I Built an AI Genetics Engine With No Biology Degree. So Did the Guy Who Cured His Dog.

Act One: Three Sentences From a Vet

Paul Conyngham is a senior data scientist in Sydney. Seventeen years of machine learning experience. Director of Core Intelligence Technologies, a consultancy in Darlinghurst. He has built recommendation engines, predictive models, and data pipelines for companies like Woolworths Group. He has never taken a biology class.

In 2019, he adopted Rosie from a shelter. A Staffordshire Bull Terrier crossed with a Shar Pei. Stocky, loyal, the kind of dog that leans her full weight against your leg when she's happy.

In 2024, large tumors appeared on one of Rosie's back legs. Mast cell cancer. The most common skin cancer in dogs, and one of the most aggressive. The vet estimated one to six months. Conyngham tried everything the conventional path offered: multiple surgeries, chemotherapy, immunotherapy. The chemo slowed the spread but failed to shrink the growths.

The tumors kept growing.

So he did what a machine learning engineer does when the known solutions fail. He started building a new one.

The Pipeline a Data Scientist Built

The first thing Conyngham opened was ChatGPT.

Not to ask it for a cure. To ask it how cancer immunotherapy works. What neoantigens are. How mRNA vaccines target specific mutations. How to design a computational biology pipeline when you've never touched one before.

ChatGPT became his tutor. His literature review assistant. His translator between the world he knew (data, models, code) and the world he needed to enter (genomics, immunology, molecular biology).

Then came the money. $3,000 to the Ramaciotti Centre for Genomics at UNSW to sequence Rosie's tumor DNA against her healthy cell DNA. The goal: identify the exact mutations driving her cancer. Not a generic "mast cell cancer" profile. Rosie's specific mutations. Her cancer's fingerprint.

With the sequencing data in hand, Conyngham turned to AlphaFold. Google DeepMind's protein structure prediction model. The same AI that won the 2024 Nobel Prize in Chemistry. He used it to model the 3D structures of the mutated proteins encoded by Rosie's tumor. To see the shapes. To find the vulnerabilities.

Then he wrote his own machine learning algorithms. Custom neoantigen selection: which of the mutated proteins would most likely trigger an immune response? Which ones would the immune system recognize as foreign? Which ones could a vaccine teach the body to attack?

He condensed months of analysis into what he described as "half a page of formulas."

Read that again. A data scientist with no biology training reduced a complex immunological analysis to half a page of formulas. Using AI tools that didn't exist three years ago.

He brought those formulas to Professor Pall Thordarson, director of the UNSW RNA Institute. Thordarson and his team reviewed the analysis, validated the approach, and designed and produced a bespoke mRNA vaccine. Custom-built for one dog. Based on her specific cancer mutations.

Professor Rachel Allavena at the University of Queensland's School of Veterinary Science administered the first injection in December 2025, at the Gatton laboratory. Conyngham had driven ten hours with Rosie to get there. A 100-page ethics application had taken three months to approve.

A booster shot followed in January 2026. Another was scheduled for March.

The Tennis Ball Shrank

The tumor on Rosie's leg had been the size of a tennis ball.

After two injections, it shrank by 50 to 75 percent. The Australian newspaper, which broke the story on March 14, 2026, reported it had "halved." Other outlets reported up to 75% reduction.

By January 2026, Rosie was jumping over fences to chase rabbits. Behavior that had disappeared entirely during her illness.

Conyngham was careful with his words. He told reporters he does not believe the treatment is a cure. He hopes it has given Rosie more time and a better quality of life. This is a single case, not a clinical trial. One dog, one vaccine, one outcome.

But the world heard something else.

Greg Brockman, president of OpenAI, shared the story on X: "The first personalized cancer vaccine designed for a dog." Trung Phan, whose posts reach millions, wrote: "Australian tech entrepreneur... used ChatGPT/AlphaFold (spent $3,000 with no biology background) to create a custom mRNA vaccine to treat his dog's cancer tumors. Unreal."

Within 48 hours, the story had been covered by outlets across five continents.

Let that land.

A data scientist, using AI tools that are free or nearly free, designed the computational backbone of a personalized mRNA cancer vaccine. For three thousand dollars. In his spare time.

The Skeptics Had a Point. And They Proved Ours.

Patrick Heizer, a biomedical engineer, posted on X to 1.5 million viewers: "Sorry to be the downer because this is an impressive story in some senses. But 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."

