Source: The Australian — "Tech boss uses AI and ChatGPT to create cancer vaccine for his dying dog" (March 13, 2026)

The Dog

Rosie is an eight-year-old staffy-shar pei cross, rescued from bushland, adopted from a Sydney shelter in 2019. In 2024, she was diagnosed with mast cell cancer — aggressive, multifocal, terminal. Her owner Paul Conyngham threw thousands at veterinary chemotherapy and surgery. The tumors slowed but didn't shrink. Months to live.

Conyngham is not a biologist. He's a data engineer — 17 years of machine learning, co-founded Core Intelligence Technologies, former director of Australia's Data Science and AI Association. When conventional treatment failed, he opened ChatGPT.

The chatbot suggested immunotherapy and pointed him to UNSW's Ramaciotti Centre for Genomics. He paid $3,000 to sequence both Rosie's healthy blood DNA and her tumor DNA. Then he did something no one expected: he analyzed the data himself.

He ran the genomic output through mutation-finding pipelines, used AlphaFold to identify the mutated proteins, then used other algorithms to match targets to potential drugs. When the UNSW scientists saw what he'd done — with no biology training — they were, in Associate Professor Martin Smith's words, "gobsmacked."

An immunotherapy drug was identified but the pharma company refused compassionate use. Dead end. Then Smith mentioned mRNA vaccines. Conyngham pivoted. He designed the mRNA sequence encoding Rosie's tumor-specific targets and sent a half-page formula to Pall Thordarson, director of UNSW's RNA Institute. Thordarson's team manufactured a custom lipid nanoparticle vaccine.

The regulatory wall was worse than the science. Conyngham spent three months, two hours every night, writing a 100-page ethics document. He never got approval. Instead, a US researcher saw a UNSW website blurb and connected him to Rachel Allavena at the University of Queensland, who already had blanket ethics approval for experimental canine immunotherapy. He drove 10 hours with Rosie for the first injection in December.

The tennis ball-sized tumor on Rosie's hock halved. Her energy returned. She jumped a fence to chase a rabbit.

Now Conyngham is designing a second vaccine — sequencing again, finding why parts of the tumor didn't respond, running the same pipeline with updated inputs. Version 2.

What This Isn't

Before going further — a reality check from people who actually do this for a living.

Patrick Heizer, a cancer researcher, responded on X: "It is ~trivially easy to make a single mRNA vaccine. It's not hard. I cure mice of various cancers with various therapeutics all the time. I've made mice lose more weight in a month than tirzepatide does in a year."

He's right. Making one personalized mRNA vaccine for one patient is not a scientific breakthrough. It's a Tuesday for a competent biology lab. Heizer has an ongoing experiment where 100% of untreated animals had to be euthanized while 100% of treated animals are seemingly unaffected. But he's still "extremely far away from proving that it works."

The dog's tumor halved. It wasn't cured. N=1, uncontrolled, in a species with a different immune system. Domain experts aren't impressed by the science. They're right not to be.

So what IS the signal?

What This Is

The signal isn't that the science worked. The signal is that the science was executed by someone who isn't a scientist, using tools that are freely available, in weeks instead of years. The design layer of personalized medicine just became accessible to non-experts.

What Conyngham demonstrated is a pipeline:

1. Sequence healthy DNA + tumor DNA ($3K, days)
2. Diff the genomes to find mutations (AI, hours)
3. Identify mutated proteins as targets (AlphaFold, hours)
4. Design mRNA encoding those targets (ChatGPT, hours)
5. Manufacture nanoparticle vaccine (lab, weeks)
6. Inject — immune system attacks tumor
7. Partial response? Sequence again, design v2, repeat

Cancer researchers already knew every step of this was possible. What changed is that steps 2-4 collapsed from "requires a PhD and six months" to "requires ChatGPT and an afternoon." The science was always there. The accessibility barrier just fell.

This matters because it reframes where the moats are. If design is commodity — and it is, Heizer confirms researchers do this routinely — then the value lives elsewhere in the chain.

