Why I built DecodeMyBio

Eight weeks at a time

“Let’s start with this one and see how you feel in a couple of months.”

I heard some version of that sentence for years.

The arithmetic was always the same. Start a medication. Wait six to twelve weeks for it to reach steady state and for anything real to surface. Notice it isn’t right. Taper off. Start the next one. Wait again.

What makes that cycle so expensive isn’t the waiting. It’s that a “no” at the end of it tells you almost nothing. It doesn’t tell you why the drug didn’t work, or whether the next one will fail for the same reason, or whether the problem was ever the drug at all. It costs you two months and returns you to the start with no more information than you had going in.

This is not a story about bad doctors. Every clinician I saw was working from what was on the table — symptoms, history, and a prescribing order that begins with whatever works for most people. That is the correct way to practise medicine when you don’t have anything else. I just wasn’t most people, and nobody had a way to check. Whether the data would have changed anything, I don’t know. It was never on the table.

Coffee made me sleepy. Not “failed to wake me up” — actively, reliably sleepy. For years that was filed under personality quirk, the kind of thing you mention at a dinner table and everyone laughs and nobody follows up. I still don’t have a confirmed answer for it, and I want to be careful about that: caffeine response is one of the areas where the popular genetic explanations run a long way ahead of the evidence. I keep it in the list because of what didn’t happen, not because I solved it.

And some years before any of it, a statin. My legs ached. Not dramatically — a dull, constant ache that made stairs annoying and made me feel like I’d aged a decade in a month. I stopped taking it. Nobody asked why, and I didn’t push, because “my muscles hurt” sounds like a complaint rather than a finding. It went into my file as intolerance, and that was the end of it.

It was years later that I learned muscle pain on a statin is one of the most-studied drug reactions in pharmacogenomics — that a gene called SLCO1B1 governs how efficiently statins are taken up by the liver, that there are published clinical guidelines describing what to do when a patient carries the reduced-function variant, and that “he didn’t get on with statins” was not the end of a story but the beginning of one nobody told me.

When I eventually read my own raw data, rs4149056 was there.

I can’t prove that’s why my legs ached. Nobody was testing for it then, plenty of people get muscle pain on statins without it, and I will never know for certain. What I know is that there was a question worth asking, and nobody asked it.

What connects the antidepressants, the coffee and the statin isn’t a diagnosis. It’s a missing question. In all three cases, the question nobody asked was whether my body could process the compound in the first place.

The same file, twice

Eventually I got sequenced — the ordinary consumer route, the same one most people take.

What came back was a tidy summary and a pie chart about my ancestry. But underneath it, available for download if you go looking, was the raw data file: hundreds of thousands of genetic markers, sitting in a text file on my laptop, that nobody had ever read to me.

So I fed it to an AI.

It was extraordinary, and that’s why I kept going. For the first time, something was reading the actual file and telling me about me — mechanisms, pathways, reasons. It felt like a level of self-knowledge I hadn’t known was available.

Then I ran the same file again.

I got a different answer.

Not wildly different. Different in the way that matters — a shifted emphasis, a claim that had softened, something present the first time that was simply gone the second. My DNA had not changed between Tuesday and Wednesday. The file was byte-for-byte identical. The output was not.

And once I started checking properly, I found the worse problem underneath the first one. Some of what it told me was wrong — and it was wrong in exactly the same calm, organised, confident voice as everything it got right. There was no tell. Nothing in the writing distinguished a well-established pharmacogenomic finding from something the model had assembled because it sounded plausible.

That’s the part that frightened me. Wrong-and-uncertain is survivable; you go and check. Wrong-and-confident is how somebody stops taking a medication they need.

So I went and found out why it happens, and the answer turned out to be structural rather than fixable. A language model generates the most plausible next piece of text. It is not, at that moment, looking anything up. Genetics has ground truth — a given marker either is or isn’t in your file, and the published evidence about it either exists or doesn’t. Those two facts are fundamentally incompatible. You can make a model sound more careful. You cannot make generation into retrieval.

That was the moment DecodeMyBio stopped being a thing I was doing for myself.

