Submission note: I wrote this article last month, when this case was first reported. It has since been covered by Astral Codex Ten for April's linkpost, and was praised by RFK Jr. in a Senate hearing on Wednesday. (RFK Jr. was seemingly unaware that the AI-powered treatment he was referring to was an mRNA vaccine, a technology he has a history of opposing). This article aims to contextualize the role of AI in Conyngham's story.

The Australian (archive link) recently reported that the entrepreneur Paul Conyngham developed a personalized mRNA vaccine which successfully treated his dog’s mast cell cancer. The buzzword-laden headline, “Tech boss uses AI and ChatGPT to create cancer vaccine for his dying dog”, attracted skepticism from AI pessimist corners of the internet. Meanwhile, popular retellings from AI optimists tended to exaggerate the role of ChatGPT, while downplaying the involvement of researchers at the University of New South Wales (UNSW),^[1] who were responsible for a large bulk of the process.

Underneath the noise, however, is a story that is entirely true and extremely interesting - though maybe not for the reasons you think.

What is a personalized mRNA vaccine, and why is it useful for cancer?

mRNA vaccines were catapulted into mainstream awareness by the COVID-19 pandemic, where they proved highly effective. The mRNA in a COVID-19 vaccine codes for a viral spike protein, which is expressed and degraded by a host dendritic cell, then presented to T cells via the major histocompatibility complex (MHC) system, ultimately training an immune response against viral antigens.

The flexibility here is impressive. Your body has a built-in process for teaching your immune system to fight stuff that you can hijack, provided your desired target has an antigen on its surface that can be used to distinguish it from normal cells. Cancer cells often carry tumor-associated antigens (TAAs) - found on healthy cells, but expressed too much or in the wrong tissues by cancer cells - and tumor-specific antigens (TSAs) - which are viral antigens or novel mutant antigens (neoantigens). An mRNA vaccine can teach the immune system to attack cancer cells, just like COVID-infected cells.^[2]

This makes mRNA vaccines a promising cancer treatment. Human trials have been published as early as 2008. But while COVID-infected cells all express approximately the same spike proteins,^[3] every individual cancer is unique, complicating the problem of vaccine design.

One approach is to identify a set of common antigens found in a specific cancer type and hope that enough work for any given patient.^[4] Off-the-shelf designs have obvious practical advantages, but their fixed antigen selections are typically far from optimal for any given patient. They are bad at targeting neoantigens, which are considered more attractive targets due to being unique to cancer cells, and thus more likely to trigger an immune response with fewer side effects. One of BioNTech’s early attempts at “off-the-shelf” melanoma vaccines was phased out after middling Phase II results, though the company still has several ongoing clinical trials for similar vaccines targeting both TAAs and viral TSAs in various cancers.

A second, sexier approach is personalized mRNA vaccines.

Every cancer is unique? No problem. We’ll just sequence the patient’s genome^[5] and design a custom vaccine.

Personalized mRNA vaccines have seen impressive results from clinical trials. In a five-year follow-up of Moderna & Merck’s Phase IIb mRNA-4157 trial, the vaccine resulted in a 49% reduction in the risk of recurrence or death among high-risk melanoma patients already on standard treatments (confidence interval 29.4-88.7%).^[6] Moderna & Merck alone have “eight Phase 2 and Phase 3 clinical trials underway across multiple tumor types including melanoma, non-small cell lung cancer, bladder cancer and renal cell carcinoma,” and they aren’t the only companies in the field. We should expect to see the conclusion of the first Phase III trials before 2030.

Until the vaccines are approved, most patients’ best hope for accessing them is getting into a clinical trial. For Conyngham’s dog Rosie, even that wasn’t an option. Conyngham had to do it himself.

How impressed should I be by Paul Conyngham?

How hard is it to make your own mRNA vaccine? The process, as described by Moderna in the supplemental material of their Phase IIb mRNA-4157 paper, looks like this:

1. Sequence patient DNA, tumor DNA, tumor RNA
2. Identify candidate neoantigens from mutations
3. Select neoantigens based on predicted immunogenicity
4. Design an optimized mRNA sequence from selected neoantigens
5. Manufacture mRNA vaccine

Steps 2-4 are entirely computational. Step 1 and 5 require a lab, which Conyngham was able to access through UNSW.

Conyngham, with the support of several UNSW researchers, followed a documented process for manufacturing a bespoke treatment too new and expensive to be offered by existing veterinary practices. That’s not easy. It shows impressive skill and determination. But we must put things into perspective: this is not a novel discovery, and no, AI has not cured cancer.