Then the key line: "What is hard and expensive is proving its BOTH safe AND effective in a randomized and controlled study in humans while ALSO manufacturing it at clinical scale and grade."

He's right. And here's the part of Rosie's story that most outlets left out of the headline: a second tumor on the same dog didn't shrink. Conyngham is now sequencing that resistant tumor to understand why. One vaccine, two tumors, two different outcomes. In one dog.

That's not a failure. That's biology being complex. And it's exactly why the gap between "promising result" and "proven therapy" is measured in years of controlled trials, reproducibility studies, and regulatory scrutiny.

This is the part we care about deeply at TPL Genetics. We built our entire evidence classification system around this exact problem. When we label a recommendation as Tier 2 (mechanistic inference) instead of Tier 1 (published pharmacogenomic evidence), we're drawing the same line Heizer is drawing. The line between "this should work based on the biology" and "this has been proven to work in controlled studies."

The difference is: we don't think acknowledging that gap means you stop building. It means you build with rigor. You label your evidence honestly. You seek funding to close the gap with real testing. And you don't present a hypothesis as a certainty.

Heizer is right that making one vaccine is the easy part. But he's answering the wrong question. The question isn't whether this is trivially easy. The question is: what happens when the tools to build it are in the hands of every technically skilled person on the planet, and the ones who take it seriously invest in proving it out?

* * *

Act Two: 108 Out of 111

I know what it feels like to stare at AI-generated results and not quite believe what you're seeing.

My name is Andrei Mellas. I'm the founder of The Peptide List and TPL Genetics. I'm a self-taught data scientist. No biology degree. No computer science degree, either. And over the past 67 days, I've built what I believe is the most rigorous evidence-based genetics platform for peptide optimization that exists.

I used the same family of AI tools that Paul Conyngham used to save Rosie.

The moment it became real for me was in February 2026. I had built a genetic panel: 111 SNPs across 12 peptide-relevant categories. Growth hormone pathways. GLP-1 and metabolic response. CYP drug metabolism. Recovery. Immune function. Cognitive performance. Longevity markers. Body composition. Pain sensitivity. The full spectrum of what determines how your body responds to peptide therapies.

The question I needed to answer: is this panel biologically real? Are these variants actually doing what the published literature says they're doing? Or did I just build a sophisticated-looking system on top of statistical noise?

So I ran the panel against Google DeepMind's AlphaGenome.

AlphaGenome is part of the same family as AlphaFold. Where AlphaFold predicts protein structures, AlphaGenome predicts the functional impact of genetic variants across tissues. It was published in Nature in January 2026. It represents DeepMind's best understanding of which DNA changes actually matter at the molecular level.

108 of 111 variants scored successfully.

Every single scored variant exceeded the 0.9 quantile threshold. Top 10% for functional impact. Across eight pharmacogenomics-relevant tissues: liver, pituitary, adipose, skeletal muscle, brain, hypothalamus, skin. Twelve scoring categories, independently validated.

I sat with that number for a while. 108 out of 111. Independently confirmed by Google DeepMind's genomic foundation model. Not because I asked DeepMind. Not because we have a partnership. Because the science converged.

Paul used AlphaFold to model his dog's cancer proteins. I used AlphaGenome to validate that my genetic panel targets real biology. Same DeepMind AI family. Same outsider approach. Same result: the tools work, and you don't need a PhD to use them.

I wrote a detailed breakdown of that validation if you want to see the full analysis.

What We Actually Built

The AlphaGenome validation was one data point. Important, but one data point. Behind it sits something larger.

TPL Genetics is a platform that analyzes your genetic data and maps it against the most comprehensive peptide evidence base we could assemble. The goal: tell you which peptides your body is most likely to respond to, and which ones might not work for you, based on your actual DNA.

Not based on what worked for your friend. Not based on what a subreddit recommended. Based on your genotype, cross-referenced against published pharmacogenomic research, mechanistic pathways, and clinical evidence.

Here's what sits under the hood, at a high level.

**A multi-model AI research approach.** We don't rely on a single AI. We use multiple frontier models as research scouts: scanning literature, identifying patterns, cross-referencing findings. Then we bring those findings to a separate evaluation layer for synthesis, fact-checking, and integration. Think of it as a research team where each member has different strengths and blind spots, and the final analysis accounts for all of them.

