Google DeepMind has released AlphaGenome 2, a genomics model that predicts how a single change in your DNA affects specific cell types. The first AlphaGenome worked at the level of broad tissues. This version works at the level of individual cell types and the states those cells are in. According to Google DeepMind, that finer view nearly doubles how many disease-linked genetic variants the model can identify.
The problem it’s trying to solve
Your heart, brain and immune cells all carry the same DNA. Yet a one-letter change in that DNA can cause disease in one type of cell and do nothing in the others. That happens because each cell type “reads” the genome differently. It switches different genes on and off depending on what it is and what it’s doing at the moment.
Most disease-linked variants don’t sit inside genes. They sit in the regulatory regions that act like dimmer switches for genes. When a variant changes one of those switches, the effect may show up only in one type of cell. Earlier models averaged across whole tissues, so they often missed those effects.
What DeepMind built
Google DeepMind says AlphaGenome 2 zooms in on one cell type at a time. It predicts how a variant changes gene regulation in that specific cellular context. Here’s what DeepMind has shared about how it works:
- Training data: more than 100 million profiles across 68 tissues, according to the company’s announcement.
- Cell-state awareness: it models cells in specific states, not just broad tissue averages.
- Generalization: DeepMind reports that it can handle entirely new single-cell experiments without retraining.
That last point matters more than it sounds. Single-cell datasets are multiplying fast, and labs don’t want to retrain a large model every time they run a new experiment.
The benchmark numbers
The headline result comes from TenK10K, a large immune-cell dataset. The task is to find variants that likely cause changes in gene expression, while keeping precision high so researchers don’t chase false leads.
- AlphaGenome 2: recovered 30% of putatively causal expression-linked variants at 90% precision
- AlphaGenome 1 (fine-tuned on single-cell data): recovered 16% at the same precision
That’s almost twice the yield at the same error rate. The comparison with a fine-tuned AlphaGenome 1 is a fair one. DeepMind didn’t just test against the old model straight out of the box. It gave the old model the same single-cell data and still beat it by a wide margin.
Why this matters for researchers
Genome-wide studies routinely turn up hundreds of variants linked to a disease. Very few of them get confirmed in the lab, because every experiment costs time and money. A model that ranks candidates more accurately, and also tells you which cell type to test in, can cut a lot of that wasted work.
Here’s what research teams can realistically do with it:
- Prioritize variants from association studies before committing to wet-lab work.
- Pick the right cell type to test, so you don’t run assays in tissue where the effect never shows up.
- Analyze new single-cell data with a model that doesn’t need retraining for it.
Access and limitations
AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API. Commercial research access is in private preview. DeepMind has also announced AlphaGenome 2 Skills for Google Antigravity, which isn’t out yet.
There are some important caveats:
- Research only. It hasn’t been validated or approved for clinical or diagnostic use. Don’t use it to make decisions about patients.
- Self-reported results. The benchmark numbers come from DeepMind’s own evaluation. Independent replication on other datasets will show how well it holds up.
- Recall is still limited. Recovering 30% of causal variants is a big step up from 16%, but it still means most of them go undetected.
Where this is heading
What stands out to me is the direction. Genomics AI is moving from asking “what does this variant do?” to asking “what does it do in this cell, in this state?” Disease biology works at that level, so it’s the right question to ask. If your lab works on variant interpretation, I’d test AlphaGenome 2 against your existing pipeline on data you already understand well before trusting it with new targets. The full announcement on the Google DeepMind blog has more detail.