Conyngham’s story is an example of AI behaving as a normal technology

Few people actually claim AI has cured cancer. More credible sources have instead advanced the narrative that Conyngham’s story is an example of AI revolutionizing healthcare. This is true to a degree, but I am skeptical of the hype. Conyngham’s story is consistent with AI as a normal technology.

Has AI contributed to the development of mRNA vaccines?

Personalized mRNA vaccines are made possible by innovation across multiple disciplines. The cost of sequencing a human genome has fallen 10,000-fold since 2000 thanks to the development and commercialization of faster, better techniques. Advancements in mass spectrometry-based proteomics contributed to the creation of peptidome datasets for the prediction of peptide-HLA presentation.^[7] mRNA vaccines themselves are built on several new techniques, including the Nobel-prize winning discovery that you can use modified nucleosides^[8] to prevent the mRNA from triggering a cell’s innate immune system.

Computational advancements are an integral part of this: algorithmic data analysis is required for everything from assembling genomes to interpreting mass spectra. However, traditional ML approaches are usually preferred over the deep learning models associated with modern AI. AlphaFold aside, there aren’t too many situations where you’d pick a stack of transformers over a traditional method. Deep learning is best fit for tasks involving high inherent complexity and large, balanced datasets, a description that matches protein folding but not that much else.

That isn’t to say deep learning is never useful. The peptidome dataset paper I linked developed HLAthena, a publicly available HLA-binding prediction model based on a simple fully connected network. Protein language models are in principle promising for predicting neoantigen immunogenicity, though it’s too early to tell.^[9] You should think of AI as a specialty tool, one component among many that made personalized mRNA vaccines possible.^[10]

Has AI empowered laymen to pursue DIY medicine?

AI optimists covering Conyngham’s story often place less emphasis on AI’s contribution to the underlying technology, and more on the idea that commercial large language models have democratized healthcare. Commercial AI services have massively lowered the barrier of entry to new fields: vibecoding has allowed non-programmers to create software, and generative music has allowed non-musicians to create songs.^[11] There are no shortage of anecdotes about ChatGPT giving useful health advice.^[12] Conyngham himself enthusiastically credited ChatGPT with aiding him in the vaccine development process, and I don’t doubt him.

However, it is important to keep in mind that Conyngham has, to quote the Australian, “17 years of experience in machine learning and data analysis” with an impressive resume. The work Conyngham had to do was computational. ChatGPT explained oncology concepts to an intelligent amateur also getting guidance from university researchers, and it assisted an experienced engineer in a complex data analysis task, but it did not eliminate the need for human expertise. Without Conyngham’s experience, money,^[13] or the support of UNSW, this project wouldn’t have been possible.

The utopian vision of DIY medicine

Over at Persuasion, Ruxandra Teslo has used Conyngham’s story as a springboard to argue for the necessity of clinical trial reforms, citing the difficulty Conyngham faced in getting approval to use his custom vaccine. Her original post was provocatively titled, “The Bureaucracy Blocking the Chance at a Cure.” Teslo brought up two other names that might be familiar to her audience: GitLab co-founder Sid Sijbrandij, who faced Kafkaesque red tape trying to conduct experimental treatments on himself, and writer Jake Seliger, who died in 2024, and (along with his wife, Bess Stillman) reported on the insanity of navigating clinical trials as a terminally ill patient.

Frustration with the loss of patient autonomy in healthcare is universal across the political spectrum. But it’s felt particularly strongly^[14] by the tech crowd,^[15] who often start with the assumption that established systems are broken, and believe themselves to be intelligent and agentic enough to fix it. “Founder Mode”, in Sijbrandij’s words. Their dream of DIY medicine is reflected in the growing popularity of direct-to-consumer health products.

In this context, the hype around commercial AI “revolutionizing healthcare” makes sense. LLM chatbots are good at lowering barriers of entry, both for navigating legal red tape and overcoming knowledge gaps. They’re usually reliable enough at summarizing literature, and generating custom explanations. If you’re a wannabe biohacker that doesn’t know much biology, AI feels like magic.

Is it magic? As noted, I’m skeptical. Accomplishments like Conyngham’s require resources and expertise most people lack. There are real downsides to relying too much on an LLM to explain ideas to you, even if you think you’re being mindful of hallucinations. I’ve met scores of otherwise intelligent people who’ve convinced themselves of ridiculous theories about cancer after an extended conversation with Claude, and I’ve had to talk them out of it.^[16] Commercial AI as it stands now cannot grant amateur biohackers the knowledge to replace their doctors. Biotech startups are far from liberating medical development from bloat and bureaucracy. Conyngham may have bypassed the system, but only by finagling backdoor access to technology that was invented and will be deployed by the traditional players.