**A three-tier evidence classification system.** Every peptide recommendation in our system is classified by the strength of its evidence. Tier 1: published pharmacogenomic research. Direct, peer-reviewed evidence linking a genetic variant to a peptide response. Tier 2: mechanistic inference. The biological pathway is well-established, the connection is scientifically sound, but the specific peptide-gene pairing hasn't been validated in a dedicated clinical trial. Tier 3: exploratory hypothesis. Emerging research, plausible mechanism, but early-stage evidence.

We label every recommendation with its tier. Transparently. Because the fastest way to lose trust in this space is to present a hypothesis as a certainty.

**A knowledge graph with 54,000 edges.** Genomic variants. Pharmacological pathways. Drug interactions. Tissue expression data. Population frequency data. Functional annotations. Clinical associations. All interconnected. All queryable. All maintained.

This isn't a spreadsheet. It's a living map of how genetics, pharmacology, and peptide science intersect.

**And the numbers that sit on top of all of this:**

All of this was built in 67 days. By a small team. With AI as the force multiplier. You can decide whether that's frightening or amazing.

Why I Build in Public

There's a tension every founder in a competitive space feels. You have three options: overshare and risk someone copying your work. Share strategically and hope the right people notice. Or stay silent, build in stealth, and pray you ship before someone else does.

I've never built anything like this before. I had no playbook. So I just started sharing what I was doing, and every assumption I had about what would happen got blown away.

The inbound has been unreal. The volume and quality of conversations I've been having with researchers, clinicians, investors, and people in the peptide space who actually know what they're talking about. Those conversations have made the platform better in ways I never could have planned for. That is worth the risk of oversharing to me.

Last week, someone compared what I built to SelfDecode, a genomics company with a team of 70 people and over $20 million invested in development. Their point was that this already exists. Fair. But here's the thing: I built a more rigorous report, with one person, for a few thousand dollars. Being capital-inefficient is not something to brag about. That reeks of incompetence, not innovation.

> "So you mean to tell me that one person can compete with, and beat, a well-funded team of 70 people? Imagine what I could accomplish with funding of my own."

Could someone read this article and try to build something similar? Sure. Go ahead. I hope you have a spare 600 hours, because that's what I've poured into this so far. And when you're done, you'll still be starting from zero on the audience, the data, the relationships, and the credibility that only comes from showing your work publicly while everyone else is hiding behind "stealth mode."

Paul Conyngham did the same thing. He documented Rosie's journey on X starting in November 2024. He didn't hide the process. He shared the wins, the setbacks, the $3,000 sequencing bill, the 10-hour drive to Queensland. And when the results came in, the world already knew the story.

Building in public isn't just a growth strategy. It's a trust strategy. In a space where "proprietary algorithm" is often code for "we made this up," showing your work is the single most powerful differentiator.

You can't fake 108 out of 111.

* * *

Act Three: The Barriers Have Collapsed

Paul Conyngham and I are not anomalies. We are the leading edge.

Three years ago, what Paul did would have been impossible. The genomic sequencing would have cost $50,000, not $3,000. AlphaFold wasn't publicly available. ChatGPT didn't exist. The idea that a data scientist could design the computational backbone of a personalized cancer vaccine was science fiction.

Two years ago, what I built would have required a team of bioinformaticians, a pharmacogenomics lab, and millions in funding. Building a 54,000-edge biomedical knowledge graph, validating a genetic panel against a frontier genomic model, classifying peptide evidence across three tiers of rigor. That was institutional-scale work.

Today, the tools are free. Or close to it.

AlphaFold is open-source. AlphaGenome's predictions are published. ChatGPT costs $20 a month. Claude, Gemini, Grok: all available. Open-source genomics pipelines are on GitHub. PubMed is free. ClinVar is free. PharmGKB is free. The raw materials of biomedical research are sitting on the internet, waiting for someone with the technical skills and the stubbornness to put them together.

The question is no longer "can outsiders do serious biotech work?"

Paul answered that question with a shrinking tumor and a dog jumping fences.

The question is: what happens when a million technically skilled people realize these tools are available?

What happens when a machine learning engineer in Lagos builds a genetic screening tool for sickle cell carriers? When a bioinformatics student in Mumbai builds a drug interaction checker that outperforms what most hospitals use? When a retired software engineer in Austin builds a peptide optimization system for veterans?

