Base editing

Base editing converts one DNA letter directly into another at a chosen position, using an enzyme that chemically modifies a base rather than a nuclease that severs the chromosome. It was built to solve a specific weakness of CRISPR–Cas9: most disease-causing mutations are single-letter substitutions, and correcting a single letter with a double-strand break is inefficient and messy. Base editors handle four of the twelve possible substitutions, which covers a large fraction of known pathogenic point mutations but by no means all of them.

§How it works

A base editor is a fusion protein. The targeting half is a catalytically impaired Cas9 that binds DNA under guide-RNA control but cuts at most one strand. The working half is a deaminase, an enzyme that strips an amine group from a base and changes what the cell's polymerases read.

Cytosine base editors, described in 2016, use a cytidine deaminase to convert cytosine to uracil.¹ Uracil is read as thymine during replication, so a C•G pair becomes T•A. Because cells actively excise uracil from DNA, the editor also carries a uracil glycosylase inhibitor to block that repair, and nicks the opposite strand so the cell rebuilds it using the edited strand as template.

Adenine base editors, described in 2017, had no natural starting enzyme: no known deaminase acts on adenine in DNA. The solution was directed evolution of a bacterial tRNA adenine deaminase until it accepted single-stranded DNA.² The product, inosine, is read as guanine, converting A•T to G•C.

Both editors act on the short stretch of DNA displaced when Cas9 unwinds the helix, giving an editing window of roughly five nucleotides within the protospacer. The window is the source of the method's characteristic error: if more than one editable base falls inside it, all of them may be converted. These bystander edits are usually silent or tolerated, but they are not always, and guide design often becomes a search for a position that puts only the intended base in range.

§Development history

Both classes came out of David Liu's laboratory, and both were engineered rather than discovered: the adenine editor in particular required many rounds of laboratory evolution before it worked on DNA at all. Successive generations improved on-target activity, narrowed the editing window, and reduced the deaminase's promiscuity. Faster variants shortened the exposure time needed, which matters because most unintended editing accumulates with duration of expression.

The first use in a person came in 2022, when a team at Great Ormond Street Hospital in London treated a teenager with relapsed T-cell acute lymphoblastic leukaemia using donor T cells in which three genes had been disabled by base editing, converting them into a CAR-T product that would not attack itself.³ The editing was done outside the body, which sidesteps the delivery problem entirely.

Editing inside the body started the same year. In July 2022 Verve Therapeutics announced it had infused VERVE-101, an adenine editor aimed at the cholesterol-regulating gene PCSK9 in liver cells, into a patient in New Zealand with heterozygous familial hypercholesterolaemia, which Verve described as the first in vivo base-editing dose in a person.⁴ That trial, Heart-1, paused enrolment in April 2024 after the sixth patient dosed at 0.45 mg/kg developed a grade 3 rise in liver enzymes and grade 3 thrombocytopenia within four days of the infusion. Both resolved within days and neither caused symptoms.⁵ The company's own investigation, including animal studies using a version of the drug with a guide RNA designed not to edit, concluded that the lipid nanoparticle was likely the primary driver of the abnormalities.⁶

§Current state

As of November 2026, base editing is in clinical trials rather than approved practice, and the trials divide into two families.

Ex vivo programmes edit cells outside the body and return them. Beam Therapeutics' sickle cell candidate, ristoglogene autogetemcel, edits the HBG1 and HBG2 promoters in a patient's own blood stem cells so that the repressor BCL11A can no longer bind there, switching production from sickle haemoglobin to fetal haemoglobin. An unplanned interim analysis of the phase 1-2 BEACON trial, published in 2026, covered 31 patients followed for a mean of 6.6 months. Fetal haemoglobin averaged more than 60% of total haemoglobin at six months among the 13 patients assessed then, no investigator-reported severe vaso-occlusive crisis occurred more than 60 days after the last transfusion, 27 of the 31 patients had an adverse event of grade 3 or worse, and one died of idiopathic pneumonia syndrome.⁷ That is a different route to the same physiological end as Casgevy, which disrupts an erythroid enhancer of the repressor gene BCL11A instead. Both approaches share the burden that limits autologous stem cell therapy generally: patients need conditioning chemotherapy to clear marrow space, and that conditioning carries real mortality and infertility risk independent of the edit. Beam has said it expects to apply for United States approval of the therapy as early as the end of 2026.⁸

In vivo programmes deliver the editor systemically, almost always to the liver, using Lipid nanoparticles. Verve's successor candidate VERVE-102 carries the same adenine editor and PCSK9 guide RNA as VERVE-101 inside a different particle, one built around a different ionisable lipid and a sugar tag, N-acetylgalactosamine, that lets it enter liver cells through the asialoglycoprotein receptor as well as the LDL receptor.⁵ Interim results from the phase 1 Heart-2 trial, published in the New England Journal of Medicine in 2026, cover 35 adults with heterozygous familial hypercholesterolaemia or premature coronary disease, each given a single infusion at one of six dose levels. Mean blood PCSK9 fell by 51% at the lowest dose and 88% at the highest; mean LDL cholesterol fell by 9% and 62% respectively, an absolute drop of 78 mg/dl at the top dose. The reductions appeared to hold through follow-up, which had reached at least a year in 15 of the 35 participants.⁹ Eli Lilly announced the purchase of Verve in June 2025 for about $1.0 billion in cash plus a contingent right to up to $3.00 a share that pays out only if a first patient is dosed in a United States phase 3 trial of VERVE-102 within ten years of closing, and completed it the following month.¹⁰¹¹ A randomised, placebo-controlled phase 2 trial with a planned 175 patients and LDL cholesterol at day 90 as its primary endpoint was registered in September 2026 by Verve, now a Lilly subsidiary, and was listed that month as not yet recruiting.¹²

