Personalized Cancer Vaccines: The ChatGPT Question

A Sydney Dog, a ChatGPT Prompt, and a 75% Tumor Shrinkage Claim

The story that went viral in early 2025 had every ingredient for a modern science headline: a dying dog, a desperate tech entrepreneur, an AI chatbot, and a miracle recovery. Paul Conyngham, a Sydney-based technologist, adopted Rosie — an eight-year-old staffy-Shar Pei cross — in 2019. By 2024, tennis ball-sized tumors had appeared on her hind leg. The diagnosis: mast cell cancer, the most common skin malignancy in dogs.¹

Conyngham spent thousands on veterinary chemotherapy. It slowed the tumors but failed to shrink them. So he did what many of us would do in 2024 — he asked ChatGPT for help. What happened next turned into international news: Conyngham used the AI to help brainstorm approaches, process genetic data, and ultimately design a personalized mRNA cancer vaccine for his dog. He connected with Páll Thordarson, director of the UNSW RNA Institute, who manufactured the vaccine. Within a month of administration, the tumor had reportedly shrunk by 75%.¹

The headlines wrote themselves: "Tech boss uses AI to cure his dying dog." Social media exploded. The implication was clear: if ChatGPT can design a cancer vaccine for a dog, human cures can't be far behind.

But here's the thing — and it's the reason I'm writing this issue. The story of Rosie is genuinely touching. The broader field of personalized cancer vaccines is genuinely promising. But the relationship between "a man typed questions into ChatGPT" and "a dog's tumor shrank" is far more complicated than the headlines suggest. Let me walk you through what the science actually shows.

How Personalized Cancer Vaccines Actually Work

Every cancer is, at its molecular core, a collection of genetic mistakes. Mutations in tumor DNA produce abnormal proteins — called neoantigens — that don't exist in healthy tissue. In theory, these neoantigens are perfect immune targets: they're unique to the tumor, which means a vaccine targeting them should activate the immune system against cancer cells while leaving healthy tissue alone.²

The pipeline works like this. First, surgeons remove a tumor sample. Next, whole exome sequencing maps the tumor's DNA and compares it to the patient's healthy genome, identifying somatic mutations — the ones unique to the cancer. Then comes the hard part: predicting which of those mutations will actually produce neoantigens that the immune system can recognize.³

This is where computational biology earns its keep. Specialized algorithms — primarily NetMHCpan and related tools — use neural networks trained on binding affinity data to predict which mutant peptides will be presented on MHC molecules (the cellular display cases that show immune cells what's happening inside).⁴ The algorithms evaluate proteasomal cleavage patterns, TAP transport efficiency, and MHC binding strength to rank candidate neoantigens. The top 10-20 candidates make the final vaccine.

The delivery vehicle is mRNA — the same technology behind COVID-19 vaccines. Synthetic mRNA encoding those selected neoantigens is injected intramuscularly. The patient's own cells take up the mRNA, produce the tumor-specific proteins, and display them to the immune system. T cells activate, proliferate, and — if everything works — hunt down cancer cells expressing those exact mutations.⁵

The mechanism is elegant. The biology is sound. The question — as always — is whether it works in actual patients, in actual clinical trials, with actual controls.

The 157-Patient Trial That Changed the Conversation

If you want to understand where personalized cancer vaccines actually stand, ignore the dog for a moment. Look at KEYNOTE-942 — the trial that pharmaceutical analysts, oncologists, and the FDA are actually watching.

This is real data. A 44% reduction in cancer recurrence, sustained at five years, in a randomized trial with a control arm. The FDA granted mRNA-4157 Breakthrough Therapy Designation — the agency's way of saying "we want this to move faster."⁷ Phase 3 trials are underway. A commercial launch is anticipated by 2029.

Meanwhile, BioNTech has been running its own program with a different approach — and targeting one of the most lethal cancers.

And there's more coming. A Dana-Farber Phase 1 trial in kidney cancer reported that all 9 patients receiving a neoantigen vaccine remained cancer-free at 34.7 months.⁹ A Harvard Wyss Institute biomaterial-based approach achieved 100% immune activation in a 9-patient trial.¹⁰ As of late 2024, there were 78 personalized cancer vaccine trials worldwide — 70 in Phase 1, 8 in Phase 2 — with over 120 RNA vaccine platform trials running across multiple malignancies.¹¹

What ChatGPT Actually Did — and What It Can't Do

Now let's return to Rosie. The headlines credited ChatGPT with designing a cancer vaccine. Let me be precise about what that means — and what it doesn't.

