not my field so I can't judge how useful it is but assuming this data can be used for drug discovery and their ToS limiting to non-commercial use only. Does DeepMind plan on selling this data to pharmaceutical companies?
the Google DeepMind PR team can't catch a break, shame it makes no sense to me - if someone's in the field maybe they could explain to the rest of us if this is a big deal, just a PR move, or nah?
I think orliesaurus meant the PR team is unable to rest because the rest of Deepmind is so prolific and keep coming out with new things for them to announce
Why not? There is no logical reason as to why this would not work, IF it works in the first place, which I don't know. In theory the problem space here is finite, so there is of course a way to predict everything. Whether this is the case right now - who knows; I probably don't think it is currently ready. But eventually it will be. And it should not be in the hands of private companies.
So it sounds like you're coming to this with very little knowledge about biology. I encourage reading up on modern challenges in pathogenic prediction, especially with regards to SNPs: on their own, with the exception of a few diseases, individual SNP predictions are meaningless in terms of actual pathogenicity.
23andMe and similar companies don't transcribe your entire genome because that would cost way more than they charge you. They just sample a few tiny sections of it.
I did fair amount of research, and it seems like they take it seriously with GDPR, HIPAA etc. and all the privacy legalities, but obviously Uncle Sam will have his way with it as it's a US based company (California). They will share your DNA if legally compelled.
There is a "permanent delete my data" option, but there's no independent audit that permanent delete actually deletes anything. If you're in EU/California they kinda have to take delete seriously due to GDPR and CCPA, so I think you're safer here.
I've been thinking about this a lot, and I'm not sure how "protectable" your DNA actually is because of how easy it is to get if somebody wants to. Obviously having your DNA in US government is bad, but they likely have it already.
As I'm an EU citizen I will get the data and use the delete option with fingers crossed that US government doesn't want to destroy me for whatever reason. Seems like they have their hands full with other kinds of destruction anyway lol
23andMe tests SNP's (single nucleotides) that are inferred to be significant in protein function/epigenitics.
Those SNP's i believe are testd from primers
so what 23andMe does is specifically on the back of previous research and afaik their data isnt technically clinically significant as most findings need confirmation or more tests.
Probably not any 23andMe haven't already told you about. They test a limited set of SNPs, balancing between ones thought useful for genealogy, ones useful for ethnicity estimates and ones thought useful for health-related things (the latter they would like to make their main selling point, the two former are really all commercial DNA services' bread and butter).
It's unlikely that they would luck into testing some unknown SNP which turned out to be relevant for disease.
This could be especially useful for people who get a VUS on a genetic test. I’m an ML engineer, not a biologist, and the precomputed Atlas meant I could just look up my own VUS and inspect the outputs directly, without setting up a genomics pipeline myself. I found it really helpful for understanding my variant better. I can also see how easy it would be to over-interpret the output, but I still think there’s a lot of value in lowering the barrier for patients to inspect and understand their own variants.
I saw a really interesting talk by Katie Pollard at ISMB this year about the limitations of variant prediction.
The gist was, can existing variation provide enough context to infer impact of variation? The answer seemed to be no.
Kind of like how frontier LLMs need to ingest larger and large amounts of text to advance. We are going to need to leverage comparative data from other species, and likely tremendous amounts of laboratory mutagenesis experiments to actually make headway on variant prediction. Nature, as it stands, just doesn't have enough human variation.
Yes, and genetic variant generally do not act in isolation. We currently focus on the small additive effects of variants because we can with small sample sizes—and 1 million humans is marginal using a GWAS cohort to dive into epistasis. But these interaction effects among variants are critical. Now almost completely deprecated.
That's an interesting statement: "The gist was, can existing variation provide enough context to infer impact of variation? The answer seemed to be no."
Is this saying that if we were to sequence every human being on the planet, we'd still be unable to explain some phenotype differences caused by variation simply becase there aren't enough humans/enough variation? Interesting, as that's the first time I've heard that claim, and it would suggest that we spend our time working on mechanistic models of variant to phenotype.
This has Demis written all over it. There is a great video of him with AlphaFold chatting with the team about releasing some results, and he asked something like “what if we just do them all?”
This has nothing to do with AlphaFold at all- not sure if you were implying that.
