DeepMind 发布 AlphaGenome Atlas，预测人类基因组约 90 亿种单碱基突变的影响

Google Deepmind has predicted what each of the roughly nine billion possible single-letter changes in the human genome would likely do inside the body. The AlphaGenome Atlas aims to help researchers spot the few relevant variants among a flood of genetic ones.

The human genome runs to about three billion DNA letters. Every person carries millions of tiny deviations from the reference sequence, usually a single swapped letter. Most of these variants are harmless. A few cause disease. Which ones can't be read directly from the DNA, and testing every one in the lab is basically impossible when there are roughly nine billion possible swaps.

The newly released AlphaGenome Atlas tries to fill that gap with predictions. For each of those nine billion changes, it offers an estimate of how the change would likely affect molecular processes across hundreds of cell types and tissues. The dataset spans one petabyte, making it more than 30 times the size of the AlphaFold database for protein structures.

It builds on the AI model AlphaGenome, introduced in 2025. The model reads DNA stretches one million letters long and predicts how strongly a gene gets read, whether regulatory proteins can bind to the DNA, and how a gene's transcript gets spliced. Until now, the model had to be queried for each variant one at a time. Now the answers are precomputed. According to the paper, each variant comes with about 27,000 individual prediction values on average.

That matters most for the roughly 98 percent of the genome that holds no blueprints for proteins. These noncoding regions act like switches and dials that decide when and in which tissue a gene is active. That's where most disease-linked variants sit, and it's also where their effects have been hardest to read.

One number for every mutation

Thousands of prediction values per variant are too much for everyday use, the team says. So Deepmind built the AlphaGenome Variant Impact Score (AVI), which boils it all down to a single number. A small neural network combines the AlphaGenome predictions with the protein model AlphaMissense and two measures of how unchanged a DNA site has stayed across millions of years of evolution. AVI works with 18 input features. The established benchmark tool CADD uses more than 150.

For almost no variant is it known for sure whether it causes harm. The team worked around this. Variants that are very rare in the population are treated as likely harmful, common ones as likely harmless, because harmful mutations spread less often across generations. Despite this indirect training, AVI beat existing tools in tests on variants that had already been clinically classified, especially in noncoding regions, according to the paper. On some tasks, the competition edged ahead. The atlas also breaks down for each variant which process drives its score, such as whether the splicing of a gene's transcript or a switch is affected.

An epilepsy case shows the payoff

A case from the GREGoR consortium, which studies unsolved rare diseases, shows how this helps in practice. A child with severe epilepsy had gone without a diagnosis despite genome sequencing. AVI pushed a variant in the gene DNM1, previously classed as unclear, to the top of the candidate list.

The AlphaGenome predictions also supplied the mechanism. The variant creates a wrong splice site during the processing of the gene's transcript, which lengthens the protein by 13 building blocks. But this happens only in a gene version that is read exclusively in the brain. That's why earlier work on blood samples had found nothing.

A lab experiment confirmed the prediction, and the researchers recommend classifying it as likely disease-causing. Looking back at cases the consortium had already solved, AVI ranked the causal variant among the top 50 candidates in 29.5 percent of cases, compared with 12.5 percent for CADD.

More signal in the noise

The atlas is also meant to push population studies forward. To find out whether rare variants in a genome region affect something like a blood value, you have to analyze many of them together, because each one alone is too rare for statistics. If harmless and effective variants get mixed together, the signal disappears in the noise.

Gareth Hawkes of the University of Exeter used the atlas to group only those variants predicted to act the same way, drawing on genome data from more than 54,000 UK Biobank participants. That turned up 22 percent more links between noncoding variants and protein levels in the blood than conventional filters did.

From the predictions, the team also derived 2,601 recurring short DNA patterns, essentially the "words" of the genome where regulatory proteins latch on.

A research tool, not a diagnosis

AlphaGenome has limits too. It doesn't know every cell type, and it misses effects that work through the amount of other regulatory proteins. The atlas and AVI are research tools, Deepmind says, and can only be one link in the chain of evidence behind a diagnosis.

The atlas is available for noncommercial use through a web portal, an API, and as a skill in Google Antigravity. A commercial version is set to follow through Google Cloud.