Google DeepMind releases AlphaGenome 2, which can predict which cell types DNA variants act on
Google DeepMind has released AlphaGenome 2, a new-generation DNA sequence prediction model that can predict which cell types a DNA mutation acts in and which genes it affects, for use in tracing disease roots and finding drug targets. The model's training data covers 68 tissues and more than 107 million cells; in the TenK10K test covering 28 types of immune cells, its recall for identifying key causal variants is 30%, compared with 16% for the previous generation after fine-tuning on single-cell data, and its ability to determine whether a variant enhances or weakens gene activity rose from 23% to 43%. The model is now open to non-commercial research and can be called via the AlphaGenome API; it will later also launch on Google Antigravity in the form of AlphaGenome 2 Skills.
Why read this
AI score 75
The original text provides a recall comparison between the new and old models in single-cell variant identification, as well as a case of reproducing a year-and-a-half experiment in half an hour, which can be used to assess the model's impact on the target discovery pipeline.
Topics
Body · AI translation
Which cells do DNA mutations actually cause disease in? Now, AI can help us investigate that too!
Google DeepMind has now released AlphaGenome 2, a next-generation DNA sequence prediction model.
It can predict which types of cells a DNA mutation will act in and which genes it will affect, and it can also help scientists trace the roots of disease and find more precise drug targets, offering new directions for targeted treatments for diseases such as cancer and diabetes.
And this time, its performance genuinely stunned scientists.
For example, the James Lee team spent a year and a half on experiments to find clues related to inflammatory bowel disease, while AlphaGenome 2 took only half an hour to reproduce the key findings in computation.
This prompted James Lee to exclaim in an interview with Nature:
They worked out the answer in 30 minutes, whereas our experiments took a full year and a half...
Even better, AlphaGenome 2 can even be used out of the box, with no retraining required.
On this point, Sarah Teichmann, co-founder of the Human Cell Atlas consortium and a professor of medicine at the University of Cambridge, could barely contain her excitement:
The most exciting thing about AlphaGenome 2 is that it can directly adapt to new single-cell data without retraining;
Even with only a small number of cell samples, it can predict how DNA variants will act in newly detected cell states.
Onlookers also exclaimed that it was awesome:
Some have even begun to imagine a future in which they live past 100, hoping AI can really help. (doge)
An AI trained on 100 million cells can finally trace DNA variants down to specific cells
By the way, why can AlphaGenome 2 help scientists trace the effects of DNA mutations?
When introducing this achievement, Google CEO Sundar Pichai raised a very interesting concept:
The cells in our bodies basically all carry the same DNA blueprint, but different types of cells, such as heart muscle cells and immune cells, interpret this blueprint differently.
In other words, for the same stretch of DNA, some cells will read certain genes from it, while others will shut them off and not use them.
Therefore, the same DNA variant may also have an effect only in a specific type of cell.
This gives scientists a headache.
Even when they already know that a certain stretch of DNA is related to a disease, it is hard to figure out exactly which cell type and which gene are the problem.
Now, AlphaGenome 2 can help.
To teach the AI to read DNA in different cells, DeepMind put it through a massive biology training program.
The training data covered 68 tissues and more than 107 million cells, including both human cells and mouse cells.
After taking in so much data, the new model's results were also quite impressive.
In the TenK10K test covering 28 types of immune cells, AlphaGenome 2 achieved a recall rate of 30% in identifying key causal variants, while the previous generation, after fine-tuning on single-cell data, scored only 16%, nearly double.
Its ability to determine whether these variants would enhance or weaken gene activity also soared from 23% to 43%.
So how exactly does it do this?
The key lies in the fact that AlphaGenome 2 changed the way it reads DNA.
(The previous generation, AlphaGenome, read in about 1 million DNA bases at a time and would simultaneously predict their molecular activity across a large number of tissues and cell lines; although this approach of handling multiple prediction tasks in one go was highly efficient, it was not very good at distinguishing subtle differences between similar cell types.)
Now, scientists can precisely specify the cell type or state they want to study and have the model specifically predict how this stretch of DNA will act under that condition.
(It is equivalent to previously grading a large stack of exam papers from different subjects all at once, whereas now you can specify the subject and examine them one by one in detail.)
△
Moreover, the new model has also learned to predict DNA methylation and can even infer molecular features that have not yet been measured, such as chromatin accessibility, based on already measured gene activity.
In this way, scientists can gradually narrow the scope of their research to further pin down exactly which stretch of DNA and which gene have quietly gone wrong in a specific cell.
OK, theory is over. Next, let's look at the most exciting part: real-world application!
From finding a needle in a haystack to precise targeting! AlphaGenome 2 reproduces a year-and-a-half experimental finding in half an hour
So how exactly does AlphaGenome 2 help scientists find disease clues?
Take the inflammatory bowel disease mentioned at the beginning.
The team of British scientist James Lee spent a full year and a half doing experiments before finally figuring out that a DNA variant hidden in a "gene desert" would cause the ETS2 gene in intestinal immune cells to become overactive, thereby triggering intestinal inflammation.
But with AlphaGenome 2, things became much simpler.
Google had it screen seven suspicious DNA variants in one go, and the model quickly locked onto the key one, accurately predicting which gene it would affect and in which type of immune cell it would act.
△
In addition, Google is also using AlphaGenome 2 in research on complex diseases such as type 2 diabetes, helping scientists screen out more suspicious clues from massive numbers of DNA variants, identify the relevant cells and genes, and then carry out targeted experiments and look for potential drug targets.
And patients who have long suffered from diseases may also have the chance to see more precise and targeted treatment options in the future.
One More Thing
The good news is that AlphaGenome 2 is already open to non-commercial research!
Everyone can call the new model through the AlphaGenome API and refer to the interactive Colab tutorial provided by Google to run experiments with their own single-cell experimental data.
Moreover, Google also plans to put it into the toolbox of Agents.
Next, AlphaGenome 2 Skills will also land on Google Antigravity, allowing everyone here to connect this DNA prediction model to their own Agent workflows.
Those interested should go try it out soon~
Reference links:
This article comes from the WeChat public account "Quantum Bit", author: Focus on Frontier Technology, authorized for publication by 36Kr.