I’m not a geneticist

I should say that plainly, because you’d be right to ask.

I’m not a geneticist, a pharmacist, or a physician. I’m someone who spent years on the receiving end of trial-and-error medicine, learned to read his own raw data, and found that the tool everyone reaches for to do that is quietly unreliable in a way most people will never notice.

And that is precisely why DecodeMyBio is built the way it is. Because I couldn’t rely on my own judgment, I had to build something that doesn’t rely on anyone’s judgment — mine included.

Nothing Decode tells you is my opinion. Concretely:

Every finding in Decode is traceable to a published guideline. Results come from CPIC and DPWG drug–gene guidelines, with allele definitions from PharmVar, variant classifications from ClinVar, and regulatory context from the FDA’s table of pharmacogenomic biomarkers in drug labeling. Each finding names the guideline it came from, and your Decode report records the CPIC dataset version and ClinVar release it was built against — so you can see not just what I say, but what I read to say it.

Your Decode result is looked up, not generated. Your star alleles, your metabolizer phenotype and the guideline that applies to it come from matching your file against curated evidence tables. No language model decides what your result is. The assistant that answers your follow-up questions is a language model, and it is fenced to the findings the lookup already produced — it can phrase them, it cannot invent them.

Your Decode answer is stable until the evidence moves — and then you hear about it. Re-running the same file does not produce a new opinion. What can change is the guideline underneath it: CPIC revises a recommendation, ClinVar reclassifies a variant. When that happens I re-run your stored results against the new version and show you the difference — what changed, what didn’t, and when. That is the opposite of getting a different answer on Wednesday for no reason.

Where the science is unsettled, your Decode result says so. A great deal of what circulates online about genetics is far more confident than the underlying evidence supports. Results carry the strength of the recommendation behind them and the review status of the classification — including “uncertain significance” and “conflicting classifications,” which is often the honest answer. Some of the most popular genetic claims on the internet are among the weakest, and I’d rather disappoint you accurately than satisfy you incorrectly.

When I get something wrong, I publish it. Evidence moves, guidelines get revised, and I have shipped mistakes — a citation pointing at the wrong paper, a drug–gene interaction described backwards, guidance attributed to the wrong guideline body. They’re in the corrections log with dates, rather than quietly edited away.

What DecodeMyBio does not do

• It does not diagnose anything.
• It does not tell you to start, stop, or change any medication or dose. Those decisions belong to you and the person who prescribes for you, and genetics is one input among many — alongside your kidney and liver function, your other medications, your history, and your symptoms.
• It does not predict disease, and it is not a clinical or diagnostic test.
• It does not confirm what it finds. Consumer arrays read a subset of the variants that define each star allele; where a defining variant isn’t genotyped, it’s assumed to be the common one. That’s standard practice for this kind of analysis, and it means rare alleles can be missed. I’ve written up the limits in full on the methodology page.
• It does not replace clinical pharmacogenomic testing. If a decision is genuinely load-bearing, that testing exists and is the right tool.

What it is for is turning a file you already own into questions worth bringing to the person who prescribes for you.

Your data

Your DNA file is about as personal as data gets, so I’d rather tell you how it’s handled in my own words than make you go and find the policy.

It’s encrypted in transit and at rest. It’s processed server-side, and no other user can reach it. I don’t sell it, rent it, or share it. I don’t use it for research. I don’t use it to train anything. It does not go to insurers or employers, and it never will — that isn’t a position I intend to revisit. You can delete it from your account whenever you want, and if you’d like a copy of your results, ask and I’ll send one.

What I do use, and I’d rather you heard it from me than found it in clause nine: this site runs Google Analytics and the Google Ads tag. They set cookies. They tell me how many people arrive and which ads actually work. And if you come here from a Google ad, a hashed — irreversible — version of your email address is shared with Google so a sale can be matched to the click that produced it.

None of that touches your genetic data. Your file and your results are never used for analytics, never used for advertising, and never shared with anyone for either purpose. The measurement is about traffic and ads. It is not about your DNA.

The full detail is in the privacy policy.