Science brief: a reading aid, unvalidated by any person. Not a prediction and not advice. Public data only.

Science brief: lab of David R. Liu

As of 2025-12-31. Target mode. Literature: prime editing. 35 abstracts read, 3 claims checked, 5 papers had no abstract and were not read.

Read this first.

• Each claim was extracted by a model from one abstract. A second, separate reader then checked it against the abstracts in this pack, and a third checked that reader. Every line below carries two exact quotes that code found in the sources.
• The corpus is each lab's three most recent last-author papers at four year-ends, abstracts only. "No independent support found" means none in this corpus. It never means not replicated.
• The order is a published sort rule (scirank-v1): highest evidence tier among verified claims, then independent supports, then fewer contradictions. It is not a score and it predicts nothing.
• Evidence tier comes from the extraction's model-system field. "Randomized" is a word rule on the abstract.

Measured, once, on planted errors (an upper bound on real ones):

• One pass over the retrieved abstracts found 110 of 121 hard planted misstatements and named the paper that shows each (91%, 95% interval 85 to 96). It flagged 2 of 423 unaltered claims (false alarms).
• By type: a flipped direction 58 of 66 (88%, 81 to 94); a number swapped for another in the same abstract 35 of 37 (95%, 83 to 100); animal and human swapped 17 of 18 (too few for a rate).
• Baselines on the same claims: a read with no literature flagged 52% and could name no paper; the Lens with a string rule found 2%; fan-out over chunks found 89%, not better than one pass.
• Planted errors are an upper bound on real ones: one word or number was changed and the true paper was in the pack. No person has checked a real case. Reader and extractor are both Claude models; scoring was by code against a key.
• Supported records, the framed read and the ranking are not measured by this test.

Supported records, graded in session by two blind agents (20 drawn from real briefs, both graders must agree): 20 of 20 held (83% to 100%). The same graders rejected 20 of 20 controls that code built to disagree with their own paper. Model-graded, not human-verified: docs/methodology/scibrief-grading.md.

Extraction: Measured accuracy on a fresh sample (agent readers, both must agree): output_type right 97% (92%–99%); model_system right 95% (90%–98%); scale right 97% (93%–99%); claim supported 98% (94%–99%); effect numbers supported 99% (92%–100%); limitations supported 93% (85%–97%).

Where a paper disagrees with a claim

None found in this pack.

Claims and their evidence

David R. Liu, Harvard University

Company, PI named on the filing: none found

• C02 AAV9-mediated in vivo prime editing corrects Atp1a3 D801N and E815K mutations in the CNS, rescuing phenotypes and extending lifespan in two alternating hemiplegia of childhood mouse models. Reported effect: Up to 48% DNA correction and 73% mRNA correction in bulk brain cortex; dramatic extension of animal lifespan.
  • animal; verified against its paper; no independent support found in this corpus; source: In vivo prime editing rescues alternating hemiplegia of childhood in mice (2025-07-21, P33)
  • Own paper:
    • claim: "up to 48% DNA correction and 73% mRNA correction in bulk brain cortex" / paper: "yielding up to 48% DNA correction and 73% mRNA correction in bulk brain cortex" — In vivo prime editing rescues alternating hemiplegia of childhood in mice (2025-07-21, P33). The paper reports the same AAV9 prime editing of Atp1a3 D801N and E815K in two AHC mouse models with identical correction figures. [R4060]
• C03 Base editors introducing G•C>A•T and A•T>G•C interruptions in CAG and GAA repeats reduce repeat expansion in patient cells and in mouse central nervous system. Reported effect: Significantly reduced repeat expansion in the CNS; the abstract gives no numeric effect size.
  • animal; verified against its paper; no independent support found in this corpus; source: Base editing of trinucleotide repeats that cause Huntington’s disease and Friedreich’s ataxia reduces somatic repeat expansions in patient cells and in mice (2025-05-26, P26)
  • Own paper:
    • claim: "Base editors introducing G•C>A•T and A•T>G•C interruptions in CAG and GAA repeats reduce repeat expansion" / paper: "significantly reduced repeat expansion in the central nervous system" — Base editing of trinucleotide repeats that cause Huntington’s disease and Friedreich’s ataxia reduces somatic repeat expansions in patient cells and in mice (2025-05-26, P26). The paper reports base editor interruptions at CAG and GAA repeats significantly reducing repeat expansion in mouse CNS and patient cells. [R4061]
• C01 Periplasmic phage-assisted continuous evolution (pPACE) can rapidly evolve disulfide-containing proteins, including YibK variants, an anti-GCN4 Ω-graft antibody, and trastuzumab variants, in the E. coli periplasm. Reported effect: Eight-fold increase in soluble expression of the Ω-graft antibody.
  • in vitro; verified against its paper; no independent support found in this corpus; source: Disulfide-compatible phage-assisted continuous evolution in the periplasmic space (2021-10-13, P29)
  • Own paper:
    • claim: "Eight-fold increase in soluble expression of the Ω-graft antibody" / paper: "leading to an eight-fold increase in soluble expression of the Ω-graft antibody" — Disulfide-compatible phage-assisted continuous evolution in the periplasmic space (2021-10-13, P29). The paper reports the same eight-fold increase in soluble expression of the Ω-graft antibody using pPACE. [R4059]