One liver programme corrects a disease-causing point mutation rather than disrupting a gene: Beam's BEAM-302, which repairs the single G-to-A change in SERPINA1 behind the severe PiZZ form of alpha-1 antitrypsin deficiency. At a late-breaking session of the European Respiratory Society congress in September 2026, investigators reported that patients given the 60 mg dose went from blood alpha-1 antitrypsin of about 5 µM at baseline to a steady-state mean near 14 µM, above the 11 µM concentration conventionally treated as protective, with more than 90% of the circulating protein in the corrected form and the mutant form down by 84%. Some patients had been followed to 18 months. Adverse events were mostly mild to moderate infusion reactions and grade 1 liver enzyme rises; one patient had a grade 3 enzyme rise that resolved without treatment while bilirubin stayed normal.¹³ Those are sponsor-funded conference data rather than a peer-reviewed paper. A cohort intended to support a potential accelerated approval filing in the United States began dosing in July 2026.⁸

What "one-time" has and has not been shown to mean

Both liver programmes are described as single-dose and permanent, and in the cells that take up an edit that is chemically true. The human record behind the durability claim is short: in the published and presented data for VERVE-102 and BEAM-302, about 18 months at most and a year or more in a minority of dosed patients, and nothing yet on the decades over which a cholesterol drug is supposed to work. An edit also cannot be stopped if it turns out to be the wrong trade against a cheap daily tablet, which is the asymmetry that makes the irreversibility of the method a safety question rather than a selling point.

The bespoke-therapy case

In 2025 clinicians at the Children's Hospital of Philadelphia treated an infant with a severe urea-cycle disorder using a base editor designed for his specific mutation and delivered by lipid nanoparticle, from diagnosis to dosing in roughly six months.¹⁴ The case is the clearest existing demonstration that an editor can be designed for one person, and the clearest illustration that regulatory and manufacturing pathways for one-patient medicines do not yet exist.

§Limitations

The four accessible substitutions are the transitions: C to T, G to A, A to G, T to C. The eight transversions require different chemistry. Glycosylase-based editors that produce C-to-G changes exist but are less efficient and less clean, and for most transversions Prime editing remains the practical option.

Targeting is still constrained by the PAM requirement inherited from Cas9, which limits how the editing window can be positioned over a given base. Cargo size is a problem for viral delivery: a base editor plus guide RNA exceeds the packaging capacity of a single adeno-associated virus, forcing split-vector designs. And systemic delivery beyond liver and haematopoietic cells remains largely unsolved, which is the same constraint that limits Targeted drug delivery generally and that the particles being tested for In vivo CAR-T therapy are designed to get around.

§Risks

Base editors avoid double-strand breaks but introduce a distinct failure mode: deaminase activity that does not depend on the guide RNA at all. Cytosine editors have been shown to cause genome-wide off-target single-nucleotide changes in mouse embryos and rice through Cas9-independent deamination of transiently exposed single-stranded DNA.¹⁵ Both editor classes can also edit cellular RNA, since the deaminases do not distinguish substrates as sharply as intended.¹⁶ Engineered variants reduce both effects but do not abolish them, and detecting scattered single-base changes is harder than detecting the indels left by nuclease off-target activity, because there is no cut site to look near.

Bystander edits raise a subtler issue for Somatic gene therapy: a silent bystander change in one cell type may not be silent in another, and the tissue is a mosaic of edit patterns rather than a uniform genotype.

In the liver trials so far, the clearest harms have come from the delivery vehicle rather than from the editing chemistry. The laboratory abnormalities that paused Heart-1 were traced to its lipid nanoparticle, and the redesigned particle still produced grade 1 or 2 infusion reactions in about a fifth of Heart-2 patients, along with transient liver enzyme elevations, even though patients were given a steroid the day before dosing and a steroid plus antihistamines on the day. Commenting on the Heart-2 results, Daniel Soffer of the University of Pennsylvania, a lipid specialist and past president of the National Lipid Association, noted that a particle which enters cells through the LDL receptor will reach tissues wherever that receptor is expressed, and that off-target risks will take years to study; Verve's chief medical officer, Scott Vafai, replied that preclinical off-target experiments had found no issues and that further clinical development is meant to establish safety.¹⁷

For heritable use, the apparent tidiness of base editing has been used as an argument that Human germline editing is now safe enough to attempt. Most of the field rejects that inference. Embryo editing still produces mosaicism, the Cas9-independent off-target mechanism operates during the exact developmental window in question, and no whole-genome assay can certify the absence of scattered point mutations in a single embryo without destroying it. The regulatory posture that followed the He Jiankui affair has not shifted, and the frameworks surveyed in Governance of human genome editing treat editor chemistry as largely beside the point.

§Outlook

The near-term question is whether permanent single-dose edits to common risk genes, such as a lifetime cholesterol reduction from one infusion, can clear a safety bar set for chronic disease rather than for fatal rare disease. That bar is high, because the comparator is a cheap daily pill and the edit is irreversible. The phase 2 cholesterol trial adds a placebo group, which the Verve and Beam trials reporting so far have not had,¹² and Heart-2 participants are to be followed for up to 15 years.¹⁷ The technical frontier is delivery to tissues other than liver, editors small enough for single-vector packaging, and reliable assays for scattered off-target deamination. The economic frontier, shared with Gene therapy for aging and every other one-time genetic medicine, is whether a treatment designed for one child can ever be paid for, a problem that runs directly into Access and inequality and that no current reimbursement system is built to solve.

§See also

• Prime editing
• CRISPR–Cas9
• Epigenome editing
• Off-target effects in genome editing
• Lipid nanoparticles
• Casgevy
• Somatic gene therapy
• Human germline editing

§References