (5'-NGG-3' for the most common Cas9) that must flank the target on the non-complementary strand. Cas9 recognizes the PAM before attempting any guide-RNA pairing. Without a PAM, no interrogation begins.. For the most widely used Cas9 — from Streptococcus pyogenes — the PAM is the three-letter sequence NGG, where N is any base and the two G’s are required. The protein scans along DNA until it finds an NGG, then and only then attempts to open the double helix to check whether the adjacent sequence matches the guide.

How many NGG sites are in the human genome? Run the calculation:

runnable · js
output
Run the code to see its output here.
How many potential interrogation sites exist in the human genome? Edit the genome size or GG frequency to explore.

Run this and the shape of the problem becomes clear. There are an estimated 100–200 million NGG sites in the human genome — roughly one every 30–60 bases on average, derived from the approximately 1-in-16 per-base GG dinucleotide frequency across both strands of the 3.2 Gb genome — and Cas9 must interrogate each one. With a perfect 20-nucleotide match, the address space (42010124^{20} \approx 10^{12}) is about 300 times larger than the genome, so the target appears essentially once. But allow even three mismatches anywhere in the guide and the combinatorics multiply the risk toward a handful of sites; allow five mismatches and hundreds of potential addresses exist. The GPS model collapses here: the precision is not programmed in as an address. It is earned, step by step, at each of those estimated ~100–200 million candidate positions.

The five-step interrogation at every PAM

Every time Cas9 encounters an NGG in the genome, it runs the same sequence of molecular tests. Think of it as a five-stage filter, where failing any stage ends the encounter in milliseconds. Passing all five stages takes minutes — and triggers the cut.

  1. Stage 1 — PAM recognition

    The PAM-interacting (PI) domain of Cas9 reads the minor groove of the NGG motif on the non-complementary strand. This recognition happens without unwinding the double helix. If no NGG is present, Cas9 diffuses onward in under a millisecond. The PAM is the entry ticket — without it, the guide RNA never gets a chance to pair.

    This matters enormously for specificity: the PAM requirement restricts interrogation to an estimated ~100–200 million sites rather than the full 3.2 billion — eliminating the vast majority of genomic positions before any base-pairing begins.

  2. Stage 2 — Seed invasion

    Once the PI domain locks onto an NGG, the REC lobe of Cas9 begins melting the DNA duplex immediately upstream of the PAM. The first 2–3 nucleotides of the guide RNA attempt base-pairing with the complementary strand. This is the seed region initiation: the R-loop forms here first, and any mismatch in these opening positions collapses the attempt before it can propagate.

    Think of it as a key entering a lock — the first few tumblers must engage before the mechanism goes further.

  3. Stage 3 — R-loop propagation

    If the seed bases pair correctly, the R-loop — a three-stranded structure in which the guide RNA displaces one DNA strand and base-pairs with the complementary strand, leaving the other strand looped out as single-stranded DNA — propagates toward the PAM-distal end of the protospacer, unzipping the DNA helix base pair by base pair across the full 20-nucleotide target. This propagation is directional — it proceeds from PAM-proximal (seed) toward PAM-distal — and mismatches near the PAM collapse it while mismatches far from the PAM are better tolerated. The fully formed R-loop spans 20 base pairs.

  4. Stage 4 — HNH conformational change

    A complete 20-bp R-loop triggers a dramatic structural rearrangement. The HNH nuclease domain — which had been held away from the DNA, exhibiting significant flexibility in the apo state — undergoes a large conformational swing to position its catalytic residue directly over the complementary (guide-paired) strand. This conformational change is the rate-limiting step for cleavage: it requires the R-loop to be stable for long enough for the domain to reposition, which takes on the order of minutes for a perfect match.

    If the R-loop is destabilized by a seed mismatch, the complex falls apart before HNH can complete its reorientation. The conformational change is a kinetic checkpoint built into the protein structure.

  5. Stage 5 — Double-strand break

    With HNH positioned over the complementary strand and the RuvC domain positioned over the non-complementary strand, both nuclease domains fire. HNH cleaves the complementary (guide-paired) strand ~3 bp upstream of the PAM; RuvC cleaves the non-complementary strand 3–8 bp upstream of the PAM. The resulting break has predominantly blunt ends, producing a double-strand break (DSB) at a predictable genomic location. The cell's DNA repair machinery then takes over — either introducing insertions or deletions (indels) via error-prone NHEJ, or, if a repair template is supplied, making a precise edit via homology-directed repair (Jinek et al. 2012).

