Everyone asks me how I did it. Here it is**
Part 1: Rosie's Story**
June 2023: About three years ago, I noticed something wrong with Rosie. She had swollen lumps on her head and leg, and it was noticeably bothering her. Over the course of the next 11 months I took her to a veterinarian three times. During each visit, the vet was dismissive and no cancer diagnosis was made.
May 2024: At this point, the swelling had become severe and started bleeding. I insisted that the vet address it. Rosie had surgery to remove the swollen masses and a biopsy was taken. This is when the cancer was finally diagnosed. Eleven months late.
The fatal prognosis left me with a choice: accept it, or throw everything I had at it. I used ChatGPT and devoured everything I could to understand cancer and what I could do about it. All the while continuing a full-time job: running an AI consulting business.
After the diagnosis and my early research, I discussed what to do with a new veterinarian. Rosie was initially put on:
July 2024: Desperate for better treatment options, I settled on a high-end veterinarian clinic. With great success, many large masses were removed by surgery. One large mass remained - it would have required amputation of the leg. I debated but decided against it.
The clinic was great but on-going treatment was unaffordable. On a friend’s recommendation, I arrived at Dr. Mina Ghaly, also a vet.
August 2024: Dr. Ghaly is one of the unsung heroes of this story. He was the only vet that was truly receptive to me - a "citizen scientist." This new vet was critical in facilitating much of what was to come. About this time my journey in understanding cancer drug development had just begun. I knew that AlphaFold had been touted for its ability to accelerate the discovery of targeted cancer therapies and I wanted to see if that was applicable to Rosie’s case. I queried ChatGPT deeper on this, and the first step it suggested was to get genomic sequencing of Rosie’s DNA and her cancer cells’ DNA.
September 2024–January 2025: With cancer continuing to spread unabated, a new therapeutic regimen was needed. ChatGPT o1 identified Professor Martin Smith at the University of New South Wales (UNSW) as the lead for genomic sequencing.
Using tissue that was collected in January, genomic sequencing employed here was not like a cheek swab with partial sequencing from 23andMe. This was a full and complete genome sequence and required specific preservation of the tissue and extracted DNA. I drove the tissue samples from the vet to the Garvan Institute myself. Pavel Bitter's team at the Garvan Institute played a critical role in extracting DNA for sequencing. The DNA was then couriered to UNSW's Ramaciotti Centre for Professor Martin Smith to do his sequencing work - one of the critical deep science steps that just a couple of decades ago would have cost over a billion dollars and is rapidly coming down in price, while getting much, much faster.
https://www.ccralliance.org/about
February–May 2025: With the raw DNA data in hand, data analysis began in earnest. Gemini, Grok, and ChatGPT were all used in arriving at a data transformation protocol. My team and I eventually found a mutation in the c-KIT gene, a known driver of mast cell cancer. AlphaFold 2 was used to model and understand Rosie’s mutated c-KIT protein - a step that would have required specialised experimental techniques like crystallography only 9 years ago.
June–July 2025: The goal was to use a ligand to switch off the c-KIT mutation and hence stop the unfettered cancer growth. With a prospective cancerous mutation identified, we now had two pathways to consider:
But first, what is a ligand ? → Any molecule that binds specifically to a larger "target" molecule (usually a protein) to serve a purpose
(A) Using genetic algorithms - a numerical optimisation technique - to discover a novel ligand. Pulling together more scripts and simulation engines, we developed a candidate that worked in simulation. It could have taken years to go from petri dishes, to mouse models and finally to dogs. After further consideration, the risks and approvals process for a novel ligand made this approach unworkable within Rosie's timeframe. But the ligand candidate still exists - and if the approvals landscape ever catches up, it may yet have its day.
(B) Docking a library of pre-existing 1M+ ligands to look for a match. After two weeks of running 1 million candidates - eureka - a match! It felt like we finally had something that could work, but then this pathway began to unravel as well. The match was for an existing patented compound. This meant that there were huge legal barriers to use or synthesise the compound.
I did however make contact with the patent holder and seek a compassionate use exemption outside their existing legal pathways. Understandably this was too high a risk for them to consider and they declined.
This was a major low point. We had come so far - sequencing the healthy and cancerous DNA, running state-of-the-art tooling to develop candidates for treatment, identifying known compounds. It really felt like we had exhausted the space and it just was not meant to be.
