Capstan Therapeutics raised $165M in 2023. Umoja raised $210M before that. Kelonia $50M. The pitch is always the same: skip ex vivo CAR-T manufacturing, deliver the CAR construct via LNP straight into the patient's T cells in vivo. Drop the $475,000 price tag to $50,000. Hospital admission to outpatient. But $400M of funding still hasn't solved the hard part — the co-design.

Why ex vivo CAR-T can't scale

Kymriah (CD19 CAR-T, Novartis 2017) was the first FDA-approved cell therapy. Launch price: $475,000 per patient. Plus apheresis. Plus 14-day manufacturing turnaround while the patient bridges on chemotherapy. Seven years later, there are fewer than 15 fully-qualified ex vivo CAR-T sites in the US. The patients who reach them are a fraction of those who need them.

In vivo CAR-T fixes the bottleneck by eliminating it. A lipid nanoparticle encapsulates the CAR mRNA plus a T-cell-specific surface targeter. The patient receives an IV infusion. The LNP transfects their T cells in vivo. T cells express the CAR, clear the tumor, and persist as memory cells. No apheresis. No manufacturing window. No $475,000 price tag.

Three molecules that have to agree

But now three molecules must be co-designed simultaneously, not one:

  1. The CAR target — which tumor antigen to eliminate (e.g., CD19 for B-cell lymphoma)
  2. The LNP surface targeter — which T-cell marker guides LNP to T cells (anti-CD8, anti-CD3, anti-CD5)
  3. The CAR construct itself — scFv + hinge + transmembrane domain + costimulatory domain + CD3ζ

The deadliest failure mode is fratricide. Pick CD7 as both the CAR target and the LNP targeting moiety. The LNP transfects T cells. T cells express anti-CD7 CAR. The CAR-expressing T cells kill each other. The product self-destructs in vivo, invisibly, before efficacy data is ever collected. The 2024 oncology pipeline has programs that should have flagged fratricide earlier.

The 4-phase co-design pipeline

In Vivo CAR-T Co-Design — CD19 + Anti-CD8 LNP (Capstan CPTX2309 Architecture) A — Atlas Safety Check CD19 Expression (Tabula Sapiens · 26 tissues) B cells HIGH ✓ T cells low NK cells none Liver / CNS none B-cell restricted ✓ VIABLE B — LNP Tropism Selection anti-CD8 ✓ SELECTED → CD8+ T cells only · no off-target anti-CD3 → T + B cells broad; B-cell off-target risk anti-CD5 → broad T cells less selective anti-CD7 ⚠ FRATRICIDE RISK T cells express CD7 → workflow refuses 4 options scored · 1 selected · 1 flagged unsafe C — Fratricide Safety Check CAR target CD19 (B cells) LNP target anti-CD8 (T cells) NO OVERLAP ✓ SAFE — emit construct If CD7 target: circles overlap → REFUSED Phase D — CAR Construct (1,485 AA · FMC63 + CD8α + CD28 + CD3ζ · Capstan CPTX2309) FMC63 scFv CD19-binding VH-VL CD8α hinge/TM transmembrane CD28 costim CD3ζ ITAM signaling → FASTA emitted (1,485 AA) Capstan CPTX2309 architecture: anti-CD8 LNP + CD19 CAR · no fratricide · FMC63 construct emitted from biology alone CD7 swap → fratricide flag → pipeline refuses construct · outputs: atlas_check.json · lnp_tropism.json · cart_construct.fasta Query: "Design in vivo CAR-T for CD19 B-cell lymphoma with anti-CD8 LNP · check fratricide · emit construct FASTA"
Figure 1 — In vivo CAR-T co-design pipeline: 4 phases from query to FASTA. The fratricide check in Phase C cross-references the CAR target's cell-type expression against the LNP-transfected cell set — if they overlap, the pipeline refuses to emit a construct. Anti-CD8 LNP → CD8+ T cells; CD19 target → B cells: no overlap, safe.

The Capstan question

Capstan's lead program (CPTX2309) is anti-CD19 mRNA delivered by an anti-CD8 LNP for systemic lupus erythematosus. The Phase 1 readout is expected later in 2025.

Run BioMate on the Capstan thesis and the 4-phase pipeline reproduces it from scratch: CD19 safe to eliminate, anti-CD8 LNP delivers to CD8 T cells, no fratricide, FMC63-based construct assembled. The agent doesn't know about Capstan's IND. It infers the design from the biology.

What ChatGPT can't do here

You can ask any general-purpose LLM to "design an in vivo CAR-T for CD19 lymphoma with LNP delivery" and it will produce a fluent answer mentioning FMC63, CD3ζ, and lipid nanoparticles. What it won't do:

  • Check the CAR target's tissue expression against an actual single-cell atlas across 26 tissues
  • Flag the fratricide risk if you swap CD19 for CD5 or CD7
  • Emit a FASTA sequence with the correct domain order and exact amino acid count
  • Refuse to assemble if the safety check fails — and explain why

BioMate's pipeline does all four, and persists each decision to an auditable JSON file. When the IND submission asks "why anti-CD8 LNP and not anti-CD3?" — lnp_tropism.json has the cell-set analysis with the Tabula Sapiens citation. Not the agent's opinion.

"The biotechs with $400M of funding are running this co-design with spreadsheets, slide decks, and a four-month cycle. The cycle should be 30 seconds."

Try it yourself

Design in vivo CAR-T for CD19 B-cell lymphoma with anti-CD8 LNP delivery

→ biomate.ai · 30 seconds · 4 phases on AWS Batch · FASTA emitted.

Swap CD19 → CD7 and watch the fratricide warning fire. Swap anti-CD8 → anti-CD3 and the cell set widens (CD4 T cells join). Swap to DLBCL indication and BioMate notes the Kymriah precedent (Phase 3, 2017 approval). The pipeline propagates every change through all 4 phases.

Further reading: Maude et al. 2018, Tisagenlecleucel (Kymriah) — NEJM; Tabula Sapiens Consortium (Chan Zuckerberg Biohub); CELLxGENE Census (CZI); FDA Approved Cell & Gene Therapy Products.

What this means for cell therapy programs

The fratricide risk in in vivo CAR-T is real, invisible until late, and easy to catch early. A 4-phase co-design pipeline that cross-references the CAR target, LNP tropism, and cell-type expression at the atlas level turns a 4-month architectural debate into a 30-second safety check — with an auditable trail that an IND reviewer can follow.