The five stages of Cas9 target interrogation — from PAM recognition to double-strand break.

The five-stage cascade is not an accident of molecular architecture. Each stage is a specificity filter layered on top of the previous one. PAM recognition eliminates the vast majority of the genome from consideration before any base-pairing occurs — restricting interrogation from 3.2 billion positions to an estimated ~100–200 million NGG sites. Seed region interrogation catches the mismatches most likely to cause off-target cuts. The HNH conformational gate adds a kinetic hurdle. The result is a system that achieves nucleotide-level specificity without any active search mechanism — purely through the physics of molecular recognition and the kinetics of R-loop formation.

The mathematical intuition for why 20 nucleotides is enough to specify a unique target:

420=1,099,511,627,7761012    3.2×1094^{20} = 1{,}099{,}511{,}627{,}776 \approx 10^{12} \;\gg\; 3.2 \times 10^9
four possible bases at each of 20 positions — the full address space of a 20-nt sequence
roughly one trillion unique sequences
three billion base pairs — about 300 times smaller than the guide's address space
The 20-nucleotide guide has roughly 1 trillion possible sequences — about 300 times more than there are bases in the human genome. A perfectly matched 20-nt guide is effectively unique.

With a perfect 20-nt guide, any given sequence appears roughly once in the human genome. The NGG PAM adds additional specificity: only the ~1-in-16 positions adjacent to an NGG can even be considered. Together, equation (1) and PAM filtering explain why well-designed guides find their target without false positives — and why mismatches change the picture fast.

The seed region: why position is everything

Not all mismatches are equal. A mismatch at position 1 — immediately adjacent to the PAM — aborts the interrogation at stage 2. A mismatch at position 20 — the far end of the guide — is often tolerated. The difference is mechanistic: the R-loop starts at the PAM-proximal end and propagates outward, so mismatches near the PAM appear early in the propagation and collapse the entire structure before it can stabilize.

The region most sensitive to mismatch — roughly the first 10 base pairs from the PAM — is called the The 8–12 PAM-proximal nucleotides of the guide RNA where mismatches are least tolerated. A single mismatch here typically reduces cleavage efficiency by more than 70% and often abolishes it entirely, because the R-loop must traverse this region first.. Jinek et al. 2012 showed that toleration of mismatches in the PAM-distal region (positions 11–20) was substantially higher, while even a single mismatch in the seed region dramatically reduced or abolished cleavage. The structural explanation arrived with the 2014 crystal structure: The Cas9 REC lobe pre-orders the seed-region nucleotides with their Watson-Crick faces exposed, ready to nucleate an RNA-DNA hybrid upon contact with the target strand. Nishimasu et al. 2014 (Cell, PMC4139937) report that the RNA-DNA heteroduplex is distorted from canonical A-form rather than being pre-formed into it — the nucleotides are pre-positioned for base-pairing, not pre-folded into a canonical helix. Any mismatch in this exposed, pre-ordered region is kinetically very costly. the protein pre-orders the seed-region nucleotides — exposing their Watson-Crick faces for base-pairing — before any DNA engagement. Any mismatch here fights against a structural scaffold that was built to expect a perfect match.

Schematic relative cleavage efficiency (%) when a single mismatch is introduced at each position counting from the PAM (position 1 = immediately adjacent to PAM, position 20 = PAM-distal). The seed region (positions 1–10) is hypersensitive; PAM-distal mismatches are better tolerated. Based on Jinek et al. 2012 (PMC6286148) and Doench et al. 2017 (PMC5427927).

The asymmetry in the chart is the R-loop propagation direction made visible. Position 1 (next to the PAM) is where the R-loop forms first — a mismatch here prevents the structure from nucleating at all. Position 20 (PAM-distal) is where the R-loop arrives last — by the time it reaches here, enough of the 20-bp helix has already formed that a single mismatch at the far end may not destabilize the whole structure sufficiently to prevent cleavage.

Go deeper: why R-loop formation is mechanically one-directional

The directionality of R-loop propagation — seed first, PAM-distal last — is not arbitrary. PAM recognition by the PI domain positions the guide RNA so that its 3’ end (which pairs with the PAM-proximal strand) is the first segment presented to the target DNA. The REC1 domain then stabilizes the growing RNA-DNA hybrid as it extends away from the PAM. This means the protein is designed to pay the energetic cost of seed-region pairing first and only invest further if the seed succeeds. Multiple mismatches in the seed region pay a compounded cost: not only do they destabilize the local hybrid, they prevent propagation to the regions where the protein would gain compensating stability from the PAM-distal pairs. The mechanism is therefore not just binary (match/no match) but a graded investment: the protein bets progressively more energy on the target as each new base pair forms, and a bad bet early is the cheapest possible failure mode.