August 2025: After two weeks of prioritising quality time with Rosie, an idea came to me - “what if it was possible to create a vaccine myself?” I spent the night with ChatGPT debating possibilities. Eventually we arrived at what looked like the best bet: a peptide neoantigen vaccine.
The vaccine approach is a fundamental shift from ligand docking; rather than merely obstructing the proteins that fuel cancer growth, vaccines enable the immune system to re-identify and actively eliminate cancerous cells.
I emailed Professor Martin Smith at UNSW's Ramaciotti Institute and he referred me to Dr. Deborah Burnett. Dr. Burnett said that while she did not have experience in peptide neoantigen vaccines, she did have experience with mRNA vaccines. This is yet another example of the ideation bouncing between humans and chat bots that was critical to the process. Following this advice from Dr. Burnett, I reverted back to ChatGPT and got confirmation that the mRNA approach would be faster.
Two pipelines were initiated in parallel:
September–mid October 2025:
Vaccine development:
Around early September, a fast-growing tumour needed to be removed. This was an opportunity for RNA extraction and sequencing.
Here we go again: tissue on ice, Pavel Bitter's team at the Garvan Institute, RNA couriered to UNSW, sequencing by Professor Martin's team at UNSW's Ramaciotti Centre.
Returning to the data analysis, we narrowed the field to 7 epitopes highly prevalent in both the mRNA and DNA. This intersection was the “smoking gun.” Because mRNA expression reveals the cell’s actual protein production, these 7 targets represent the most viable candidates for triggering a robust immune response.
To give you a flavour of being in my shoes, try asking your favourite LLM chat bot to list 7 random epitopes. For illustration purposes only, here are 7 random epitopes I got from Gemini:
And, here is one vaccine construct that I got from Gemini by merely passing the epitopes listed above and asking it to create one. Note that the construct is not unique, and depends on what one is optimising for:
PADRE—GPGPG—WITQCFLPVFLAQP—GPGPG—GVGSPYVSRLLGICL—AAY—KVAELVHFL—AAY—FLWGPRALV—AAY—SLLMWITQC—AAY—KIFGSLAFL
Try creating your own vaccine construct using your preferred chat bot of choice !
For the Rosie-specific case, I utilised Gemini 2 Pro to architect the multi-epitope vaccine construct, incorporating optimised linkers and adjuvants. The resulting sequence was then subjected to a secondary heuristic refinement via Grok 3, ensuring structural stability and minimising junctional immunogenicity.
DISCLAIMER: The vaccine construct listed above is strictly for educational purposes and is not indicated to be used for the purposes of drug development or as part of any treatment protocols.
Ethics approval (September–mid October 2025):
Navigating the ethics approval process was a second full-time job, demanding 120 hours of paperwork alongside my actual career. To manage the load, I turned back to chat bots for a massive tactical shift. I used them to decode the subtle legalese and opaque regulatory language necessary to secure permission for the trial.
The University of New South Wales was able to develop the vaccine but did not have a process for a trial of this "n=1" kind. To create a new process would have taken until mid-2026. Rosie did not have time.
My efforts to this point had already been considered unusual. UNSW and some other local media published them in July 2025 (albeit without the global fanfare). However, some experts were watching. I had managed to get the attention of Dr. Mari Maeda of the Canine Cancer Alliance in the USA - arguably the world's foremost authority on canine cancer.
Dr. Maeda connected me with Professor Rachel Allavena at the University of Queensland. Professor Allavena had an existing compatible trial approval. On hearing the news over a Google meet, I acted nonchalantly - glad the trial could go ahead. My real reaction after the call, was fist pumping so hard Rosie's fear of thunder was almost provoked.
October–November 2025: With the trial paperwork completed, Professor Pall Thordarson and his team at UNSW could now commence manufacturing the vaccine. After 6 weeks of their expert work, the vaccine was ready.
December 2025: Go time. Rosie jumped into the car and we headed north. We took the 10-hour drive to Queensland to commence treatment. However, unbeknownst to me, the relevant campus was another four hours' drive out of Brisbane.
Finally at the University of Queensland’s School of Veterinary Science, we were ready to proceed:
-
A specifically designed mRNA vaccine - a precision-guided weapon to strike the cancer, developed with a pipeline designed by ChatGPT o1, implemented for candidate development by Gemini Pro 2, and with the final design validated by Grok 3 Thinking.
