Nobel laureate Hassabis has delivered another phenomenal masterpiece, and AlphaGenome 2 deciphers 98% of the human genetic "dark matter".

DeepMind drops another bombshell late at night!

Just now, Nobel laureate Demis Hassabis personally leads the team, leading Google DeepMind to officially release the new-generation DNA foundational model AlphaGenome 2.

It is worth noting that AlphaFold, which made him a legend in the field back then, only decoded the 2% protein-coding regions in human DNA.

The remaining full 98% non-coding regions have long been known as the "dark matter" in the genome. These regions not only control the "regulatory switches" for gene activation and deactivation, but also hide over 90% of human disease-causing mutations.

This time, DeepMind released a 78-page heavyweight technical report in one go, launching a full-scale offensive against this "dark matter" with AlphaGenome 2, bringing four core breakthroughs.

• From "observing the entire organ" to precise "observation at the single-cell level"

The model has mastered data of 68 human tissues and over 107 million single cells, which can clearly show the respective responses of the same DNA mutation in different cells, and the success rate of identifying disease-causing mutations is directly doubled compared with the previous generation.

• No biological experiments required, pure computation reproduces years of research achievements from *Nature*

Without touching a single reagent, the model accurately identified the hidden mutation that causes severe enteropathy, and even clearly calculated in which type of immune cell the mutation takes effect.

• For unseen new lesions, the model can understand them with only dozens of cells as input

Even for lesion samples that the model has never seen before, researchers do not need to retrain the model. They only need to upload sequencing data of dozens of cells to quickly predict the consequences caused by mutations.

• Free million-level variant library is open to public, and will soon be integrated into Gemini 4

The official has 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 to single cell, disease-causing mutation recognition rate doubles directly

To decipher this genomic "dark matter", DeepMind previously launched the first-generation AlphaGenome, but it had an obvious shortcoming: it could only make general predictions based on "the entire organ tissue".

However, a piece of human tissue is usually mixed with dozens of different types of cells, and mixing them together can only calculate the "average signal", which leads to the characteristics of a small number of diseased cells being easily covered up.

This also explains a confusing phenomenon that has long existed in the genetics community —

It is clearly the same immune-related DNA mutation, which increases the expression of the IL20RA gene in lung tissue, but reduces it in esophageal tissue. Sequencing based on the entire organ has never found the reason behind this.

It was not until AlphaGenome 2 zoomed the perspective to a single cell that the underlying reason was completely clarified.

It turns out that in epithelial cells, this mutation destroys the regulatory site responsible for activating genes, resulting in significantly decreased gene expression.

On the contrary, in vascular endothelial cells, it destroys the site responsible for inhibiting genes. Without restriction, gene expression rises sharply instead.

Since epithelial cells account for the majority in the esophagus and vascular endothelial cells are more abundant in the lungs, when the average value is calculated by mixing the entire organs, the two opposite effects cancel each other out.

It is not difficult to see that when the prediction accuracy is refined to the single-cell level, the recognition accuracy of the model has achieved a substantial leap.

In the TenK10K benchmark test covering 28 types of human immune cells, even after special single-cell fine-tuning on the first-generation model, it could only find 16% of disease-causing mutations.

In contrast, AlphaGenome 2 not only increases this proportion to 30%, but also raises the accuracy rate of judging "whether a mutation enhances or weakens gene activity" from 23% to 43%, the two core indicators almost double at the same time!

Even for the dynamic process of human hematopoietic stem cells differentiating into red blood cells and B cells step by step, the gene change heat map it draws is exactly the same as the real measured results in the laboratory.

Pure AI simulation calculation, zero experiments reproduce years of *Nature* achievements

If doubling the benchmark test scores is only the basic skill, its performance on real intractable diseases is what really shocks biologists.

In 2024, *Nature* once published a sensational medical research on inflammatory bowel disease.

To find out a disease-causing clue hidden in the non-coding region, the scientific research team worked hard in the laboratory for several years, repeatedly carrying out tedious experiments such as CRISPR gene knockout.

Finally, they barely confirmed that a tiny gene mutation crossed the ultra-long distance of 290,000 base pairs, over-activating the ETS2 gene, thus triggering severe intestinal inflammation.

This time, the DeepMind team input the same segment of DNA into AlphaGenome 2. Then, the model did not conduct a single biological experiment throughout the process, and reached exactly the same conclusion only through calculation!

In this process, the model not only immediately spotted the disease-causing mutation from 7 highly similar candidate sites, but also skipped the adjacent genes with closer distance, accurately locking the distant ETS2.

Moreover, it further calculated the key detail that "this mutation only works in intestinal macrophages and monocytes, and other immune cells are almost not affected".

In addition to finding the cause of disease, its performance on the approved marketed CRISPR gene therapy targets for sickle cell disease is also very outstanding.

The model not only accurately found the regulatory site 58,000 base pairs apart in key cells in the early stage of red blood cell development, but also recognized the core GATAA sequence targeted by gene editing word for word.

No need to retrain the model, only dozens of cells are fed to understand unknown lesions

In the front line of clinical practice, researchers often encounter a practical problem.

For many rare diseases or new infections, only dozens of diseased cells can be collected from patients, and this amount of data is far from enough to retrain a large model.

To solve this pain point, AlphaGenome 2 specially added a cell encoder outside the backbone network.

With this design, no matter what new unseen lesions are encountered, researchers only need to upload sequencing data of dozens of new cells, and the cell encoder can extract the corresponding cell state features on the spot.

Subsequently, these features will be directly transmitted to the backbone network. In this way, the entire large model does not need to be retrained, and can directly predict the impact of gene mutations in these new cells.

To verify this capability, the team selected a batch of brain tissue samples from 190 Parkinson's disease patients that the model had never seen before for testing.

Without any targeted training, AlphaGenome 2 successfully identified 41% of disease-causing mutations, far exceeding the similar model Scooby (only 17%) specially trained with this batch of patient data.

In another batch of samples from COVID-19 mRNA vaccine recipients, it not only keenly identified the rare T cells that were awakened only after the booster shot, but also followed the clues to find a very hidden severe COVID-19 mutation.

The data shows that this mutation has almost no effect in unactivated T cells. But once the immune system is activated, it will significantly inhibit the expression of antiviral genes, thus increasing the risk of infection.

Moreover, this architecture can also help laboratories save a large amount of testing funds.

Researchers only need to measure the most conventional single-cell gene expression data, and the model can automatically infer a number of high-cost indicators such as single-base DNA methylation, which is equivalent to completing multiple data items by testing only one item.

7 million variant library is open for free, and will be integrated into Gemini 4 as Skills soon

In addition to releasing the model and paper, DeepMind this time also pre-calculated two large-scale databases for global researchers and made all of them publicly available for free.

It not only includes the gene regulation map of 640 types of real human cells, but also pre-calculates the specific impact scores of nearly 7 million gene variants under 510 cell states.

At present, relevant APIs and interactive Colab tutorials are open for free for non-commercial scientific research.

At the end of the official blog, DeepMind also released a heavy preview.

The team will soon package AlphaGenome 2 into Skills, integrate it into its own Agent development platform Google Antigravity, and let Gemini 4 call it directly!