I had a quick look at Helisoma. The concept is pretty interesting — especially being able to take existing raw DNA data and turn it into something more practical instead of just staring at a huge list of genetic markers 😅
I also like that the result isn't locked inside another app and can be used with tools people already use, like ChatGPT or Claude.
The privacy side would probably be my biggest question before uploading something as sensitive as DNA data, but I appreciate that they address deletion and data ownership on the site.
Overall, definitely one of the more unusual AI + health ideas I've come across recently.
Regarding privacy, let’s do some math. Your genome is ~3.2 billion base pairs. A consumer DNA chip reads about 650k of those positions, so the file you have covers only ~0.02% of your genome. From that file, Helisoma extracts, sends and stores on the backend just a few hundred markers (and we don’t store the file itself), which works out to around 0.00001% of your genome — a hundred-thousandth of a percent. We don’t need more data because for the vast majority of positions there’s simply no solid, replicated research linking them to everyday wellness traits yet — science can read the letters, but for most of them it can’t say what they mean for you.
And even for that 0.00001%, you can delete your data, or even your whole account, at any time. It’s a hard delete: the records are actually removed from the database, not just flagged as deleted while sitting there forever.
The strongest part is the emphasis on evidence quality. You’re not presenting the DNA output as equally reliable across findings, which makes the product’s trust problem much more visible.
This was actually my biggest fear when designing the report. So, curious, when you see the low-confidence sections, did the honesty make you trust the rest more, or does it make you wonder why those findings are in the report at all? Still not sure about it
From the interviews I've done so far, users split into two camps: those who trust every number without asking about sample size or replication, and those who trust nothing and verify each line. Interestingly, both wanted the low-evidence findings included, cause a lot of the fun traits people are most curious about (chronotype, caffeine response, etc), sit exactly in that zone
The thing I found most interesting was that the problem wasn’t with the LLM. ~
It worked with information that had a certain format.
It's easy to overlook that difference. A model can adequately explain something but may still give fluctuating replies if the underlying data is messy or troublesome.
I also like that decision, because it gives us a sense of the strength of the evidence, rather than every finding being equally reliable. This is especially important when people seek answers to make decisions about themselves.
There is a larger lesson here too: sometimes, the opportunity is not to build smarter models. The model will find it easier to understand the information.
I took a 23andMe test more than ten years ago and had no idea how to make use of the results afterward. This is a great way to keep using that data. Thank you! It’s a fascinating product.
About
We track everything: sleep, HRV, steps. But that data describes what happened, not what your body is built like. What if your DNA joined the picture and your routine was planned around your own genes?
13 Comments
I had a quick look at Helisoma. The concept is pretty interesting — especially being able to take existing raw DNA data and turn it into something more practical instead of just staring at a huge list of genetic markers 😅
I also like that the result isn't locked inside another app and can be used with tools people already use, like ChatGPT or Claude.
The privacy side would probably be my biggest question before uploading something as sensitive as DNA data, but I appreciate that they address deletion and data ownership on the site.
Overall, definitely one of the more unusual AI + health ideas I've come across recently.
Thank you, really appreciate the kind words!
Regarding privacy, let’s do some math. Your genome is ~3.2 billion base pairs. A consumer DNA chip reads about 650k of those positions, so the file you have covers only ~0.02% of your genome. From that file, Helisoma extracts, sends and stores on the backend just a few hundred markers (and we don’t store the file itself), which works out to around 0.00001% of your genome — a hundred-thousandth of a percent. We don’t need more data because for the vast majority of positions there’s simply no solid, replicated research linking them to everyday wellness traits yet — science can read the letters, but for most of them it can’t say what they mean for you.
And even for that 0.00001%, you can delete your data, or even your whole account, at any time. It’s a hard delete: the records are actually removed from the database, not just flagged as deleted while sitting there forever.
The strongest part is the emphasis on evidence quality. You’re not presenting the DNA output as equally reliable across findings, which makes the product’s trust problem much more visible.
This was actually my biggest fear when designing the report. So, curious, when you see the low-confidence sections, did the honesty make you trust the rest more, or does it make you wonder why those findings are in the report at all? Still not sure about it
That’s exactly the tension I was getting at. I’m curious what you hear from the next few users on that tradeoff.
From the interviews I've done so far, users split into two camps: those who trust every number without asking about sample size or replication, and those who trust nothing and verify each line. Interestingly, both wanted the low-evidence findings included, cause a lot of the fun traits people are most curious about (chronotype, caffeine response, etc), sit exactly in that zone
That split is interesting. Would be happy to continue the conversation — what’s the best email to reach you on?
analyzed my dna test from myheritage. got a lot of new info!
Glad you liked it!
Great product. Keep it up!
thanks a lot!
The thing I found most interesting was that the problem wasn’t with the LLM. ~
It worked with information that had a certain format.
It's easy to overlook that difference. A model can adequately explain something but may still give fluctuating replies if the underlying data is messy or troublesome.
I also like that decision, because it gives us a sense of the strength of the evidence, rather than every finding being equally reliable. This is especially important when people seek answers to make decisions about themselves.
There is a larger lesson here too: sometimes, the opportunity is not to build smarter models. The model will find it easier to understand the information.
I took a 23andMe test more than ten years ago and had no idea how to make use of the results afterward. This is a great way to keep using that data. Thank you! It’s a fascinating product.