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!
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.
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.
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.
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.
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.
Alpha Fold has continued to impact the field of protein networks, but I do hear that not every one of the deep learning biology models from Google/Deep Mind and others have made equivalent impact or had as lasting relevance in their respective domains..some have performed more poorly than other available models. I'd love to learn more about this, but this has mostly come from little snips of conversations here and there, in person and online, but I haven't seen anything comprehensive in terms of evaluating their impacts overall
Is this just Google precomputing Alpha genome values - which were already accessible via API and making them available as another API (presumably more broadly)? Or is there actually new information?
People are upvoting this because it has the “Alpha______” prefix. Meanwhile, everyone in the field of genomics knows that AlphaGenome provides essentially zero improvements over the previous SOTA, Borzoi…
> a database that predicts the effects of every possible single nucleotide variant in the human genome. We used the AlphaGenome AI model to pre-calculate the regulatory impact of all 9 billion single-letter genetic changes, resulting in a massive, 1-petabyte dataset.
This is for a database, no? While Borzoi is a model?
> Here, we introduce Borzoi, a model that learns to predict cell-type-specific and tissue-specific RNA-seq coverage from DNA sequence.
> Meanwhile, everyone in the field of genomics knows that AlphaGenome provides essentially zero improvements over the previous SOTA, Borzoi…
Can you elaborate on this? I'm confused why Google would build something that provides zero improvements over SOTA, Borzoi... as you mention. I'm not familiar with this field, just curious.
Speaking as somebody who has worked within Google Research before: the researchers are under tremendous pressure to publish SOTA and sometimes they juice their results a bit to look competitive when they can't match. This is not uncommon in the field- it's remarkably easy to edit a paper to make yourself look good by omitting information.
The use of an exact quote in an ungrammatical fashion is a bit of a language model smell. I can’t help but be reminded of the purely nonsensical AI interview answers. “It’s a pleasure to meet you, Chick Bongo”
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.
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?”
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.
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 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.
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.
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.
It's "shareholder-linked" (probably not "driven"), but yes, upvoted. Nonetheless, even when with grave faults in the path, Google has managed to give us marvellous, paramount free resources (Google Street, Google Museum...), and would have given more (Google Books - if the negotiation had succeeded). I am grateful.
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.
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
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?
Don't be put off by the box asking for your "affiliation". I wrote "None", clicked submit and it took me straight to the Atlas.
Videos. [2] is for the scientists to start using AlphaGenome Atlas from AntiGravity.
1. https://www.youtube.com/watch?v=U0aToL5C-bQ
2. https://www.youtube.com/watch?v=b2qw3rDNX0Q
Not a word about promoter sequences.
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!
Eye color is not discrete Mendelian. That's only correct to the first order.
Also Mendelian has little to do with promotor sequences or differential transcription in deviations.
In another HN thread about this AlphaGenome Atlas, someone has posted a link to:
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.
Yep. Sequence-to-function models are still very limited. AlphaGenome Atlas, despite the flashy branding, is unlikely to provide significant benefit to researchers.
May be that's the reason for the alpha naming. We are waiting for a stable release (just kidding).
Can this be used with a 23andMe genome to find pathogenic mutations?
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.
A few as in tens of thousands.
23andMe used a custom Illumina Infinium microarray designed around segments of particular interest.
https://www.illumina.com/products/by-brand/infinium.html
fwiw sequencing your entire genome only costs $400 or so from providers like sequencing.com
1 reply →
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.
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.
Not really, no.
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.
2 replies →
Alpha Fold has continued to impact the field of protein networks, but I do hear that not every one of the deep learning biology models from Google/Deep Mind and others have made equivalent impact or had as lasting relevance in their respective domains..some have performed more poorly than other available models. I'd love to learn more about this, but this has mostly come from little snips of conversations here and there, in person and online, but I haven't seen anything comprehensive in terms of evaluating their impacts overall
This may not be anything new but it makes using several Google/DeepMind resources a lot less painful.
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.
Is this just Google precomputing Alpha genome values - which were already accessible via API and making them available as another API (presumably more broadly)? Or is there actually new information?
That’s my reading. (That this is a cached database of Alpha values)
People are upvoting this because it has the “Alpha______” prefix. Meanwhile, everyone in the field of genomics knows that AlphaGenome provides essentially zero improvements over the previous SOTA, Borzoi…
> a database that predicts the effects of every possible single nucleotide variant in the human genome. We used the AlphaGenome AI model to pre-calculate the regulatory impact of all 9 billion single-letter genetic changes, resulting in a massive, 1-petabyte dataset.