Conyngham is a tech entrepreneur with a data science background. He used ChatGPT to brainstorm possible treatment approaches for mast cell cancer, to discuss the concept of neoantigen targeting, and reportedly to help process some of Rosie's genetic data.¹ He then partnered with Páll Thordarson, a world-class RNA chemist who directs the UNSW RNA Institute, to actually manufacture the mRNA vaccine.

Here's the critical distinction: ChatGPT is a general-purpose language model. It is not a bioinformatics pipeline. The real work of neoantigen prediction — MHC binding affinity modeling, proteasomal cleavage analysis, TAP transport efficiency scoring — requires specialized tools like NetMHCpan, which is a neural network trained specifically on peptide-MHC binding data from mass spectrometry experiments.⁴ ChatGPT has never been trained on these datasets. It cannot run these algorithms. It cannot access patient genetic data.

And there's a more fundamental problem. A 2023 study in The American Journal of Medicine found that 47% of references ChatGPT generates are entirely fabricated, 46% are authentic but contain inaccuracies, and only 7% are both authentic and accurate.¹² When asked for genetic information numbers, ChatGPT has been documented to simply invent them.

None of this diminishes what Conyngham achieved. He identified a promising therapeutic approach, assembled the right collaborators, and his dog appears to be responding. That's remarkable initiative. But the narrative that "ChatGPT designed a cancer vaccine" obscures the actual expertise involved — Thordarson's RNA chemistry, the sequencing infrastructure, the immunological knowledge — and overstates what a general-purpose LLM can do in a field that demands specialized computational tools.¹³

Real AI is transforming cancer vaccine design. Machine learning models like NetMHCpan, Gritstone's EDGE platform, and BioNTech's proprietary algorithms are genuinely accelerating neoantigen prediction.^4,14 But these are purpose-built systems trained on immunological datasets — not consumer chatbots. The distinction matters, because conflating the two could lead people to make dangerous decisions about their own cancer care based on ChatGPT conversations.

Why a Dog's Cancer Matters for Human Medicine

Here's where Rosie's story gets genuinely interesting — not because of ChatGPT, but because of a research paradigm called One Health.

Dogs develop spontaneous cancers that are genetically and immunologically similar to human cancers. Unlike laboratory mice — where tumors are artificially implanted into inbred animals with no genetic diversity — pet dogs develop cancer naturally, in genetically diverse populations, living in the same environments as their owners.¹⁵ This makes them arguably the best preclinical model for human cancer treatment outside of humans themselves.

The National Cancer Institute runs a Comparative Oncology Trials Consortium across 19 academic veterinary sites specifically because dog cancer data translates to human medicine more reliably than mouse data.¹⁶ Checkpoint inhibitors, HER2-targeted immunotherapy, and PI3K inhibitors have all been tested in dogs as a stepping stone to human trials.

Canine oncology already has a few immunotherapy tools. Gilvetmab, an anti-PD-1 antibody for dogs, received conditional USDA approval around 2023 for mast cell tumors and melanoma. STELFONTA was conditionally approved by the FDA in 2020 for non-metastatic mast cell tumors.¹⁵ And a five-year universal cancer vaccine trial for dogs is currently running across sites in Wisconsin, California, and Colorado, targeting lymphoma, osteosarcoma, and hemangiosarcoma.

So Rosie's case — a personalized mRNA vaccine for canine mast cell cancer — sits within a legitimate research ecosystem. The study design (n=1, no controls, concurrent chemotherapy) can't tell us whether the vaccine worked. But the concept of personalized neoantigen vaccination for canine cancers, as a translational pathway to human medicine, has real scientific merit. Thordarson himself noted that "what Rosie is teaching us is that personalized medicine can be very effective," while acknowledging that companies including Moderna are already further along in human trials.¹

Everything the Headlines Left Out

I want to be fair to Conyngham and Thordarson. They never claimed this was a clinical trial. But the media coverage treated an n=1 case as if it were proof of concept, and that requires some corrections.

Then there are the broader field-level challenges. No personalized neoantigen cancer vaccine has yet received full FDA or EMA approval as of early 2026.⁷ Current costs exceed $100,000 per patient, though AI-optimized pipelines could bring this down to roughly $10,000.¹¹ Manufacturing timelines remain measured in weeks — tumor sequencing, neoantigen prediction, mRNA synthesis, quality control — which is a problem for aggressive cancers. And there's a regulatory tangle: the EMA classifies these as gene therapy products while the FDA calls them therapeutic vaccines, which creates different approval pathways in different jurisdictions.¹⁷

The most important limitation is also the most boring one: we need Phase 3 data. KEYNOTE-942's Phase 2b results are strong but modest in sample size. BioNTech's pancreatic program is only Phase 1. The field is promising, but "promising" and "proven" are different words for a reason.

Two Stories, Two Levels of Evidence