-
Supporting drugs to weaken the cancer's defences - without these, mRNA vaccines effectively just bounce off or are rendered useless before they can hit the cancer.
January 2026: Three weeks in, the cancerous areas swelled up - actually a good sign. Pseudoprogression. It means the T-cells are swarming.
Six weeks in, two cancerous areas were shrinking.
February 2026: The two areas, specifically around Rosie's legs, were returning to what appeared normal. Residual flat bumps from now presumably dead cancer cells remained, i.e., scar tissue.
Despite the tumors on Rosie’s legs receding in size, there was still a growing mass on Rosie's rear. This mass was surgically removed and sent again for genomic analysis.
March 2026: March 8th - the data from the non-responsive tumour arrived, analysis commenced.
March 10th - early signs suggested there were differences in the non-responsive cancer compared to the one the vaccine was designed for.
March 11th - the story broke in Australia.
March 12th - the story ignited X.
The last week has been a whirlwind. Three years ago I saw that Rosie needed help. I did everything I could. I was fortunate to have a background in machine learning and win the trust of world leading experts who saw something worth pursuing.
The pace at which these events transpired could not have been possible without AI chat bots.
Part 2: The AI Story
Let me address the debate about what AI actually did.
This was not an upload of DNA to ChatGPT with the prompt: "Please make a vaccine with no mistakes."
It took months. It required ChatGPT, Gemini, and Grok - each for different things at different stages.
The hardest work done by AI - the core science: turning raw genomic data into a vaccine prescription.
We started with ~300 gigabytes of raw sequencing data - whole genome sequencing of Rosie's tumour and a matched normal blood sample, plus later RNA sequencing of the tumour. The final output was roughly half a page: an mRNA vaccine construct encoding 7 neoantigen targets.
As a non-biologist, I relied on chat bots to help me design a bioinformatics pipeline. I also worked with chat bots extensively to implement it.
These are the specific tools for people wanting to understand the process. For most readers this list will probably raise more questions than answers. It took me months to wrap my head around them, and I had the benefit of working on this for over a year at that point. So without further ado, the critical components:
-
BWA-MEM for alignment
-
GATK Mutect2 (a Bayesian somatic model) for variant calling
-
Ensembl VEP for annotation
-
pVACseq with NetMHCpan-4.1 - a neural network that predicts whether a mutant peptide will bind to Rosie's specific immune molecules (DLA alleles)
What I have come to understand is that these are the same class of tools used in human precision oncology pipelines.
We then validated the DNA-predicted targets against RNA sequencing data to confirm the mutations were actually being expressed by the tumour. Only epitopes with confirmed RNA expression were retained. These epitopes became the vaccine construct that I sent to the lab for synthesis.
The implementation and use of these tools are probably the substance of graduate and postdoc activity in various institutes. I do not think everyone should have to know what I know to be able to use these tools. I think this was far more complex than it needed to be.
So what did the chat bots actually do?
They were indispensable - but not in the way people assume.
-
Planning & pipeline design: I used ChatGPT and Gemini to map out the entire bioinformatics workflow - what tools to use, in what order, what the inputs and outputs should be.
-
Troubleshooting: When tools failed (and they failed constantly - dependency conflicts, reference genome incompatibilities, annotation format mismatches), we used chat bots to debug. The dog genome is far less well-annotated than the human genome, and many tools assume human data. Every step had friction.
-
Candidate filtering & construct design: Grok helped filter our epitope candidates and later helped convert the final confirmed targets into the mRNA vaccine construct - the actual sequence with linkers, UTRs, and codon optimisation for canine cells.
-
Treatment protocol design: ChatGPT and Gemini helped design the multimodal protocol - working through drug interactions, timing, immunological synergies, and the phased rollout of distinct therapies. This involved weeks of iterative analysis across pharmacology, tumour immunology, and veterinary oncology.
-
Education: I am not a biologist. The chat bots taught me what cancer is at a molecular level, what neoantigens are, what MHC/DLA molecules do, what truncal vs branch mutations mean, what the tumour microenvironment is, and why any of this matters.
-
Ethics, compliance & coordination: The chat bots helped me write the ~100 pages of ethics approval documents, navigate the regulatory landscape across Australian states, and coordinate with multiple universities.
The vaccine was not given alone.
