DeepMind Releases DNA Foundation Model AlphaGenome 2
Google DeepMind has released a new-generation DNA foundation model, AlphaGenome 2, and published a 78-page technical report, with the focus shifting to the non-coding regions that make up 98% of the human genome. The model uses single-cell data from 68 tissues and over 107 million cells; on the TenK10K benchmark, its pathogenic mutation identification rate rose from 16% to 30%, and its accuracy in determining whether gene activity is enhanced or weakened rose from 23% to 43%. The official release freely provides impact scores for nearly 7 million genetic variants across 510 cell states, and it plans to package the model as Skills integrated into Google Antigravity for use by Gemini 4.
Why read this
SelectedAI score 80
The original text presents three capability changes: single-cell precision, zero-experiment reproduction, and training-free adaptation, which can be used to assess progress in DNA foundation models.
Topics
Body · AI translation
DeepMind just dropped another bombshell late at night!
Just moments ago, Nobel laureate Hassabis personally took the lead, guiding Google DeepMind to officially release the next-generation DNA foundation model, AlphaGenome 2.
You should know that AlphaFold, the model that made him a legend in one stroke, actually only cracked the protein-coding regions that make up 2% of human DNA.
The remaining full 98%—the non-coding regions—has long been called the "dark matter" of the genome. Not only does it house the "regulatory switches" that turn genes on and off, it also harbors more than 90% of human pathogenic mutations.
This time, DeepMind released a hefty 78-page technical report in one go, leading AlphaGenome 2 in a full-scale assault on this "dark matter," bringing four core breakthroughs.
From "looking at whole organs" to precisely "looking at single cells"
The model has thoroughly absorbed data from 68 human tissue types and over 107 million single cells, and can see how the same DNA mutation responds differently in different cells, doubling the success rate of pinpointing pathogenic mutations compared with the previous generation.
No biological experiments needed—purely computational reproduction of years of Nature findings
Without touching a single drop of reagent, the model precisely identified the mutation behind severe intestinal disease, and even worked out exactly which type of immune cell the mutation acts in.
When encountering unseen new lesions, just a few dozen cells are enough to understand them
Even for lesion samples the model has never seen, researchers don't need to retrain the model—they only need to upload sequencing data from a few dozen cells to rapidly predict the consequences of the mutation.
Free open access to a million-scale variant library, soon to be integrated into Gemini 4
Officials have publicly released an impact score library covering millions of gene variants, and will soon package the model as Skills to integrate into Google Antigravity, for Gemini 4 to call at any time.
Precise down to single cells, pathogenic mutation identification rate directly doubles
To decipher this genomic "dark matter," DeepMind had previously launched the first-generation AlphaGenome, but it had an obvious shortcoming: it could only make broad predictions at the level of "whole organ tissue."
However, a piece of human tissue is often a mix of dozens of different cell types, and mixing them together only yields an "average signal," which easily masks the characteristics of a minority of pathogenic cells.
This also explains why the genetics community has long had a puzzling phenomenon—
Clearly the same immune-related DNA mutation raises IL20RA gene expression in lung tissue, yet in esophageal tissue it does the opposite and lowers it; sequencing whole organs never found the reason.
It wasn't until AlphaGenome 2 zoomed in to the single-cell level that the underlying reason was fully uncovered.
It turns out that in epithelial cells, this mutation destroys the regulatory site responsible for activating the gene, causing gene expression to drop significantly.
Conversely, in vascular endothelial cells, what it destroys is the site responsible for suppressing the gene; once the restriction is lost, gene expression instead rises sharply.
Since epithelial cells predominate in the esophagus while vascular endothelial cells are more abundant in the lung, when whole organs are mixed together and averaged, the two opposing effects cancel each other out.
It's easy to see that once prediction precision is refined to the single-cell level, the model's identification accuracy achieves a major leap.
In the TenK10K benchmark covering 28 types of human immune cells, even when the first-generation model was specifically fine-tuned for single cells, it could only identify 16% of pathogenic mutations.
By contrast, AlphaGenome 2 not only raised this proportion to 30%, but also increased accuracy from 23% to 43% when judging "whether a mutation strengthens or weakens gene activity"—both core metrics nearly doubled!