(the scientific contriution is welcome, but it's not particularly significant)
I don’t think Demis played a big role in this. It was mainly Ziga Avsec who developed Enformer (the first actually decent sequence-to-function model), and then AlphaGenome.
Imagine cellular activity as an industry zone, its not just what you can or can not make, its also 'for what concentrations of chemical species, what transcription rates should be used' so apart from the discrete Mendelian aspects (like what eye color or what have you) there is also a concensus sequence and deviations from consensus. They mention the dataset captures non-coding DNA, which should imply promoter sequences. Will it be possible to query the atlas for joint probabilities of promoter and putative target protein occurence in human genomes?
Personalized medicine could never credibly take off as long as promoter sequences were excised before sequencing!
I don't want to drag the discourse away from this achievement, but I hate how this is announced with blatant corporate advertising (our internal model, here are the benchmarks, gpt astra TM yours now for the low low price of £200pcm). I just didn't think Navier-Stokes falling would be sponsored by McDonald's.
Still. I am crying right now. Navier-Stokes is solved.
That study has done in reality what the AlphaGenome Atlas does in fiction, but instead for a human they have done it for one of the simplest viruses.
So they have fuzzed the virus by mutating one by one each position of its DNA.
And various dedicated AI models all made poor predictions of the results of that experiment, which casts doubts about the value of the AlphaGenome predictive map.
A virus is much simpler than a human, but even for that simple virus the effects of most of the mutations could not be predicted. A half of the mutations had harmful effects, and for a half of those it is unknown for now why they were harmful.
For a human the uncertainty about the effects of a mutation will be far greater than for one of the simplest viruses.
Yep. Sequence-to-function models are still very limited. AlphaGenome Atlas, despite the flashy branding, is unlikely to provide significant benefit to researchers.
And it makes a lot of sense why they are limited. DNA is not an instruction set. It's more like a heavily encrypted dataset where the encryption key is the totality of physics and biology. The interactions with the physical world that result in the end product of life are enormously (it would seem hopelessly) complex.
For a machine intelligence to turn DNA sequences into organisim phenotype prediction requires modelling all that in latent space.
I imagine that is going to take a monumental amount of example data
It seems like we can skip much of the expensive modelling and use evolutionary conservation data to shortcut building a full latent space that captures all salient interactions. It's unclear to me whether we truly need to model the entire latent space. And given that biology develops in a generative way with feedback, it may be that attempting to model this using a static latent space is unproductive.
> And various dedicated AI models all made poor predictions of the results of that experiment, which casts doubts about the value of the AlphaGenome predictive map
I would not group AlphaGenome into the pile of failed predictions of other models. AlphaGenome deserves to get evaluated based off its own merits.
Why should we discount this model just because other models that are already thought to be worse weren’t very accurate? I’m not in this field at all, so curious if I’m missing something here.
Just because a virus genome is small doesn't make it simple, actually quite the opposite (think of it as an obfuscated, compressed package, that is hugely variable with no checksum). In the case of this virus there are even overlapping open reading frames, which is something you would never find in a eukaryotic genome, and almost the entire genome is protein coding, whereas only 2% of the human genome is protein coding.
The real value of AlphaGenome is not the effect of SNPs on protein coding regions (there are other tools for that, like AlphaFold), but rather identifying regulatory elements, such as promoters or alternative splicing patterns or microRNAs, within the ~98% of the genome that hasn't been well characterised yet.
That 98% is vastly underexplored, so a tool like this could help researchers interested in expression profiles or alternative splicing patterns of a protein, identify the source. Obviously not every "important" SNP will be consequential, but it helps narrow the search for that needle in a haystack.
There are AI labs headed by marketing CEOs that fake metrics, have an utter disregard for humanity and make up "AGI is imminent" propaganda for their IPO, and then there are AI labs headed by actual scientists that do actual science for humanity without constantly trying to put themselves into the spotlight.
89 comments
[ 64.9 ms ] story [ 922 ms ] thread1. https://www.youtube.com/watch?v=U0aToL5C-bQ
2. https://www.youtube.com/watch?v=b2qw3rDNX0Q
https://deepmind.google/blog/alphagenome-atlas-a-predictive-...
They are only announcing a cache. The origin for the cache is not discussed.
