Prime Editing Went from 0.8% to 49% Liver Efficiency in One Dose. Three Papers Solved Three Bottlenecks in a Single Week.
David Liu's lab published three Nature papers simultaneously — improving the guide RNA, the enzyme, and the delivery vehicle. Together they achieved a 63-fold improvement in mouse liver editing efficiency and cured phenylketonuria in a disease model with a single injection.
Zero point eight percent. That was it. The best anyone could manage with lipid nanoparticle–delivered prime editing in a living mouse liver, a number so low it barely registered above background noise and so discouraging that more than one lab privately questioned whether the approach would ever work outside a petri dish. The technology that could theoretically fix 89% of known disease-causing mutations was, in practice, trapped — by unstable RNA, poorly folding enzymes, and nanoparticles that couldn’t coordinate three molecular cargoes at once. The researchers at the Broad Institute had cataloged every failure mode. They just hadn’t solved all three at the same time.
Now they have. In three papers published within days of each other — two in Nature Biotechnology and one in Nature Nanotechnology — David Liu’s group reports a systematic dismantling of the barriers that kept prime editing out of the body. Their optimized system achieved 49% indel-free editing in bulk mouse liver from a single 2 mg/kg intravenous dose, a 63-fold improvement over their starting point. Along the way, they cured phenylketonuria in a humanized mouse model and introduced a next-generation enzyme family called PE8, designed not by directed evolution but by an AI protein design tool called ProteinMPNN.
The Three Bottlenecks
Prime editing was invented in 2019 and immediately recognized as the most versatile precision gene-editing tool ever built. Unlike CRISPR-Cas9 nucleases, which cut both DNA strands and rely on the cell’s own repair machinery (a chaotic process), or base editors, which can swap only certain single letters (C→T, A→G), prime editing can install any point mutation, small insertion, or small deletion without double-strand breaks. It works by nicking one strand, then using a reverse transcriptase enzyme to copy a corrected sequence from a guide RNA template directly into the genome.
Beautiful mechanism. Terrible delivery.
Getting prime editors to work outside a petri dish requires cramming three distinct molecular cargoes into a delivery vehicle small enough to slip through cell membranes and coordinated enough to arrive at the right stoichiometry: a massive 6,500-nucleotide mRNA encoding the prime editor protein, a ~150-nucleotide engineered guide RNA (epegRNA) that tells the editor where to cut and what to write, and a ~100-nucleotide nicking guide RNA (ngRNA) that nicks the opposite strand to bias repair toward the edit. Each cargo has different chemistry, different stability, and different optimal packaging conditions that interact in ways nobody had systematically untangled. When the whole system was loaded into lipid nanoparticles — the same fatty bubbles that delivered COVID mRNA vaccines to billions of arms — it barely functioned, with editing efficiency in mouse liver stuck below 1%.
Liu’s group identified three specific failure points. The guide RNA degraded too fast. The enzyme didn’t express well enough in mammalian cells. And the nanoparticle packaging wasn’t optimized for the stoichiometric complexity of the three-cargo system. Each paper attacks one bottleneck.
Paper 1: The Guide RNA Was Dying Too Fast
The first study, led by postdoctoral researchers Holt Sakai and Sarah Pierce, tackled the fragility of the prime editing guide RNA. Since 2021, nearly every lab in the field has protected their pegRNAs with the same 37-nucleotide tail, a pseudoknot motif called tevopreQ1, that shields the RNA’s vulnerable 3′ end from cellular exonucleases. It works adequately but nobody had ever systematically looked for something better, because nobody had a good way to test thousands of alternatives in parallel.
Sakai and Pierce built one. Their system, PE-PRISM (PE pooled reporter for identification of structured motifs), is a lentiviral library that pairs each candidate RNA motif with a matched genomic target site, allowing high-throughput sequencing to read out editing efficiency for every motif simultaneously across multiple prime editors and edit types. They threw 2,858 candidates at the wall — pseudoknots from viral frameshifting signals, motifs plucked from human signal recognition particles, plant pathogens, retrotransposon initiation elements, G-quadruplexes. Anything with the right structural geometry to shield a vulnerable RNA tail. Then two rounds of targeted mutagenesis evolved the best hits further.
