Base editing and prime editing: precision without breaks

Standard CRISPR/Cas9 produces a double-strand break, which triggers cellular repair. The repair options — NHEJ (error-prone, useful for disruption) and HDR (precise but inefficient and largely restricted to dividing cells) — leave a gap: many therapeutic edits would be single-base corrections, where HDR-efficiency is too low and NHEJ would scramble the sequence.

Base editors address the gap by avoiding double-strand breaks entirely. They use a catalytically-impaired Cas9 (a 'nickase' that cuts only one strand, or a fully dead 'dCas9' that doesn't cut) fused to a deaminase enzyme — a protein that chemically changes one DNA base into another.

The principle: bring Cas9 to a target site using the standard sgRNA, expose a small window of single-stranded DNA where the deaminase can act, and let the deaminase chemically convert the base. The cell's normal DNA repair completes the conversion across both strands, producing a permanent single-base change without a double-strand break.

The first base editors were developed by David Liu's laboratory in 2016 (cytosine base editors) and 2017 (adenine base editors). Both have moved into clinical trials within several years of their development.

Two families of base editors are in clinical and preclinical use, each producing a different chemical change.

Cytosine base editors (CBE). A cytidine deaminase converts cytosine (C) to uracil (U). DNA repair then treats U as if it were T, so the net change after a round of replication is C → T on the edited strand (and G → A on the complementary strand).

The most-used cytidine deaminases are APOBEC1 and engineered variants. The chemistry is established — APOBEC family deaminases are part of normal cellular antiviral defense.

Adenine base editors (ABE). An adenine deaminase converts adenine (A) to inosine (I). DNA repair treats I as if it were G, so the net change is A → G on the edited strand (and T → C on the complementary strand).

Naturally occurring adenine deaminases that act on DNA do not exist; the field engineered one (TadA, originally an RNA-editing enzyme) by directed evolution to act on single-stranded DNA. The engineered TadA variants (TadA-8e, TadA-CD) are the basis of all clinical ABE candidates.

What this covers and doesn't. The four conversions (C→T, G→A, A→G, T→C) are called transitions (within the purine class A/G or within the pyrimidine class C/T). Together they cover roughly 30% of pathogenic single-nucleotide variants catalogued in disease databases. The other 70% — the eight possible transversions (across classes, e.g., C→A or A→T) — are not addressable by these base editors. Engineered editors targeting some transversions exist (C→G base editors) but are less mature.

A base editor exposes a small region of single-stranded DNA (often 5–8 nucleotides wide) where the deaminase can act. Every cytosine (for CBE) or adenine (for ABE) within that window is potentially editable.

The consequence: if the target has multiple C's or A's near each other, the editor may edit bystanders along with the intended base. A target with multiple C's in the window can come out as a mixture of edits — some cells get only the intended C→T, some get the intended plus a bystander C→T, some get only the bystander.

Mitigations:

• Narrower-window editors — engineered variants that activate only over 2–3 nucleotides, sharply reducing bystander risk.
• Target selection — choose a guide RNA so that the intended base is in the active window and other editable bases are not.
• Outcome analysis — assess all possible edited outcomes (intended, bystander, mixed) by next-generation sequencing of the target region in treated cells.

For therapeutic applications, the requirement is usually that the intended edit dominates and the bystander edits are either silent (no protein change), present-but-tolerated (the bystander does not cause disease), or sufficiently rare. This is a target-by-target analysis that conditions whether a given disease is base-editable in practice.

Base editors share Cas9's standard genome-DNA off-target landscape (the editor can be brought to off-target sites by guide mismatches and edit cytosines or adenines there). But base editors introduce two additional off-target categories that standard CRISPR does not have.

Cas9-independent DNA off-targets. The deaminase domain can act on single-stranded DNA produced during normal cellular processes (transcription, replication) at sites the editor was never directed to. CBEs in particular have shown low-level genome-wide deamination of cytosines independent of guide RNA in early variants. Engineered versions with weaker deaminase activity have reduced this substantially.

RNA off-targets. Some deaminases also edit RNA. APOBEC-based CBEs and TadA-based ABEs both showed RNA off-target editing in early generations. Engineered variants with deaminase mutations that reduce RNA activity (e.g., TadA-CD, SECURE-BE3) have largely addressed this.

Quantifying the off-target landscape involves whole-genome and whole-transcriptome sequencing of edited cells, with the deep coverage needed to find rare events. The reported off-target rates of current-generation base editors at therapeutic doses are low — often below the detection limit of standard sequencing — but the testing standards continue to evolve as more therapies enter clinical trials.