In particular, the question of whether to trust the predictions is not addressed. For that, I think the citation is from January:
https://www.nature.com/articles/s41586-025-10014-0
I'm comfortable programming but others who also do mol bio may be less so or may not recognize when Claude is going off the rails.
23andMe used a custom Illumina Infinium microarray.
https://www.illumina.com/products/by-brand/infinium.html
There is a "permanent delete my data" option, but there's no independent audit that permanent delete actually deletes anything. If you're in EU/California they kinda have to take delete seriously due to GDPR and CCPA, so I think you're safer here.
I've been thinking about this a lot, and I'm not sure how "protectable" your DNA actually is because of how easy it is to get if somebody wants to. Obviously having your DNA in US government is bad, but they likely have it already.
As I'm an EU citizen I will get the data and use the delete option with fingers crossed that US government doesn't want to destroy me for whatever reason. Seems like they have their hands full with other kinds of destruction anyway lol
Those SNP's i believe are testd from primers
so what 23andMe does is specifically on the back of previous research and afaik their data isnt technically clinically significant as most findings need confirmation or more tests.
It's unlikely that they would luck into testing some unknown SNP which turned out to be relevant for disease.
The gist was, can existing variation provide enough context to infer impact of variation? The answer seemed to be no.
Kind of like how frontier LLMs need to ingest larger and large amounts of text to advance. We are going to need to leverage comparative data from other species, and likely tremendous amounts of laboratory mutagenesis experiments to actually make headway on variant prediction. Nature, as it stands, just doesn't have enough human variation.
Is this saying that if we were to sequence every human being on the planet, we'd still be unable to explain some phenotype differences caused by variation simply becase there aren't enough humans/enough variation? Interesting, as that's the first time I've heard that claim, and it would suggest that we spend our time working on mechanistic models of variant to phenotype.
Very excited to see that happen here.
That is outright scary. Science is being slurped up here.
Imagine cellular activity as an industry zone, its not just what you can or can not make, its also 'for what concentrations of chemical species, what transcription rates should be used' so apart from the discrete Mendelian aspects (like what eye color or what have you) there is also a concensus sequence and deviations from consensus. They mention the dataset captures non-coding DNA, which should imply promoter sequences. Will it be possible to query the atlas for joint probabilities of promoter and putative target protein occurence in human genomes?
Personalized medicine could never credibly take off as long as promoter sequences were excised before sequencing!
Also Mendelian has little to do with promotor sequences or differential transcription in deviations.
I don't want to drag the discourse away from this achievement, but I hate how this is announced with blatant corporate advertising (our internal model, here are the benchmarks, gpt astra TM yours now for the low low price of £200pcm). I just didn't think Navier-Stokes falling would be sponsored by McDonald's.
Still. I am crying right now. Navier-Stokes is solved.
https://www.science.org/content/blog-post/mutate-em-all-and-...
which comments the results of this study:
https://www.biorxiv.org/content/10.64898/2026.07.25.740675v1
That study has done in reality what the AlphaGenome Atlas does in fiction, but instead for a human they have done it for one of the simplest viruses.
So they have fuzzed the virus by mutating one by one each position of its DNA.
And various dedicated AI models all made poor predictions of the results of that experiment, which casts doubts about the value of the AlphaGenome predictive map.
A virus is much simpler than a human, but even for that simple virus the effects of most of the mutations could not be predicted. A half of the mutations had harmful effects, and for a half of those it is unknown for now why they were harmful.
For a human the uncertainty about the effects of a mutation will be far greater than for one of the simplest viruses.
For a machine intelligence to turn DNA sequences into organisim phenotype prediction requires modelling all that in latent space.
I imagine that is going to take a monumental amount of example data
I would not group AlphaGenome into the pile of failed predictions of other models. AlphaGenome deserves to get evaluated based off its own merits.
The real value of AlphaGenome is not the effect of SNPs on protein coding regions (there are other tools for that, like AlphaFold), but rather identifying regulatory elements, such as promoters or alternative splicing patterns or microRNAs, within the ~98% of the genome that hasn't been well characterised yet.
That 98% is vastly underexplored, so a tool like this could help researchers interested in expression profiles or alternative splicing patterns of a protein, identify the source. Obviously not every "important" SNP will be consequential, but it helps narrow the search for that needle in a haystack.