Three winners emerged. They consistently outperformed tevopreQ1 across cell types and delivery methods, including lipid nanoparticle injection into living mice. The champion, eSBRMV1-A — derived, improbably, from soil-borne rye mosaic virus — boosted liver editing from 17% to 26% compared to the old motif when every other variable was held constant. Fifty-three percent more editing from swapping a few dozen nucleotides at the tail end of a guide RNA, which is the molecular equivalent of upgrading a car’s air filter and finding 50 extra horsepower under the hood that was always there, waiting for the airflow to stop being the limiting constraint.
Paper 2: The Nanoparticle Needed a Complete Overhaul
The second study, led by graduate student Allen Jiang and postdoctoral researcher Ana Cristian, went after the delivery vehicle itself. Their approach was methodical in a way that reveals how much room previous studies had left on the table. They varied one parameter at a time across five stages, measuring editing efficiency in mouse liver at every step, and the cumulative improvement was staggering. Pure optimization.
| Stage | Change Made | Avg. Liver Editing | Fold vs. Baseline |
|---|---|---|---|
| s0.1 (baseline) | Chemically modified pegRNA + PEmax | 0.8% | 1× |
| s0.2 | Switch to epegRNA (tevopreQ1 motif) | 3.8% | 4.8× |
| s1 | Switch from PEmax to PE6c enzyme | 11% | 14× |
| s2 | New eSBRMV1-A guide RNA motif | 26% | 33× |
| s3 | Optimized RNA ratios + HPLC-purified mRNA | 49% | 63× |
The final stage is the one that should keep competitors awake at night. When the team replaced their in-house mRNA with commercially produced, HPLC-purified mRNA from GenScript — same coding sequence, same protein, but cleaner manufacturing that removes double-stranded RNA contaminants from in vitro transcription — editing jumped from 26% to 49%. Not a new enzyme, not a new guide RNA — just better manufacturing. That 1.9-fold improvement from purification alone, a technical detail that sounds like a footnote in a methods section, nearly doubled the entire system’s performance and should worry any gene therapy company that doesn’t have pharmaceutical-grade mRNA production absolutely locked down.
The therapeutic payoff came in a humanized mouse model of phenylketonuria (PKU), a metabolic disease in which a mutation in the PAH gene prevents the liver from breaking down phenylalanine, causing brain damage if untreated. Current treatment: a lifelong restricted diet so austere that most patients struggle to follow it even with dietary counseling, and those who slip face progressive neurological damage from phenylalanine accumulation in the brain. The team’s optimized PE-LNP system achieved 15% bulk liver editing at the disease-causing R408W allele — well above the estimated 5–10% correction threshold needed for phenotypic rescue. Blood phenylalanine dropped to curative levels.
One injection. Disease corrected.
Paper 3: AI Redesigned the Enzyme, and It Got Better
The third paper introduces something conceptually striking, a collaboration between directed evolution and artificial intelligence that may reshape how all gene-editing enzymes are optimized going forward. Led by graduate students Allen Tao and Nicholas Krasnow alongside Sakai and Jiang, this study started from a frustrating observation: the lab-evolved PE6 enzymes that Liu’s group painstakingly developed through phage-assisted continuous evolution (PACE) were catalytically superior to their predecessors but expressed at 1.5 to 2-fold lower levels in mammalian cells. Evolution had optimized the engine at the cost of the chassis, a pattern familiar to anyone who has ever seen a phage-evolved enzyme gain catalytic speed while losing the ability to fold properly in the crowded, chaperone-dependent interior of a mammalian cell.