Kinetic proofreading: dwell time as the cleavage gate

The seed region filter is powerful but not absolute — some mismatches outside the seed still produce R-loops. How does Cas9 handle those? The answer is in the clock: a fully formed, stable R-loop keeps the HNH domain in the pre-cleavage conformation long enough for the domain to complete its rotation and fire. An unstable R-loop — caused by a mismatch that destabilizes the complex without collapsing it immediately — shortens the dwell time below the threshold for HNH repositioning.

This was measured directly in living cells by Ma et al. in 2016, using fluorescence recovery after photobleaching on cells expressing tagged dCas9 complexes. Ma et al. 2016 measured residence times of Cas9-guide RNA complexes by FRAP in living human cells. A perfect-match guide showed a residence time of 206 ± 4.5 minutes. A guide with a single mismatch at the -5 seed position showed a residence time of 1.4 ± 0.6 minutes — a 147-fold reduction. Ma H et al., J Cell Biol 214:529, 2016. PMC5004447. A perfectly matched guide RNA held Cas9 at its target for 206 minutes on average. A single mismatch at position -5 in the seed region dropped the dwell time to 1.4 minutes — a 147-fold reduction. At most off-target sites the dwell time is under one second.

runnable · js
output
Run the code to see its output here.
Kinetic discrimination — how dwell time separates on-target from off-target. Data from Ma et al. 2016 (J Cell Biol 214:529, PMC5004447). The threshold for HNH rotation is estimated at ~2 minutes.

Run this and the number that comes back — 147 — is the kinetic selectivity factor. The perfect-match complex holds Cas9 at the target for more than three hours; the seed-mismatch complex holds it for less than two minutes. The HNH domain, which requires a stable R-loop to complete its rotation, never gets the chance to fire at the seed-mismatch site. Specificity is not written into a list of off-limits addresses. It is enforced by how long the molecule stays put.

  1. A guide RNA has a mismatch at position 3 from the PAM. What most likely happens?

  2. What does Cas9 do at most NGG PAM sites in the genome?

  3. The double-strand break occurs at what position relative to the PAM?

Check your mental model of Cas9 specificity.

Off-targets: when the scissors slip

Perfect selectivity is a theoretical ideal. In practice, the number of off-target cuts a given guide RNA produces varies enormously — from essentially zero to over 150 detectable sites — depending on guide sequence design. The method that revealed this was Genome-wide Unbiased Identification of DSBs Evaluated by Sequencing — a method that tags double-strand breaks with an oligonucleotide and then maps all insertions genome-wide. Tsai et al. 2015 (Nat Biotechnol), PMC4320685., developed by Tsai et al. in 2015. By tagging every double-strand break in living cells with a short DNA oligonucleotide and then sequencing all insertion sites, GUIDE-seq revealed off-target cleavage events with indel mutation frequencies as low as 0.03%.

What GUIDE-seq found was sobering: wildtype SpCas9 targeting VEGFA generated 134 detectable off-target sites, with individual indel frequencies ranging from 0.03% to 60.1%. Good guide sequences had far fewer; poorly designed ones had more. The position of mismatches — seed versus PAM-distal — explained most of the variation.

Off-target sites detected by GUIDE-seq for a representative guide targeting VEGFA. Engineered high-fidelity Cas9 variants reduce off-target site counts by roughly 82–87% compared to wildtype SpCas9. Source: Frontiers 2023 off-target review (fbioe.2023.1143157).

The engineered high-fidelity variants — eSpCas9, SpCas9-HF1, hypaCas9, and others such as evoCas9 — all work by the same principle: weakening Cas9’s grip on PAM-distal DNA, so that only perfect-match R-loops achieve the stability needed for HNH firing. They do not touch the seed-region filter, which already works well; they tighten the stage 4 kinetic gate. The tradeoff is modest reductions in on-target efficiency for the hardest targets, but for clinical applications the extra specificity is worth it.

The clinical safety bar is high. The FDA’s approval of Casgevy required a post-marketing study specifically to assess long-term off-target editing risks, because the significance of any undetected low-frequency edits in patients’ hematopoietic stem cells won’t be known for years. The mechanism is safe enough to approve; the long-term epidemiology requires further study. Both things are true at once.