The chat bots also helped design a multimodal treatment protocol - not just the vaccine, but the entire immunological strategy around it:
-
A personalised mRNA neoantigen vaccine - encoding 7 targets designed to train Rosie's T-cells to recognise and kill her specific cancer cells
-
A tyrosine kinase inhibitor - targeting Rosie's c-KIT mutation, providing foundational anti-tumour pressure and modulating the tumour microenvironment
-
A PD-1 checkpoint inhibitor - removing the brakes from Rosie's T-cells, unleashing their full killing potential against the cancer
I have seen a lot of commentary online suggesting it was really just the PD-1 inhibitor doing the heavy lifting, or just the tyrosine kinase inhibitor. This misunderstands how cancer actually works. A tumour does not just sit there waiting to be killed - it actively builds a suppressive microenvironment around itself, corrupting nearby immune cells, growing its own blood supply, and hiding from the immune system.
If this were Star Trek, the cancer tumour microenvironment is the ship's shields. You can fire as many photon torpedoes - mRNA trained T cells - as you want. If the shields are up, nothing gets through.
No single therapy punches through cancer’s “shields” alone. The tyrosine kinase inhibitor and the PD-1 inhibitor take the shields down - but from different angles.
The TKI targets the tumour's supply infrastructure. Cancer needs blood vessels to feed itself and it actively grows new ones - a process called angiogenesis. The TKI blocks this, starving the tumour of nutrients and oxygen. It also directly inhibits the c-KIT mutation driving the cancer's growth signal. In Star Trek terms, it is disabling the enemy's engines and cutting power to life support.
The PD-1 inhibitor works on a different problem entirely. Cancer cells display a protein on their surface that sends a "stand down" signal to any mRNA trained T-cell that approaches - essentially telling the immune system's soldiers to holster their weapons. The PD-1 inhibitor blocks that signal, so the T-cells can no longer be fooled into backing off. If the TKI is taking down the shields from the outside, the PD-1 inhibitor is disabling them from within - shutting down the tumour's ability to cloak itself from the very cells sent to destroy it.
Firing these two together: the TKI & PD-1 - the shields come down and the mRNA trained T-cell photon torpedoes get through.
The sequencing and timing of these therapies mattered enormously - you can not give immunosuppressant treatments, TKI’s & PD-1’s alongside an immune-activating mRNA vaccine. ChatGPT and Gemini helped me work through the pharmacological interactions, tapering schedules, and phased rollout across weeks of back-and-forth.
What the chat bots did NOT do:
They did not collect samples. They did not isolate or sequence the DNA. They did not physically manufacture the vaccine. They did not administer it. Many brilliant scientists were required - including Professor Pall Thordarson at the UNSW mRNA Institute who manufactured the vaccine, Professor Rachel Allavena & Dr. José Granados at the University of Queensland who administered it, and Professor Martin Smith who provided expert guidance on the bioinformatics throughout.
In summary:
The chat bots empowered me as an individual to act with the power of a research institute - planning, education, troubleshooting, compliance, and yes, real scientific design work in converting genomic data to a vaccine prescription and designing the treatment protocol around it. But they worked alongside humans at every step. The combination is what made it possible.
https://www.abc.net.au/listen/programs/healthreport/dog-owner-ai-cancer-treatment/105462166
Part 3: What is next
Since Rosie's diagnosis, I have done everything I can to help her. But the reaction to this story - the outpouring of support, the thousands of messages from people with dogs in distress - has made something clear.
Rosie received a fully individualised, multimodal mRNA cancer protocol. One dog, one vaccine, designed from scratch. Three months in, she is showing strong signs of improvement.
That was the easy part.
What I keep coming back to is this: I do not know why I had to be this obsessed to get here. There are so many unnecessary barriers. So many techniques and tools that could be made far easier to access. The science exists. The AI exists. The gap is in making it reachable.
For years, I have discussed with friends what kind of AI company I should start. About 3 months ago, one of them said - "why are you even asking that question? Surely it is helping other people do what you did for Rosie." At the time, I was not convinced. I had done a lot, and it had given me more options to try, but there had been no breakthrough. Now, three months into treatment, Rosie is improving. And the response from the community has been overwhelming.
Now the hard part.
I have spent the last week speaking to everyone involved to understand whether it really is possible to make this process more scalable.
We believe it is.
This will be the next chapter. It started with one dog. It will not end with one.
I will have more to share soon.
Best,
Paul