The team fed the PE6 enzyme structures into ProteinMPNN, a deep learning model trained on experimentally determined protein structures. ProteinMPNN redesigns amino acid sequences to optimize folding stability and solubility while preserving overall three-dimensional architecture. Crucially, the researchers constrained the redesign: any residue within 15 angstroms of the catalytic active site, or with greater than 25% evolutionary conservation, was locked in place and could not be changed. Everything else was fair game.
The results defied conventional protein engineering intuition. ProteinMPNN introduced up to 202 amino acid substitutions in a single enzyme — rewriting up to 40% of the protein sequence in ways that would make any biochemist who has spent months on a single point mutation weep with a complicated mixture of envy and terror. These are not the careful, one-residue-at-a-time substitutions of classical protein engineering. They are wholesale redesigns. Staggeringly radical ones. Yet 95% of the 174 synthesized designs retained functional editing activity, and 30% outperformed the already-optimized starting enzyme. The best Ec48-derived variant showed 2.1-fold average improvement across 700 pathogenic ClinVar mutations.
The team designated these new enzymes PE8a, PE8c, PE8d, and PE8max. They validated them across seven disease-relevant targets: Bloom syndrome, Crigler–Najjar syndrome, Tay–Sachs disease, sickle cell disease, and familial hypercholesterolemia — and across three clinically relevant delivery methods: mRNA electroporation in primary human T cells and blood stem cells, engineered virus-like particles, and lipid nanoparticle injection in mice. PE8 consistently outperformed PE6 and PE7. Every time.
What makes this paper matter beyond its immediate results is the method it validates: a general-purpose pipeline in which directed evolution generates the catalytic gain and AI-guided redesign rescues the biophysical properties that evolution predictably wrecks. Directed evolution through PACE excels at finding catalytic improvements but destroys biophysical properties. AI-guided redesign through ProteinMPNN excels at recovering stability and expression without touching catalysis. The two approaches are complementary at the amino acid level — PACE mutates residues near the active site, ProteinMPNN mutates residues far from it — and the combination produces enzymes that neither method alone could reach. True synergy. Sixteen thousand eight hundred prime editing experiments across 700 pathogenic variants confirmed the generality of that complementarity.
The Math on What This Unlocks
The ClinVar database catalogs roughly 120,000 pathogenic or likely pathogenic human genetic variants. Prime editing can theoretically address approximately 89% of known pathogenic point mutations — about 107,000 variants — because it handles all 12 possible single-nucleotide substitutions plus small insertions and deletions. Base editing, by comparison, is limited to C→T and A→G transitions, covering a smaller slice. But theoretical capability means nothing if the editor cannot reach the tissue at sufficient efficiency.
Most monogenic liver diseases have correction thresholds between 5% and 30% of hepatocytes for clinical benefit: phenylketonuria around 10%, familial hypercholesterolemia around 5%, ornithine transcarbamylase deficiency around 5–10%, alpha-1 antitrypsin deficiency around 5–10%. At 49% bulk liver editing from a single dose, this PE-LNP system clears the therapeutic bar for essentially every known monogenic liver disease with a defined correction threshold. And because LNPs naturally traffic to the liver after intravenous injection, coating themselves with apolipoprotein E and entering hepatocytes via LDL receptors, no exotic targeting is needed for this organ.
The cost story changes everything. Approved AAV gene therapies run $2.1 million (Novartis’s Zolgensma for spinal muscular atrophy) to $3.5 million (CSL Behring’s Hemgenix for hemophilia B) per patient. LNP-based mRNA manufacturing was industrialized during the COVID-19 pandemic — Moderna produced over 1 billion doses in 2021 alone. While a prime editing therapeutic would require far more rigorous per-patient quality controls than a vaccine, the underlying manufacturing platform is mature, scalable, and an order of magnitude cheaper per dose than viral vector production. Analysts at Regeneron and Intellia have estimated that LNP gene-editing therapeutics could reach manufacturing costs below $50,000 per dose at scale.