36 years: from accident to medicine

The path from CRISPR’s discovery to its first approved therapy covers 36 years and contains one of the cleaner examples in modern science of a phenomenon discovered by accident, ignored for nearly two decades, then recognized as adaptive immunity, then converted into a programmable tool, then deployed in human patients — each step requiring a different kind of scientist.

Event 1 of 8: 1 Jan 1987, Ishino 1987

1 Jan 1987

Ishino 1987

Yoshizumi Ishino discovers unusual tandem repeats in the E. coli iap gene while sequencing a gene for alkaline phosphatase isozyme conversion. He notes them as curious and unexplained. Journal of Bacteriology 169:5429.

1 Jun 2001

CRISPR named

Francisco Mojica and Ruud Jansen propose the acronym CRISPR — Clustered Regularly Interspaced Short Palindromic Repeats — to unify the numerous acronyms then in use for similar repeat sequences across species. (Wikipedia: CRISPR)

1 Feb 2005

Mojica 2005

Francisco Mojica proposes that CRISPR spacer sequences derive from bacteriophage and plasmid DNA and function as a microbial adaptive immune memory. Journal of Molecular Evolution 60:174.

23 Mar 2007

Barrangou 2007

Barrangou et al. prove CRISPR immunity experimentally: adding or deleting spacers matching phage sequences directly controls Streptococcus thermophilus resistance. Science 315:1709.

28 Jun 2012

Jinek 2012

Jinek, Chylinski, Fonfara, Hauer, Doudna, and Charpentier demonstrate programmable Cas9 and engineer the single guide RNA (sgRNA), simplifying the system to a two-component tool. Science 337:816.

27 Feb 2014

Crystal structure

Nishimasu et al. solve Cas9 in complex with sgRNA and target DNA at 2.5 angstrom resolution, revealing the bilobed architecture, PAM-interacting domain, HNH and RuvC nuclease domains, and the seed region mechanism. Cell 156:935.

7 Oct 2020

Nobel Prize 2020

The Royal Swedish Academy of Sciences awards the Nobel Prize in Chemistry to Emmanuelle Charpentier and Jennifer A. Doudna for the development of a method for genome editing.

8 Dec 2023

Casgevy approved

FDA approves Casgevy (exagamglogene autotemcel) for sickle cell disease — the first CRISPR-based gene therapy approved in the United States. December 8, 2023.

CRISPR: 36 years from a mysterious repeat sequence to an approved gene therapy.

Start with the accident. In 1987, Yoshizumi Ishino and colleagues at Osaka University were sequencing the E. coli iap gene — the gene responsible for converting one form of alkaline phosphatase into another — when they noticed something odd downstream of the coding sequence: five copies of a 29-nucleotide sequence, nearly identical to each other, separated by 32-nucleotide spacers that had no obvious resemblance to anything in the databases. Ishino Y, Shinagawa H, Makino K, Amemura M, Nakata A. Nucleotide sequence of the iap gene, responsible for alkaline phosphatase isozyme conversion in Escherichia coli. J Bacteriology 169:5429, 1987. PubMed 3316184. The repeats were mentioned in passing as unusual; no function was proposed. They mentioned the repeats in the discussion section as unusual and moved on, and for years the finding drew little attention.

Similar repeats were turning up in archaea as Francisco Mojica was studying the salt-tolerant organism Haloferax mediterranei in the early 1990s. Mojica spent years cataloguing these interspersed palindromic sequences, and in 2003 submitted a paper to several high-profile journals proposing that the spacer sequences were derived from bacteriophages — that bacteria were storing genetic memories of past viral infections and using them as a defense system. The paper was repeatedly rejected by major journals before finally appearing in the Journal of Molecular Evolution in February 2005. Mojica FJ, Diez-Villasenor C, Garcia-Martinez J, Soria E. Intervening sequences of regularly spaced prokaryotic repeats derive from foreign genetic elements. J Mol Evol 60:174, 2005. PubMed 15791728. The hypothesis was correct. The spacers matched phage sequences. The repeats were an immune system. Reviewers thought the idea was not significant enough to publish in a major journal.

Two years later, Rodolphe Barrangou and colleagues at Danisco — a food-science company with practical interest in how dairy bacteria resist phage infection — published the experimental proof in Science. Barrangou R, Fremaux C, Deveau H, Richards M, Boyaval P, Moineau S, Romero DA, Horvath P. CRISPR provides acquired resistance against viruses in prokaryotes. Science 315:1709, 2007. Penn State pure.psu.edu. They showed that Streptococcus thermophilus cells that acquired new spacers from phage DNA became resistant to those specific phages — and cells whose spacers were deleted became susceptible again. The spacer sequence was the memory; the immunity was real. The mechanism was still unknown: who was doing the cutting?