The structural lesson: every editing technology trades off scope, efficiency, and specificity differently. Base editors gain efficiency and avoid DSBs but inherit the chemistry-specific off-target risks of their deaminase component.

Prime editing, introduced by Liu's laboratory in 2019, takes a different approach. The editor is a Cas9 nickase fused to a reverse transcriptase (RT), an enzyme that synthesizes DNA from an RNA template.

The sgRNA is replaced by a prime editing guide RNA (pegRNA), which encodes:

• A standard spacer (to direct Cas9 to the target),
• A 'primer binding site' that matches the cut DNA on one strand,
• An RNA template encoding the desired edit.

The mechanics:

1. The Cas9 nickase cuts one DNA strand at the target.
2. The cut strand's 3' end base-pairs with the primer-binding-site portion of the pegRNA.
3. The RT extends the 3' end using the pegRNA's edit-encoding template as a template — writing the desired edit into the DNA.
4. Cellular repair resolves the resulting branched structure, propagating the edit to the complementary strand.

The scope: prime editors can perform all 12 base-to-base transitions and transversions, plus small insertions and deletions (up to ~50 nucleotides), all without a double-strand break. This is a substantial scope advantage over base editors (which cover only the four transitions) and over HDR (which is inefficient).

The trade-off is efficiency. Prime editing rates are typically 5–50% in cell lines and lower in primary cells — better than HDR but worse than well-optimized base editors. The technology is younger and is still being engineered for higher efficiency and broader cell-type applicability.

The structural landscape of programmable genome editing in 2026:

The practical decision-making for a therapeutic edit:

• Disrupting a gene — Cas9 + NHEJ. Established, efficient.
• C→T or A→G correction — base editor of the appropriate flavor. Higher efficiency than HDR, no DSB.
• Other single-base substitution or short indel correction — prime editor. Broader scope at the cost of lower efficiency.
• Larger insertions or replacements — HDR (when efficient enough), or emerging integrase-based systems (e.g., PASTE, paste integrase systems for kilobase-scale insertions).

The choice depends on the specific edit needed, the cells involved, and the off-target tolerance of the application. Different therapeutic programs pick different editors based on this profile.

Multiple base-editing and prime-editing therapeutic candidates are in clinical trials as of writing. Examples representative of the structural landscape:

• Base editing in cardiovascular disease. Adenine base editor candidates targeting PCSK9 (encoding a protein that regulates LDL cholesterol) aim to permanently lower LDL by inactivating a single allele of the gene. Trials with in vivo LNP delivery to the liver are running, with reported substantial and persistent LDL reductions in early data.
• Base editing in hematopoietic disease. Ex vivo adenine base editing of patient stem cells for sickle-cell disease (an alternative approach to Casgevy's NHEJ strategy), targeting the same fetal-hemoglobin induction pathway with a base edit at a different regulatory site.
• Base editing in T-cell engineering. Multiplex base editing of donor T cells to produce allogeneic CAR-T therapies that evade host rejection — disrupting endogenous T-cell receptors and immune-recognition genes with multiple simultaneous edits.
• Prime editing in monogenic liver disease. Candidates correcting specific point mutations causing inherited metabolic disorders.

The field's clinical trajectory has been driven by:

• In vivo LNP delivery to liver working reliably enough to support several candidate programs.
• Ex vivo editing of HSCs maturing as a manufacturing platform.
• Engineered editor variants with successively better profiles — narrower windows, lower off-target activity, higher efficiency — feeding new generations of candidates approximately every 18–24 months.

What is not yet broadly demonstrated: in vivo editing of CNS, muscle, or many other tissues; large-scale precise replacements via HDR or integrase systems; long-term safety data spanning a decade or more in treated patients.

Three structural points for the cursus.

• Editing has a scope-vs-efficiency frontier. Cas9 + NHEJ is the highest efficiency but only disrupts. Base editing is high-efficiency for the four transition substitutions. Prime editing has broader scope but lower efficiency. HDR is the broadest in principle but the most efficiency-constrained in practice. New technologies extend the frontier but don't collapse the trade-off.
• Off-target effects must be evaluated per editor and per target. Each editor has its own off-target categories (DNA, RNA, bystander); the bar for clinical use is empirical assessment in the relevant cells, not extrapolation from one technology to another.
• Delivery and editor choice interact. In vivo LNP delivery to liver supports certain editor families well; ex vivo HSC editing supports others. The editor and the delivery vehicle are picked together based on the target tissue and the edit type.

The next lesson moves out of genome editing into a different domain of modern molecular medicine: GLP-1 receptor agonists, and the broader question of how a single hormone-receptor axis can produce such large and varied physiological effects.