Framed read (unvalidated: one model's synthesis of the records above)

Claim that carries the most weight: C02. C02 is the only claim here that pairs an animal-tier result with a disease-relevant in vivo correction figure (up to 48% DNA correction and 73% mRNA correction in bulk brain cortex), so any case that this lab's editing work matters for therapy rests on it. The other two claims concern a protein-evolution method and a repeat-expansion approach, and neither supplies a correction figure of this kind.

What would have to hold:

• The correction figures in C02 come from the lab's own paper, and R4060 is the only record covering them; no independent paper in this pack addresses them.
• Bulk brain cortex measurements (R4060) reflect correction in the cell types that drive the phenotype; no record covers cell-type resolution.
• The claimed lifespan extension in C02 is covered by no record, since R4060 quotes only the correction figures.
• The C03 result (R4061) holds beyond the abstract's qualitative wording, as no record gives an effect size.
• The C01 result (R4059) extends beyond one protein; no record covers the other proteins named in the claim.

Questions a data room would have to answer:

• (C02) What are the per-animal DNA and mRNA correction values behind 'up to 48% DNA correction and 73% mRNA correction', how many animals per model, and what were the median and lowest values?
• (C02) What bystander edits, indels and off-target edits were measured in Atp1a3 and elsewhere in brain tissue, and what were the results for the D801N and E815K models separately?
• (C02) What survival data, group sizes and controls support the lifespan extension, and which phenotype rescues were scored blind?
• (C03) What numeric repeat-expansion reductions, in which mouse model and tissues, and at what vector dose, underlie 'significantly reduced repeat expansion in the central nervous system'?
• (C01) Does the eight-fold increase in soluble expression of the omega-graft antibody translate into improved binding or function, and was it seen in the trastuzumab variants too?

Cannot say:

• Whether any result here has been reproduced by a lab other than the originating one: all three records come from the claim's own paper and none is independent.
• Whether the editing results would carry to larger animals or humans, or whether delivery, dose and safety are adequate.
• How this lab compares with other prime-editing groups, since the records contain no comparator.
• Anything about IP position, company formation or team plans.

What was not shown

• Claims with no checked record: none.
• Records proposed: 3; refused by code (a quote not found, or a number of the reader's own): 0; removed by the checker: 0.
• Papers without an abstract, not read: 5.

Run record

• run: sci-860a38965b; reader arm: B1R; spec: sci-0.1; rank rule: scirank-v1
• reader prompt sha256: c04daba6743bf488eaeb3427cec02860a257d2c06ef9f58cf307195206620e03; model: claude-sonnet-5-5, medium effort
• extraction: claude-haiku-5-5, prompt extract_v0.3; no human has verified it