The cutting machinery was identified in the years that followed. By 2011, Emmanuelle Charpentier’s laboratory had characterized a small RNA called tracrRNA as an essential component of the type II CRISPR system in Streptococcus pyogenes — a molecule that paired with the crRNA (the spacer-containing guide) and together directed the Cas9 protein to its target. Charpentier and Doudna began collaborating, and in June 2012 they published the result that unlocked everything.

Emmanuelle Charpentier and Jennifer A. Doudna have discovered one of gene technology’s sharpest tools: the CRISPR/Cas9 genetic scissors. Using these, researchers can change the DNA of animals, plants and microorganisms with extremely high precision.

Royal Swedish Academy of Sciences, Nobel Prize in Chemistry 2020 source

What Jinek, Chylinski, Fonfara, Hauer, Doudna, and Charpentier did in the 2012 Science paper was not merely characterize the system — they rebuilt it. They demonstrated that the two-RNA system (crRNA plus tracrRNA) could be fused into a single chimeric molecule, the The sgRNA (single guide RNA) invented in Jinek et al. 2012 fuses the crRNA spacer sequence with the tracrRNA scaffold into one continuous molecule, simplifying the system from two required RNA components to one. This engineering move made CRISPR-Cas9 accessible to any laboratory. Science 337:816, PMC6286148. (sgRNA), dramatically simplifying the system. They showed Cas9 could be programmed with different sgRNA sequences to cut different DNA targets in vitro. They showed that the PAM was required on the double-stranded DNA target but not on single-stranded DNA or RNA substrates. They identified that both the HNH and RuvC domains were required for complete double-strand cleavage, with each cutting one strand. It was published online on June 28, 2012.

The Nobel Prize followed eight years later — 2012 discovery, 2020 award — a notably swift recognition for a field-defining technology.

In the blood: how the first approved therapy uses the mechanism

Sickle cell disease is a single-letter mistake in the gene for beta-globin, one of the two protein chains that make up adult hemoglobin. The mutant chain — hemoglobin S (HbS) — polymerizes when deoxygenated, deforming red blood cells into rigid crescents that block small blood vessels, causing severe pain crises (vaso-occlusive crises, or VOCs), organ damage, and shortened life expectancy.

There is a second hemoglobin, mostly dormant in adults: fetal hemoglobin (HbF), which does its job during gestation and is then silenced after birth by a transcription factor called BCL11A. BCL11A is a transcription repressor that silences the gamma-globin genes encoding fetal hemoglobin (HbF) shortly after birth, as part of the hemoglobin switching program. Bauer et al. 2013 identified the erythroid-specific enhancer of BCL11A as the key regulatory region. Science, PMC4018826. People with naturally high HbF levels — due to hereditary polymorphisms — are substantially protected from sickle cell disease because HbF does not sickle and its presence in red blood cells dilutes and inhibits HbS polymerization. This suggested a therapeutic strategy: turn BCL11A down in red blood cell precursors, and HbF will come back.

The challenge was turning it down only in the right cell type. BCL11A serves important functions in the brain and immune system; knocking it out systemically would cause serious harm. The key was a regulatory element discovered by Bauer et al. in 2013: an erythroid-specific enhancer of BCL11A that contains a binding site for the transcription factor GATA1 and is active only in red blood cell precursors. Disrupt this enhancer and BCL11A expression drops in erythroid cells — but not in neurons or lymphocytes.

  1. A Cas9 ribonucleoprotein complex with a guide RNA targeting the GATA1 binding site in the BCL11A erythroid enhancer is delivered to the patient's hematopoietic stem cells ex vivo. The guide makes a precise DSB at the enhancer.

    Source: CADTH Reimbursement Review (NBK614940).

  2. NHEJ repair of the DSB introduces an indel that disrupts the GATA1 binding site. Without GATA1, BCL11A transcription in erythroid cells falls. The protein is depleted specifically in red blood cell precursors.

    Source: Bauer et al. 2013, PMC4018826.

  3. With BCL11A reduced, the gamma-globin genes — normally repressed — are de-repressed. Fetal hemoglobin (HbF) is produced in large quantities in the patient's red blood cells.

    Source: CADTH Reimbursement Review (NBK614940); Frangoul et al. 2021, NEJM.