Where The Moats Actually Are

Heizer nailed the answer in a follow-up: "The scale point isn't about any individual vaccine, but when 100,000 other people come knocking wanting theirs."

Making one is trivially easy. Making 100,000 unique ones — each with patient-specific mRNA, each at GMP quality, each with full QC, each within a 60-day window from biopsy to injection — is an unsolved manufacturing problem. And proving each is safe and effective in randomized controlled human trials is a regulatory problem that doesn't have a framework yet.

STEP              MOAT?   WHY
────────────────────────────────────
Tumor sequencing  NO      $3K, commodity
Mutation ID       NO      AI, free — even non-experts can do it now
mRNA design       NO      Trivially easy (domain experts confirm)
LNP delivery      YES     Patented formulations, but expiring 2029
GMP manufacturing YES     Physical plant, per-patient QC, 60-day clock
Regulatory path   YES     No framework exists yet — first mover defines it

This is a Friction Frontier pattern. AI broke the old friction (drug design). It created new friction at the manufacturing and regulatory bottleneck. The science is commodity. Manufacturing and regulation are the moats. Companies that own the remaining friction win.

Can the FDA Even Approve This?

This is the hardest question. The science works — Heizer confirms it's trivially easy to design, Moderna's Phase 2b confirmed it works in humans. But the FDA doesn't approve science. It approves products. And a personalized mRNA cancer vaccine is, by definition, a different product for every patient.

The mechanical problem: FDA's entire framework is built around approving a SPECIFIC DRUG for a SPECIFIC INDICATION. One molecule. One Phase 3. One BLA. One approval. Keytruda is Keytruda — the same molecule goes into every patient. But a personalized mRNA vaccine encodes 20-34 neoantigens unique to YOUR tumor. Patient A gets a completely different molecule than Patient B. How do you run Phase 3 on a drug that doesn't exist until after the patient is enrolled?

What Moderna is actually testing: This is the key. INTerpath-001 is not a trial of a specific drug. It's a trial of a PROCESS: "does the pipeline of sequencing your tumor, algorithmically selecting neoantigen targets, encoding them in mRNA, manufacturing an LNP vaccine, and injecting it — does that process produce better outcomes than Keytruda alone?" 1,089 patients, each getting a unique mRNA sequence, all manufactured under the same validated process. The "product" is the pipeline itself.

The priors — has FDA ever done this?

CAR-T (2017-2022): The strongest precedent. Six approved products. Each patient's dose is unique — their own T-cells, harvested, modified, expanded, reinfused. FDA approved the manufacturing PROCESS plus the specific modification target (CD19, BCMA). One BLA per target/indication. Each individual patient's dose within that BLA didn't need separate review.

BUT — each CAR-T targeting a DIFFERENT antigen required a NEW BLA. Kymriah (CD19, B-ALL) is a separate approval from Abecma (BCMA, myeloma). Six products, six BLAs, six sets of clinical data.

CAR-T is also now expanding beyond oncology. A recent case series showed CD19 CAR T-cell transfer was feasible, safe, and efficacious across three different autoimmune diseases — SLE, inflammatory myositis, and systemic sclerosis — with all patients achieving clinical response and completely stopping immunosuppressive therapy. FDA is developing comfort with individualized biologics across disease categories.

This is the exact precedent for personalized mRNA cancer vaccines. CAR-T proved the FDA CAN approve individualized products where every dose is unique. It also proved each indication still needs its own BLA.

COVID strain changes (2020-2025): FDA allowed variant-updated COVID boosters via "strain change" supplements rather than full new BLAs. This is the CLOSEST analogy to platform approval — same manufacturing process, different mRNA sequence, abbreviated review. But the platform was validated across billions of doses with known viral targets. Cancer neoantigens are algorithmically selected and patient-specific. The FDA's comfort level with "trust the algorithm, don't review each sequence" is untested.

