Google DeepMind’s AlphaGenome model has been published in Nature, taking in a million letters of DNA at a time and predicting how small changes in that sequence affect the way genes are switched on and off. The paper reports that the model matched or beat the strongest external tools on 25 of 26 variant effect prediction tasks.
What happened
On 28 January, Nature published the peer reviewed paper “Advancing regulatory variant effect prediction with AlphaGenome”. DeepMind first shared the model as a preprint in June 2025, so the news this week is not a surprise launch. It is the moment independent reviewers signed off on the work and the full method became part of the scientific record.
According to the paper, the model processes “1 Mb of DNA sequence and predicts thousands of functional genomic tracks up to single-base-pair resolution.” In plain terms: one megabase, or roughly a million DNA letters, goes in, and thousands of predicted biological signals come out, each one detailed down to a single letter.
The authors compared AlphaGenome with competing models on two families of tests. It came out ahead on 22 of 24 genome track evaluations, and it matched or exceeded the best external approaches on 25 of 26 tasks that ask what a specific mutation does.
To show the model on a real problem, the team looked at cancer-linked variants near a gene called TAL1 in T-cell leukaemia.
How it works
Genes, the stretches of DNA that code for proteins, are only part of the story. Much of the rest of the genome, sometimes called its dark matter, contains the switches and dimmers that decide when, where and how strongly each gene is used. A single changed letter in one of those switches can matter as much as a change inside a gene, yet it is much harder to interpret.
A helpful way to picture it: think of the genome as a very long musical score. The genes are the notes, and the regulatory regions are the tempo marks and volume directions written around them. A typo in a volume direction does not change any note, but it can change how the whole piece sounds. AlphaGenome is trained to read the score and predict how the performance changes when one of those directions is altered.
The “tracks” in the paper are the model’s predictions of many measurable signals along the DNA, such as how active a region is. Reading a full megabase at once matters because a switch can sit far away from the gene it controls. A model with a short window can miss that link entirely.
For a variant, the model effectively compares two versions of the sequence, the normal one and the mutated one, and reports how its predicted tracks differ. That difference is the predicted effect of the mutation.
By the numbers
| Item | Figure | Source |
|---|---|---|
| DNA read in a single pass | 1 Mb, about a million letters | Nature |
| Finest prediction resolution | Single base pair | Nature |
| Genome track evaluations where it led | 22 of 24 | Nature |
| Variant effect tasks matched or exceeded | 25 of 26 | Nature |
| Example disease studied | T-cell leukaemia, variants near TAL1 | Nature |
| Nature publication date | 28 January 2026 | Nature |
Why it matters
Most disease-linked genetic variants lie outside genes, in exactly the regions that are hardest to read. When a patient’s genome is sequenced, doctors and researchers often find changes in these areas and simply cannot say whether they are harmless or important. A tool that predicts their effect gives scientists a way to rank which variants deserve a closer look in the lab.
What caught my attention is the breadth of the comparison. Beating rivals on one benchmark is common in AI papers. Leading on 22 of 24 track evaluations and staying at or near the top on 25 of 26 variant tasks suggests a single general model rather than a stack of narrow ones.
It is important to be precise about what this is. AlphaGenome is a research model that makes predictions. It does not diagnose patients, and the paper does not present it as a clinical tool. Its output is a hypothesis that still needs checking with experiments.
In my view, the TAL1 example is the right kind of demonstration: a known cancer-related region where the model’s reading can be compared with what biologists already understand.
What comes next
With the paper now peer reviewed, the method is open to scrutiny from the wider genomics community, and the natural next step is for other groups to test the predictions against their own laboratory data. Areas the research highlights as likely beneficiaries are rare disease diagnosis and cancer research, where interpreting variants outside genes is a daily bottleneck. How quickly that happens will depend on how well the predictions hold up in those independent tests.
Sources
- Nature, “Advancing regulatory variant effect prediction with AlphaGenome”, 28 January 2026, https://www.nature.com/articles/s41586-025-10014-0
- Google DeepMind, blog post introducing AlphaGenome, June 2025, https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/
- Nature Asia, press release on the AlphaGenome paper, January 2026, https://www.natureasia.com/en/info/press-releases/detail/9222