  4. HbF fills red blood cells alongside HbS. HbF does not sickle and inhibits HbS polymerization when deoxygenated. Cells with sufficient HbF levels resist the rigid, crescent deformation that causes vessel occlusion.

    Source: Frangoul et al. 2021, NEJM.

  5. Red blood cells no longer block small vessels during deoxygenation. Severe pain crises — the defining clinical event of sickle cell disease — stop occurring in patients with sufficient HbF levels.

    Source: Frangoul et al. 2021 (NEJM); Wikipedia — Exagamglogene autotemcel; 29 of 31 evaluable patients VOC-free (93.5%).

The molecular chain linking CRISPR editing to sickle cell protection in Casgevy — from the genome edit to the patient outcome.

The process in a patient is ex vivo: stem cells are harvested from the patient’s bone marrow, edited in the laboratory, and re-infused after the patient receives conditioning chemotherapy to ablate the original bone marrow. The edited cells engraft in the marrow and produce edited red blood cells for the rest of the patient’s life — in principle. The CRISPR edits are in the patient’s own stem cells, not in a virus or foreign construct.

7 rows
CLIMB-SCD-121 trial outcomes — Casgevy (exagamglogene autotemcel) for sickle cell disease. Sources: Frangoul et al. 2021 (NEJM); Wikipedia — Exagamglogene autotemcel.
Patients treated (SCD)44
Evaluable for primary endpoint31
VOC-free for at least 12 consecutive months29 of 31 (93.5%)
CRISPR edit targetBCL11A erythroid enhancer (GATA1 binding site)
Off-target editing observed (in vitro)None detected
FDA approval date (SCD)December 8, 2023
Post-marketing safety requirementLong-term off-target monitoring study

Twenty-nine of 31 evaluable patients — 93.5% — went at least 12 consecutive months without a severe vaso-occlusive crisis, a condition that had previously sent most of them to hospital multiple times per year. The therapy is not yet a cure in the strict sense: the long-term durability of the edit in hematopoietic stem cells is not yet known, and the conditioning regimen required beforehand carries its own risks. But the molecular chain — from a bacterial immune system, to a Nobel-winning biochemistry demonstration, to an ex vivo gene therapy approved at the FDA — is intact and running in patients.

The off-target analysis in the CLIMB-SCD-121 trial showed no off-target editing in in vitro studies of the edited cells. The FDA required ongoing post-marketing surveillance because the edited cells will live in patients’ bone marrow for decades, and the significance of any low-frequency edits that current assays cannot detect remains unknown. This is the honest state of the safety science: the short-term data are encouraging; the long-term data are still being collected.

What the mechanism teaches

Here is the whole primer as one mental model. Cas9 does not navigate by address. It navigates by interrogation. It samples an estimated ~100–200 million candidate positions in the genome, running the same five-stage test at each one: PAM present? Seed region compatible? Full 20-bp R-loop stable? HNH domain repositioned? Both strands cleaved? Pass all five stages and you get a precise, permanent edit. Fail any one stage — even by a single mismatch in the wrong position — and the complex dissociates in milliseconds, leaving the DNA untouched.

The seed region is the sharpest filter: the first ten bases adjacent to the PAM are interrogated in a structural context that pre-expects a perfect match, and any mismatch there collapses the R-loop before the kinetic commitment to cleavage. The dwell time is the gate: 206 minutes at a perfect target versus 1.4 minutes at a single seed-mismatch site, a 147-fold kinetic fence that keeps the HNH domain from firing at the wrong place. And the off-target landscape is a property of guide design — not the technology — so that a well-chosen guide targeting a unique seed sequence can produce zero detectable off-target cuts, while a poorly chosen one might produce 134.

The path from Ishino’s accidental observation in 1987 to Casgevy’s approval in 2023 took 36 years and required the work of microbiologists, structural biologists, biochemists, and clinicians who mostly did not know they were building toward each other’s conclusions. The bacterial immune system did not care about genome editing. The guide RNA did not know it would eventually direct cuts in human hematopoietic stem cells. The mechanism was what it was — elegant, specific, based on molecular interrogation rather than navigation — and the applications followed from the mechanism.

The seed region, the R-loop, the HNH conformational gate, the 147-fold kinetic fence between on-target and off-target dwell — none of this was designed by a human engineer. It was optimized over millions of years by bacteria that needed to destroy phage DNA without destroying their own. The fact that it is now cutting human genes with therapeutic intent is a consequence of understanding the mechanism well enough to redirect it.