The regulatory frameworks haven't caught up yet. The institutional gatekeepers are still adjusting. But the tools don't care about credentials. They care about the quality of the question you ask them.

This is the moment. Not next year. Not when some company announces it at a conference. Right now. Today. The convergence of AI and biology is happening in apartments, home offices, and veterinary clinics. Not just in pharma R&D departments.

And the people building at this intersection are not waiting for permission.

What We Don't Know

Honesty matters more than narrative momentum. Here's what we don't know:

**Paul's vaccine is n=1.** One tumor in one dog responded to one vaccine. There was no control group. No blinding. No peer review yet. The results are extraordinary and encouraging, but they are not clinical evidence. They are a proof of concept. A very, very promising proof of concept.

**Personalized mRNA vaccines for humans face enormous regulatory hurdles.** The FDA pathway for individualized therapies is complex, expensive, and slow. What works for a dog under veterinary ethics approval is not directly transferable to human medicine. The science may be ready. The regulation is not.

**AI-assisted research is only as good as its inputs.** Large language models hallucinate. Genomic databases have biases toward European populations. Protein structure predictions have confidence intervals. Every AI-assisted pipeline needs human experts in the loop. Conyngham had Professor Thordarson and Professor Allavena. I built our platform alongside published research, validated databases, and evidence-based practices. The AI accelerates the work. It does not replace the rigor.

**The peptide optimization space is early.** Pharmacogenomics applied to peptide therapy is a young field. The research base is growing rapidly, but there are gaps. We label those gaps clearly in our evidence tier system, and we'll keep updating as the science matures.

What we do know: the tools work. The science converges. And the people willing to build with these tools are producing results that would have been unthinkable five years ago.

**Explore TPL Genetics:** See how your DNA maps to peptide response. Evidence-based. Transparent methodology. Built with AI, validated against DeepMind. Get Your Genetics Report →

If You're Building at This Intersection

TPL Genetics sits at the convergence of three massive trends: the $60 billion GLP-1 market (Goldman Sachs projects $100 billion by 2030), the explosion of consumer genomics, and the AI revolution in drug discovery and personalized medicine.

We've built the infrastructure:

• **827+ providers** tracked and profiled across the peptide industry
• **102 peptides** with comprehensive evidence-based profiles
• **17,000+ unique visitors** and growing, with 300+ daily active users
• **54,000-edge knowledge graph** connecting genomics, pharmacology, and clinical evidence
• **111-SNP genetic panel** validated against Google DeepMind's AlphaGenome at 97.3%
• **Three-tier evidence classification** that sets the standard for transparency in peptide recommendations
• **48 evidence-based articles** driving organic traffic and establishing authority

The weight loss peptide market is exploding. Semaglutide and tirzepatide are household names. But personalized peptide therapy, matching the right compound to the right patient based on their genetics, is the next frontier. We're building it.

If you're an investor or builder working at the intersection of AI and biotech, I'd like to talk. Not a pitch deck meeting. A conversation about where this space is going and how we can build something significant together.

**Reach out:** andrei@thepeptidelist.com

The Moment

Paul Conyngham spent $3,000 and twelve months. He used ChatGPT as a tutor, AlphaFold as a microscope, and his own ML skills as the engine. He built something that professors at two Australian universities validated and administered. A tumor the size of a tennis ball shrank by 75%. A dog named Rosie is jumping fences again.

I spent 67 days and every AI tool I could get my hands on. I built a knowledge graph, a genetic panel, an evidence classification system, and a platform that 17,000 people have used to make better decisions about peptide therapy. Google DeepMind's own model confirmed that 108 of our 111 genetic variants target real biology.

Neither of us has a biology degree.

Both of us have results.

The tools are here. The barriers are gone. The only question left is what you're going to build.

FAQ

**Did ChatGPT actually design the cancer vaccine?**

No. ChatGPT served as a research assistant and biology tutor for Paul Conyngham. It helped him understand cancer immunology, plan his computational pipeline, and navigate unfamiliar biological literature. The actual mRNA vaccine was designed and produced by Professor Pall Thordarson's team at the UNSW RNA Institute. AI accelerated the research; human experts built the vaccine.

**Is Rosie cured?**

Conyngham himself says he does not believe the treatment is a cure. The tumor shrank significantly (50-75%), and Rosie's quality of life improved dramatically. She's received two injections with a third planned. But this is a single case, not a clinical trial, and long-term outcomes are unknown.