Platform Technology Designation (FDORA 2022): Created post-pandemic. FDA's own draft guidance cites mRNA/LNP as a candidate platform technology. BUT: the first product must be approved the hard way — full BLA, full Phase 3. After that, subsequent products can cross-reference CMC (chemistry, manufacturing, controls) data from the first approval. Clinical efficacy data still required per indication. No designation has been granted yet. Moderna is the obvious first applicant — after melanoma approval.

CMC Flexibilities (January 2026): Three specific changes that directly help personalized vaccines. Sponsors no longer need three PPQ (Process Performance Qualification) lots — a major bottleneck when each lot is unique. Process validation can leverage prior knowledge. Release specifications can be revised post-approval. FDA explicitly acknowledging the old framework doesn't work for individualized products.

Plausible Mechanism Pathway (February 2026): New draft guidance for ultra-rare individualized therapies where randomized trials are infeasible. But this targets genetic disorders with known molecular cause, NOT cancer. Wrong analogy — cancer is common, even if the treatment is individualized.

What happens first approval through last:

FIRST APPROVAL (melanoma):
  Full BLA. Full Phase 3 (INTerpath-001, 1,089 patients).
  Full CMC review of the manufacturing process.
  Full review of the AI/bioinformatics neoantigen selection pipeline.
  40+ QC tests per patient batch.
  This is the HARD one. Estimated 2028, 45-55% cumulative probability.

SECOND APPROVAL (NSCLC):
  New Phase 3 required (INTerpath-002, enrolling, data ≈2030).
  BUT with Platform Technology Designation after melanoma:
    → CMC cross-referenced (don't re-prove the manufacturing process)
    → Algorithm validation cross-referenced
    → Rolling review eligible
  Clinical efficacy still needs its own data.

THIRD+ APPROVALS:
  Each incrementally easier on CMC.
  Each still requiring Phase 3 clinical efficacy data.
  Compounding advantage for first mover.

What FDA will NOT do (10% probability): Approve the process as a blanket "platform" where any cancer type can be treated via supplement. Cancer is too heterogeneous. The algorithm that selects neoantigens may work in melanoma (high mutational burden, immunogenic) and fail in pancreatic cancer (immunologically cold). FDA won't trust a blanket process approval across fundamentally different tumor biology.

What FDA will likely do (50% probability): Hybrid model. Approve process-level components (manufacturing, LNP delivery, QC framework) via Platform Technology Designation. Require indication-level clinical evidence (Phase 3 per cancer type). Each subsequent approval is faster and cheaper but not automatic. This is the CAR-T model applied to mRNA.

The Conyngham paradox: A dog owner can design and manufacture a personalized cancer vaccine in weeks. Getting it approved for human use will take 5-10 years of Phase 3 trials across multiple indications. The science is a decade ahead of the regulation. That gap is where both the friction and the investment thesis live.

The 2011 problem: FDA's only existing guidance for therapeutic cancer vaccines is from 2011 — before personalized neoantigen vaccines existed. There is no specific guidance for personalized cancer vaccines in the 2026 CBER agenda. The regulatory framework for this category is being built in real-time, and the first approved product will DEFINE it.

Five Paths

We went hunting for the trade across five parallel research tracks. Every obvious vehicle was either dead, marginal, or too big to have edge.

Path 1: Arcturus Therapeutics (ARCT, $6.68, 90% idio)

Self-amplifying RNA platform — lower dose per patient means easier personalized manufacturing. Only approved sa-mRNA vaccine globally. Looked like the perfect small-cap vehicle.

Dead on arrival. Oncology work is preclinical mouse data only. No IND, no trials, no timeline. Their oncology partner (Achilles Therapeutics) voluntarily liquidated in March 2025. Management isn't mentioning oncology on earnings calls. Revenue halving. CSL wrote down $430M. CEO hasn't bought stock since 2022 despite 53% decline. The 32% short is well-constructed.

Path 2: Maravai LifeSciences (MRVI, $3.08, 69% idio)

Makes CleanCap — the capping technology used in virtually all mRNA manufacturing. Picks-and-shovels.