This is for a database, no? While Borzoi is a model?
> Here, we introduce Borzoi, a model that learns to predict cell-type-specific and tissue-specific RNA-seq coverage from DNA sequence.
https://www.nature.com/articles/s41588-024-02053-6
Any model can be expressed as a database.
1 reply →
> Meanwhile, everyone in the field of genomics knows that AlphaGenome provides essentially zero improvements over the previous SOTA, Borzoi…
Can you elaborate on this? I'm confused why Google would build something that provides zero improvements over SOTA, Borzoi... as you mention. I'm not familiar with this field, just curious.
Speaking as somebody who has worked within Google Research before: the researchers are under tremendous pressure to publish SOTA and sometimes they juice their results a bit to look competitive when they can't match. This is not uncommon in the field- it's remarkably easy to edit a paper to make yourself look good by omitting information.
1 reply →
The use of an exact quote in an ungrammatical fashion is a bit of a language model smell. I can’t help but be reminded of the purely nonsensical AI interview answers. “It’s a pleasure to meet you, Chick Bongo”
1 reply →
They did benchmark the model and beat the SOTA on every metric though
Is it so bad to have another entrant, especially with the resources Google could bring to bear?
Imagine if the Apple EV had actually happened, you think the EV enthusiasts would roll their eyes like you are?
Are you really taking a holier than thou approach on a google article?
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.
Interestingly the labs led by marketing CEOs have much better models (astra is A LOT better than gemini).
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?”
Very excited to see that happen here.
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.
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)
Why are you excited about this?
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.
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.
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.
I really love Deep Mind, its genuinely focused on using AI to make the world a better place.
Be careful when loving shareholder-driven endeavors; they're a single executive decision away from breaking your heart.
It's "shareholder-linked" (probably not "driven"), but yes, upvoted. Nonetheless, even when with grave faults in the path, Google has managed to give us marvellous, paramount free resources (Google Street, Google Museum...), and would have given more (Google Books - if the negotiation had succeeded). I am grateful.
9 replies →
Be careful when loving government/philanthropy/ngo-driven endeavors; they're a single executive decision away from breaking your heart.
3 replies →
No AI company is doing that. There is some benefits but that’s not what any AI for profit will ever focus upon
DeepMind objectively is doing that. The world is not so black and white as "evil profit-hungry corporations will never do anything good".
19 replies →
I am freaking out. This is a huge moment.
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.
This post is not about Navier-Stokes. That is in a different thread. Glad you're cry-happy though.
Wait, is NS solved? I thought I was just a singularity thing?
There is an announcement by OpenAI https://news.ycombinator.com/item?id=49613262
So I will know which DNAs to change to become a wolverine! yay!
Anyone know if you can use an indel VCF file with this?
Cool. But is this actually advanced biology, or just making predictions about biology more accurate? Those aren't the same thing.
I'm convinced AI labs are absolutely hallucinating with their random names. It doesn't even mean anything anymore. Can we go back to numbers?
could this be used with a nebula genomic sequence to find pathogenic sequences?
This Google blog post is a distilled version of a Deep Mind blog post:
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
Appreciate the load-bearing info.
[dead]
[dead]
hmm gate keeping the database to elitist institutions and private businesses, I'm excited for the future!
All controlled by an adCompany.
That is outright scary. Science is being slurped up here.
unleash gemini 4 stop saving dario
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?
Just wait a few months, if it is a big deal that will likely be obvious by then
i like this way of thinking ;)
1 reply →
What are you even talking about? What do you mean catch a break? They’ve made massive contributions over the last decade.
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
3 replies →
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?
Individual SNP predictions are not useful for drug discovery. Pharma might license this, but mainly out of fear of missing out.
Probably. I know Google has partnerships with Pharma: https://www.merck.com/news/merck-and-google-cloud-partner-to...
This is just typical enterprise "use AI" infrastructure, not hyper specialized drug research.
My understanding is that they already are, via Isomorphic Labs
It's a bigger deal for diagnostics than for drug development.
Hopefully. Tired, as a shareholder, to see the giveaways that deepmind is doing.
Yeah, screw enabling life changing medical research, won't someone please think of the investors' financial returns??
1 reply →
Astroturfing, eh? This is verbatim the same comment as another user in this thread. Strange...
5 replies →
Google doing work to uncover the pandoras box of genetics? How long before they shove this under the rug...
An atlas of the human genome from the company whose other atlas still routes me into a lake.