What’s Still Missing
None of this has been tested in primates. Mouse liver editing efficiencies do not always translate to larger animals — doses scale nonlinearly with body weight, immune responses differ, and hepatocyte turnover rates vary across species in ways that can dramatically change the persistence of edits over time. Verve Therapeutics’ base editor achieved approximately 60–70% editing at PCSK9 in non-human primate liver with a single LNP dose, which is comparable to prime editing’s 49% in mice, but the comparison is fragile because primate studies consistently require higher doses per kilogram to achieve equivalent outcomes, and nobody knows yet where prime editing’s dose-response curve flattens in larger animals.
The liver monopoly is the bigger constraint. LNPs go to the liver because physics and biology conspire to send them there — apolipoprotein E adsorption, fenestrated hepatic endothelium, LDL receptor–mediated endocytosis, all of which are features of liver biology that no amount of nanoparticle engineering has reliably overcome for other organs. For the roughly 70% of monogenic diseases that manifest outside the liver, these three papers change nothing. Brain, muscle, kidney, lung — each requires a different delivery solution. Alternative vehicles exist (engineered AAVs, extracellular vesicles, polyplexes) but each drags its own baggage of immunogenicity, limited packaging, and poor tissue penetration.
Immunogenicity remains an open question that no single-dose cure can entirely sidestep. The Cas9 protein originates from Streptococcus pyogenes, and pre-existing antibodies against it have been detected in 58–78% of humans in some studies. A single-dose cure sidesteps the redosing problem, but even a single dose must avoid neutralization. And off-target editing, while reported to be low in these studies, requires comprehensive genome-wide assessment in primate models before clinical translation can proceed. No shortcuts there.
The strongest case against excitement here is that base editing, a simpler and more efficient technology, has already reached patients. Verve dosed its first patient with an in vivo LNP base editor in 2023. Prime Medicine is running ex vivo prime editing trials for chronic granulomatous disease. The bar for in vivo prime editing isn’t just scientific; it’s commercial. Companies must justify the added complexity of prime editing over base editing for each indication, and for many single-nucleotide diseases, base editors already work. Prime editing’s unique value appears in mutations that base editors cannot touch: transversions, insertions, deletions, and multi-nucleotide changes. That is still tens of thousands of pathogenic variants, including the PKU mutation corrected in this study.
The Bottom Line
The timeline math is rough but directional. Base editing went from mouse proof-of-concept to first-in-human compassionate use in approximately three to four years. If in vivo prime editing follows a similar trajectory, the first patient could be dosed around 2029 or 2030. GLP toxicology studies, manufacturing scale-up, and IND-enabling work take two to three years; Phase 1 adds another one to two.
What these three papers establish is not a clinical-stage product. It is something more important: proof that the in vivo prime editing problem is an engineering problem, not a fundamental science problem. Every bottleneck that was identified has a solution. The guide RNA needed better armor; they found it in soil-borne rye mosaic virus pseudoknots. The enzyme needed better folding; an AI retrained on the Protein Data Bank provided it. The delivery vehicle needed to be treated as a multicomponent system, not a black box; systematic optimization and pharmaceutical-grade manufacturing delivered a 63-fold improvement.
For the roughly 300 million people worldwide living with a monogenic disease, the question was never whether gene editing could theoretically fix their mutation. The question was whether anyone could get an editor into their cells, at the right efficiency, through a scalable delivery system, in a single dose. For liver diseases, these papers answer yes. For everything else, the harder work is just beginning. But someone carrying a PAH R408W mutation who has spent their entire life avoiding protein should know: in a mouse that shares their disease, one injection was enough.
If you work in gene therapy and use in vitro–transcribed mRNA, the single most actionable finding here is the 1.9-fold improvement from HPLC purification — start purifying more aggressively, because your current pipeline is leaving half its potential on the floor. If you or someone you love has PKU, familial hypercholesterolemia, or another monogenic liver disease, watch for three things: primate data from Liu’s group (likely within 18 months), IND filings by Prime Medicine or a partner company, and Phase 1 recruitment announcements for PE-LNP liver therapeutics. No one can enroll in a clinical trial for this today — but the distance between today and that day just shortened considerably.