Even for the dynamic process of human hematopoietic stem cells progressively differentiating into red blood cells and B cells, the gene-change heatmaps it draws match the real laboratory measurements exactly.
Purely AI simulation and computation, zero experiments to reproduce years of Nature findings
If doubling the benchmark scores was just basic skill, then its performance on real complex diseases is what truly stunned biologists.
In 2024, Nature published an inflammatory bowel disease study that caused a sensation in the medical world.
To pin down a pathogenic clue hidden in a non-coding region, the research team toiled in the lab for years, repeatedly performing cumbersome experiments such as CRISPR gene knockout.
In the end, they barely managed to confirm that a tiny gene mutation, across an ultra-long distance of 290,000 bases, overactivated the ETS2 gene, thereby triggering severe intestinal inflammation.
This time, the DeepMind team fed the same DNA segment into AlphaGenome 2. Then, without performing a single biological experiment throughout, the model reached a completely consistent conclusion purely through computation!
In the process, the model not only spotted the pathogenic mutation at a glance among 7 highly similar candidate sites, but also skipped the nearer genes beside it and precisely locked onto the distant ETS2.
Moreover, it further worked out the key detail that "the mutation only acts in intestinal macrophages and monocytes, with other immune cells almost unaffected."
Beyond finding the cause of disease, its performance on the target of the already-approved CRISPR gene therapy for sickle cell disease is equally impressive.
The model not only precisely located, in key cells from the early stages of red blood cell development, a regulatory site 58,000 bases away, but also identified the GATAA core sequence targeted by gene editing, character for character.
No need to retrain the model—just feed it a few dozen cells to understand unknown pathologies
When they reach the clinical front lines, researchers often encounter a real-world challenge.
For many rare diseases or emerging infections, only a few dozen diseased cells can often be collected from a patient—far too little data to retrain a large model.
To address this pain point, AlphaGenome 2 adds a dedicated cell encoder outside the backbone network.
With this design, no matter what unseen new pathology is encountered, researchers only need to upload sequencing data from a few dozen new cells, and the cell encoder can extract the corresponding cell-state features on the spot.
These features are then passed directly to the backbone network. In this way, the entire large model can directly predict the effects of genetic mutations in these new cells without retraining.
To validate this capability, the team selected brain tissue samples from 190 Parkinson's disease patients that the model had never seen for testing.
Without any targeted training, AlphaGenome 2 successfully identified 41% of the pathogenic mutations, far exceeding the comparable model Scooby (only 17%), which was specifically trained on this patient data.
In another batch of samples from COVID-19 mRNA vaccine recipients, it not only keenly identified rare T cells that were awakened only after the booster shot, but also followed the trail to find a very hidden severe COVID-19 mutation.
The data show that this mutation has almost no effect in unactivated T cells. But once the immune system is activated, it significantly suppresses the expression of antiviral genes, thereby increasing the risk of severe infection.
Moreover, this architecture can also save laboratories a substantial amount of testing costs.
Researchers only need to measure the most routine single-cell gene expression data, and the model can automatically infer multiple high-cost metrics such as single-base DNA methylation—equivalent to measuring one item and completing multiple datasets.
Free release of a library of 7 million variants, soon to be packaged as Skills and integrated into Gemini 4
In addition to releasing the model and paper, DeepMind this time also precomputed two large-scale databases for researchers worldwide and made them fully public for free.
These include not only gene regulatory maps of 640 real human cell types, but also precomputed specific impact scores for nearly 7 million genetic variants across 510 cell states.
Currently, the relevant API and interactive Colab tutorials are freely available for non-commercial research.
At the end of the official blog post, DeepMind also dropped a major teaser.
The team will soon package AlphaGenome 2 as Skills, integrate it into its own agent development platform Google Antigravity, and have Gemini 4 call it directly!
Imagine: on one side, the Gemini 4 brain, skilled in literature study and scientific reasoning; on the other, AlphaGenome 2, which sees through hundreds of millions of single cells.
When the two are combined, the trial-and-error and screening that originally required years of hard work in the lab can be run through in advance on a computer via a few rounds of agent calls.
In this way, the cycle for humanity to figure out complex diseases and develop new drugs will also be greatly shortened!
References:
This article is from the WeChat public account "Xinzhiyuan", author: ASI Revelation, published by 36Kr with authorization.
AI Trends