Marginal. CleanCap is real but not irreplaceable — enzymatic capping works, big pharma insourced. Revenue down 79% from peak. $244M term loan due October 2027 with negative operating cash flow. Must refinance in 19 months. Oncology exposure is indirect. Insider buying is strong ($1.88M by CEO + director), but the debt maturity and indirect thesis exposure make it weak.

Path 3: mRNA CDMO Landscape

Nobody manufactures personalized mRNA at commercial CDMO scale. COVID manufacturing and personalized manufacturing are completely different businesses — unique sequence per patient, full QC per batch, 60-day turnaround. DNA template production (≈30 days) is the primary bottleneck. Moderna and BioNTech do it in-house. The CDMO infrastructure for third parties doesn't exist.

Danaher (DHR) is the only company to have manufactured a personalized mRNA therapeutic (CRISPR for UCD infant, May 2025, NEJM). But 26% idio — you're buying the S&P. The interesting companies (BioCina, ReciBioPharm, BioNTech's RNA Printer) are all private.

Path 4: LNP Patent Landscape

Arbutus/Genevant own foundational LNP composition patents. Validated by a $2.25B Moderna settlement (March 3, 2026). Core patents expire April-June 2029.

Critical finding: the settlement covers infectious disease vaccines ONLY. Moderna's personalized cancer vaccine uses the same SM-102 LNP but oncology is explicitly excluded. Latent patent liability on the lead program. No injunctions sought — damages only — so it won't be blocked, but the financial exposure is real.

The Race

Five paths searched. One horse found.

Moderna/Merck is the only realistic near-term candidate. INTerpath-001 (Phase 3, adjuvant melanoma, n=1,089) is fully enrolled, with interim data expected H2 2026. Five-year data from the Phase 2b showed 49% reduction in recurrence or death (HR 0.51), sustained. FDA granted Breakthrough Therapy Designation.

BioNTech/Roche is in trouble. Their personalized neoantigen platform failed first-line melanoma (Imcode-001), hit a safety hold in bladder cancer (Imcode-004), and withdrew from NSCLC. Three indications, three setbacks. 3-5 years behind. Everyone else — Gritstone (failed), OSE, Nykode — is Phase 1-2.

Scenarios

TAM at maturity:

US new cancer diagnoses:          2M/year
Immunotherapy-eligible (≈60%):    1.2M
Adoption at maturity (30-40%):    400K patients/year
Treatment cost per cycle:         $25K
US TAM at maturity:               $10B
Global (US = 40% pharma spend):   $25B

Probability of first US approval:

Scenario analysis (MRNA):

                    Revenue      EV         vs Current ($10B EV)
Bull (25%):         $8-12B      $40-60B    +300-500%
  Melanoma approved, pipeline expansion to NSCLC/CRC/others

Base (45%):         $4-6B       $16-24B    +60-140%
  Positive interim data, reclassification begins, slow rollout

Bear (30%):         $2-3B       $6-9B      -10 to -40%
  Phase 3 disappoints, class effect questioned, COVID decline continues

Probability-weighted expected return: +56%.

Five risks stress-tested:

The Thesis

The Rosie story isn't a scientific breakthrough. Domain experts do this every week. What it revealed is where the moats are and aren't in the value chain of personalized medicine.

Design is commodity — a dog owner proved it, researchers already knew it. Making one personalized vaccine is trivially easy. Making 100,000 unique ones at GMP quality, proving each is safe and effective in controlled human trials, getting FDA to approve a process that produces a different product for every patient — that's the hard problem. The moat is in manufacturing and regulatory navigation, not in the science.

The market prices Moderna as a declining COVID franchise. The reality is that they're running the only Phase 3 for a personalized cancer vaccine, with $10B cash, a Merck partnership, and 5-year data showing 49% reduction in cancer recurrence. The catalyst (interim data) is 6-9 months away. The narrative hasn't shifted yet.

The science is trivially easy. The regulation is brutally hard. The gap between those two facts is where the thesis lives.