**What is AlphaGenome and how is it different from AlphaFold?**

Both are AI models from Google DeepMind. AlphaFold predicts 3D protein structures from amino acid sequences. It won the 2024 Nobel Prize in Chemistry. AlphaGenome predicts the functional impact of genetic variants across human tissues. It was published in Nature in January 2026. Paul used AlphaFold for protein modeling; TPL Genetics used AlphaGenome for genetic panel validation. Same AI family, different applications.

**What does "108 out of 111" mean?**

TPL Genetics built a panel of 111 genetic variants (SNPs) relevant to peptide response. When validated against AlphaGenome, 108 of those variants scored successfully, with every scored variant in the top 10% for functional impact across eight tissues. This independently confirms that the panel targets biologically meaningful genetic variations. Read the full validation analysis here.

**How does TPL Genetics use AI?**

We use multiple AI models as research tools: scanning scientific literature, identifying patterns in genomic data, cross-referencing findings across databases, and synthesizing evidence. AI accelerates work that would take a traditional research team months. But every recommendation is grounded in published research and classified by evidence strength. AI is the force multiplier, not the decision-maker.

**What is the evidence tier system?**

Every peptide recommendation in TPL Genetics is classified into one of three tiers. Tier 1: published pharmacogenomic research with direct clinical evidence. Tier 2: mechanistic inference based on well-established biological pathways. Tier 3: exploratory hypothesis based on emerging research. We label every recommendation transparently so you know exactly how strong the evidence is.

**Can I use TPL Genetics to find out which peptides are right for me?**

Yes. TPL Genetics analyzes your genetic data and maps it against our evidence base to show which peptides your body is most likely to respond to based on your DNA. The report includes evidence classifications, pathway explanations, and actionable insights. Visit our provider directory to find a qualified provider.

**Is this type of AI-assisted biotech research safe?**

AI-assisted research requires the same rigor as traditional research, plus additional validation to account for AI limitations like hallucination and data bias. Paul Conyngham's vaccine went through a 100-page ethics application and three months of review before a single injection was administered. TPL Genetics validates findings against published databases, peer-reviewed literature, and independent models like AlphaGenome. The tools accelerate research; they don't replace the safeguards.

* * *

_Disclaimer: This article is for educational and informational purposes only. It does not constitute medical advice, diagnosis, or treatment. Peptide therapies should only be pursued under the supervision of a qualified healthcare provider. The case of Paul Conyngham's dog represents a single anecdotal outcome under veterinary ethics approval and should not be interpreted as clinical evidence for human medicine. TPL Genetics provides genetic information and educational content; it does not prescribe or recommend specific treatments. Always consult with a healthcare professional before making decisions about peptide therapy. Past results do not guarantee future outcomes._

DNA Kit

Want to know which peptides match your genetics?

Our $250 DNA kit analyzes 111 SNPs across 15 pharmacogenes plus peptide-response markers. Get a personalized report on which peptides fit your biology.

Get your DNA kit

Free PDF Guide

Want the complete picture?

Download our free 56-page Essential Guide to Peptides. Covers the science, safety, and how to pick a legit provider.

Get the Free Guide

No spam. Unsubscribe anytime. We respect your privacy.

Ready to Find a Provider?

Browse our directory of verified peptide providers.

Browse Providers

Related Articles

**We Tested Our Genetic Panel Against DeepMind's AlphaGenome. 108 of 111 Held Up.** \\
\\
108 of 111 variants. 8 tissues. Every scored SNP in the top 10% for functional impact. Independent confirmation that our panel targets real biology. **Which Peptides Actually Work? Evidence-Based Guide (2026)** \\
\\
520 adverse events. 17 deaths. $70 billion in sales. One Zagreb lab. And the evidence hierarchy that separates real science from Reddit mythology. **FDA Peptide Regulations 2026: Is Your Peptide Legal?** \\
\\
A comprehensive analysis of current FDA peptide restrictions, enforcement reality, and future regulatory changes **Peptides for Weight Loss: 5 Expensive Mistakes to Avoid (2026)** \\
\\
The GLP-1 revolution changed everything,but most people are still making the same expensive mistakes

This website is for informational purposes only and does not constitute medical advice. Consult a licensed physician before using any peptides. Provider listings do not constitute endorsements. None of the statements on this site have been evaluated by the FDA.