The utopian vision of DIY medicine is far from being achieved, and may never^[17] be achieved in a way that works for average people. But its central principle of increasing patient autonomy is worthwhile. AI can help us achieve that, if used correctly.

Addendum

Did it work?

Maybe. The tumor shrunk and the dog's symptoms improved, but I'm not sure how confident we can be in attributing this to the vaccine. If anyone has a better analysis, I'll link it here.

I want to help make open source mRNA vaccine design tools easier to access.

Contact me.

1. ^{^}

   See their reporting here: https://news.unsw.edu.au/en/meet-the-man-who-designed-a-cancer-vaccine-for-his-dog

2. ^{^}

   Here is a good overview that I used as a source for most of this article: “Leveraging mRNA technology for antigen based immuno-oncology therapies.”

3. ^{^}

   When a variant pops up with a mutated spike protein, it often reduces the efficacy of existing vaccines, which is part of what boosters try to address.

4. ^{^}

   Many recent (and less recent) advancements in cancer treatment are hyper-specific in this way, working only for cancers with particular mutations or even for patients with particular hereditary diseases. Despite what charlatans with Theories of Everything may try to sell you, it’s unlikely there are any straightforward, broadly effective cancer treatments lying undiscovered. Real progress in oncology is found in increasingly precise targets and increasingly precise techniques: a war waged through attrition.

5. ^{^}

   Normal DNA, tumor DNA, and tumor RNA. Recall that the pipeline goes DNA → RNA → protein; not all mutant proteins identifiable through DNA are even expressed, making tumor RNA useful.

6. ^{^}

   What’s stopping personalized mRNA vaccines from working 100% of the time? One broader limitation is that not all tumors are immunogenic: there’s a reason why trials are focused on specific types of cancer known for triggering immune response, such as melanoma. In the context of this specific trial, treatments may fail because tumors mutate to stop expressing antigens or develop other immune-evasion tactics, or because T cells have difficulty penetrating solid tumors, or other reasons. Better neoantigen selection and faster development loops can overcome some challenges, but not all of them. Cancer is complicated.

7. ^{^}

   A peptide (protein fragment) must bind to HLA (MHC) molecules on the cell surface so it can be presented to T cells. Likelihood of presentation is determined by peptide-HLA binding affinity, among other factors, and is a key factor in the broader problem of predicting neoantigen immunogenicity. When designing a vaccine, you want to select neoantigens that will trigger an immune response.

   People have different HLA alleles, which affects how their cells present antigens and is ideally accounted for in the process of vaccine design. This is another advantage that personalized mRNA vaccines have over off-the-shelf ones. Dogs have an analogous system, though I don’t know whether Conyngham’s neoantigen identification process involved MHC typing.

8. ^{^}

   You know how DNA is made of A/T/C/G bases, and RNA is made of A/U/C/G? You can swap out bases with similar molecules, U for Ψ for example, to create mRNA that can still be translated but isn’t recognized by the cell as foreign. This technique is used by most major COVID vaccines.

9. ^{^}

   We’re bottlenecked by data. It would be great to throw the immunogenicity problem into a massive pile of linear algebra and call it a day, but that’s not easy with a dataset limited to ~1100 experimentally validated human neoantigens. Instead, we have to find a way to leverage the generalized knowledge encoded in protein language models, or simplify our approach. Moderna describes their neoantigen immunogenicity prediction algorithm as “a deterministic machine learning (ML) algorithm” that only aims to predict likelihood of presentation. My best guess is that Moderna uses small, simple neural networks at most.

10. ^{^}

    Another issue with AI is that directly using the outputs of black-box models for anything directly affecting a patient, including mRNA antigen selection, makes people uneasy. This problem is not impossible to overcome: the recent FDA approval of the ArteraAI Prostate Test is evidence of that.

11. ^{^}

    You can complain about these things if you want, and they do have flaws, but you can’t deny they’re useful. I listen to AI-generated music almost every day.

12. ^{^}

    And less useful advice.

13. ^{^}

    The entire process took “tens of thousands of dollars”, according to the Australian. The commonly cited $3k figure was only for the initial gene sequencing.

14. ^{^}

    Among the relatively wealthy, that is. Those with fewer resources tend to be at least as concerned about accessing any default standard of care as they are with the limitations of it, and one could say people have strong feelings on the matter.

15. ^{^}

    Silicon Valley culture? There’s gotta be a better term for this.

16. ^{^}

    Why this happens despite users being aware of LLM limitations, and how to prevent it, deserves a post of its own. For now, the best generic advice is to always remain skeptical, and get a foundational, non-primarily-LLM-mediated education in biology if you’re serious about it.

17. ^{^}

    Barring a weird singularity.