Q4 2024: Verve Therapeutics dosed the first patient with VERVE-102 — the second iteration of their in vivo PCSK9 base-editing program, after VERVE-101 hit a liver enzyme signal in 2023. VERVE-102 is the most-watched gene-editing trial in the world. Can BioMate recover the VERVE-102 architecture from "loss-of-function PCSK9 for hypercholesterolemia" — in 5 phases? Yes.

Why VERVE matters for the field

Pharma spent 30 years on PCSK9 inhibition. Repatha (Amgen, 2015) and Praluent (Sanofi/Regeneron, 2015) brought monoclonal antibodies to market — but both require biweekly injection. Inclisiran (Novartis, 2021) brought siRNA at twice-yearly dosing. Verve's pitch: one shot, for life.

The mechanism: an adenine base editor (ABE8.8) + PCSK9 guide RNA + liver-tropic LNP. The ABE makes a single A•T→G•C edit at the exon 1/intron 1 splice donor, disrupting the canonical GT dinucleotide (GT→GC). PCSK9 pre-mRNA fails to splice, the transcript is degraded by nonsense-mediated decay, and the protein is never expressed. LDL drops permanently. One infusion — done.

VERVE-101 (2022) saw two patients with elevated liver function tests in Phase 1b. Verve paused and pivoted to VERVE-102 — a reformulated LNP designed to reduce hepatotoxicity. That reformulation decision is what makes VERVE-102 interesting, and it's exactly what BioMate's CMC flag surfaces.

What changed between VERVE-101 and VERVE-102

ProgramLNP platformEditorClinical update
VERVE-101Acuitas ALC-0315 ionizable LNPABE8e (ABE8.8) — splice donorPaused — LFT elevation Nov 2023
VERVE-102Reformulated hepatotropic LNP (lower hepatotoxicity)ABE8e (ABE8.8) — splice donorFirst patient dosed Q4 2024
VERVE-201ANGPTL3 target (same platform)ABE8eIND filed 2024

The 5-phase information flow

In Vivo Base Editing — VERVE-102 Reconstruction (PCSK9 LoF) A Atlas PCSK9 liver-restricted B ABE8.8 adenine base editor C gRNA splice donor NGG PAM D LNP-IV hepatotropic ionizable lipid E CMC Flag LFT signal VERVE-102 Phase C — ABE8.8 Splice-Donor Edit (PCSK9 exon1/intron1 junction) Before Edit EXON 1 GT ↑ A on antisense INTRON 1 Canonical splice donor active PCSK9 pre-mRNA splices PCSK9 protein expressed ↑ hypercholesterolemia ABE8.8 A→G ABE8.8 Deamination ABE 8.8 A deam. G inosine read as G Antisense strand A•T → G•C (permanent) 20-nt protospacer · NGG PAM After Edit EXON 1 GC ✗ broken INTRON 1 NMD mRNA degraded PCSK9 eliminated ↓↓ LDL-C ↓↓ ABE8.8 + PCSK9 exon1/intron1 splice-donor gRNA + hepatotropic LNP-IV + LFT flag → VERVE-102 · matched dosed program Musunuru et al. Nature 2021;593:429–434 · Lee et al. Circulation 2023 · Same pipeline: CFTR → "wrong platform, need inhaled/AAV" Outputs: editor_selection.json · grna_design.json · off_target_screen.json · delivery_platform.json · cmc_flags.json
Figure 1 — In vivo base editing pipeline: 5 phases from query to IND-ready design. Phase B selects ABE8.8 (adenine base editor) for the PCSK9 splice-donor strategy — the mechanism VERVE-101/102 actually uses (Musunuru et al., Nature 2021; Lee et al., Circulation 2023). Phase E surfaces the VERVE-101 LFT signal — the CMC reformulation flag that defines VERVE-102. The same pipeline applied to CFTR flags "wrong delivery platform" because lung tissue requires inhaled or AAV delivery, not LNP-IV.

The gap between fluent and grounded

The Verve / Beam / Prime base-editing literature is fragmented across Liu lab Nature papers (2016–2017 original CBE/ABE chemistry), Verve corporate releases (2022, 2023 program updates), FDA briefing documents, and investor decks from Editas, Beam, Prime, and Intellia.

A reasoning LLM may correctly select ABE as the editor class and identify the splice-donor strategy — these are knowable from the published literature. What it won't do is flag the VERVE-101 LFT signal as a CMC class effect for hepatic LNP-IV, because that signal lives in a corporate press release from November 2023 and the LLM's knowledge of it is diffuse and unverified. The difference between "ABE + splice donor" and "ABE + splice donor + VERVE-101 LFT class-effect flag" is the difference between a design and an IND-ready design.

BioMate grounds Phase E against a curated knowledge base that includes corporate program updates. When the workflow runs for PCSK9 loss-of-function, the agent reads VERVE-101's 2023 pause and emits the reformulation flag as a structured CMC note. That's the gap between fluent and grounded.

"VERVE-101's failure cost roughly $200M to learn. The next program shouldn't have to pay that tuition."

Why it generalizes

The same 5-phase pipeline — different inputs, different answers:

  • ANGPTL3 → VERVE-201 design (also hepatocyte-restricted, ABE8e, same LNP platform)
  • HMGCR → competitive flag: "rosuvastatin already covers this at low cost; LDL-lowering bar for a new modality is high"
  • CFTR → delivery rejection: "lung tissue, LNP-IV is wrong platform — inhaled or AAV required"
  • DMD → "muscle tissue, LNP-IV has poor muscle tropism — systemic delivery unsolved"

The pipeline's intelligence isn't in knowing each program. It's in applying the same logic — tissue restriction → editor class → delivery feasibility → CMC precedent — to any target.

Try it yourself

Base editing strategy for PCSK9 loss of function in hypercholesterolemia

→ biomate.ai · 30 seconds · 5 phases · VERVE-102 architecture recovered.

Further reading: Komor et al. 2016, CBE original chemistry — Nature (PMC); Gaudelli et al. 2017, ABE original chemistry — Nature (PMC); Verve Therapeutics pipeline (vervetx.com); FDA gene therapy guideline (FDA.gov).

What this means for gene editing programs

A program like VERVE-101 fails at Phase 1b because of a CMC signal that was partially predictable from prior LNP hepatotoxicity literature. A grounded pipeline that surfaces that precedent at design time — before the IND is written — turns a $200M lesson into a one-line CMC flag in a JSON output. That's the purpose of grounding, and the reason general-purpose LLMs can't substitute for it.