Using AI to Interpret Your Genetic Data? Be Careful
Article at a Glance
- Nutrigenomics and clinical genetics are particularly high-risk domains for AI errors.
- AI tools can help with basic gene calls, like identifying the presence of MTHFR or COMT SNPs or whether a user carries a given “mutation” of interest.
- These tools are high risk for anyone looking for deep insight into their genome. The tools can hallucinate confidently in ways that can lead people down the wrong path.
At a birthday party for my daughter’s friend a few weeks ago, and making small talk about kids and technology, I asked a Mom her concern level for the future of AI. She was at a 7 on a 10 scale.
I am around 3 or 4.
AI is a tool we use at Gene Food and it enhances our ability to serve customers. I’m proud of the AI meal planner we built which allows users to map out weeks of personalized meal plans based entirely on their genetic data and labs. It’s cool, and it’s useful.
All of us are hopeful that AI tools will advance the ability to treat disease and open up new horizons across many industries. So, put me, at least partially, in the techno optimist camp with Andreesen and Horowitz. There seems to be good things on the horizon, but I don’t buy the AGI stuff at all, and look, we still need to have human oversight of these tools.
Case in point is the interpretation of sensitive genetic data. We have written previously on the blog about why you might think twice about uploading raw genetic data to ChatGPT, and today I want to offer a concrete example of how an AI tool can get a subtle detail wrong, and do so confidently, which can throw off the ability to map the genome in a truly sophisticated way.
Because I believe that much of what we see with the carnivore diet and very strict elimination diets are rooted in environmental sensitivity, we are doing more with detox at Gene Food. For example, even our polygenic risk score on testosterone predispositions factors in an NR2F2 SNP.
Unlock Your Personalized Nutrition & Supplement Report
Gene Food uses a proprietary algorithm to divide people into one of twenty diet types based on genetics. We score for cholesterol and sterol hyperabsorption, MTHFR status, histamine clearance, carbohydrate tolerance, and more. Where do you fit?
As we researched the GWAS research on genotypes that may show greater sensitivity to pesticides like glyphosate, some SNPs started to emerge. Our AI tools told us that PON1 Q192R (rs662) was a good candidate for a glyphosate detox analysis, but there is one big problem that our geneticist flagged in our manual review process.
The short answer is the AI was wrong, because rs662 isn’t linked with glyphosate toxicity.
The longer answer is that this is indeed an important variant for organophosphate oxon metabolism – it’s just that glyphosate isn’t an oxon. Therefore, we could broaden the scope of the report to ‘organophosphate pesticides detox,’ and include the SNP, but the original AI framing was off. The issue is so complex, 99.99% of users wouldn’t know to question the AI, they would go about their lives with the incorrect interpretation.
When I asked the AI, why they messed up, here is what they said.
The AI tool made a classic hallucination-by-plausible-association error. This is a textbook example of why nutrigenomics and clinical genetics are particularly high-risk domains for AI errors. The model connected two true facts through a superficially logical bridge that breaks down at the biochemistry level. It knew the category membership (organophosphate) but missed the functionally critical subcategory distinction (oxon vs. non-oxon substrate specificity).
I won’t name the AI that gave us the hallucination, but it’s an advanced model. The message is important, by the AI’s own admission, genetics is an area that is high risk for misinterpretation.
You can’t just dump thousands of SNPs of raw data into an AI and trust the outcome. If you are using AI tools hoping for deep insights into genetic data, tread with caution.
