Google DeepMind has released AlphaGenome 2, a new version of its AI system for reading DNA, built specifically for cell-level research. The focus is a subtle but crucial one: a genetic variant does not do the same thing in every cell of the body, and predicting its effect means predicting it in the cell type where it actually acts.
Most disease-linked variants sit outside genes, in the regulatory DNA that controls when genes switch on and off — and that control logic differs from tissue to tissue. A model that reads those patterns cell by cell gives researchers a way to test, in a computer, which variants are worth years of laboratory work.
AlphaGenome 2 continues DeepMind's run of applying its strongest models to biology's hardest data — following the path that took its protein-folding work to a Nobel Prize.
The background
Most of the genome is not genes. The protein-coding regions that dominated early genetics are a small fraction of human DNA; the rest is dense with regulatory sequences — switches and dimmers that control when, where and how strongly genes turn on. When large studies link DNA variants to diseases, most of the implicated variants turn out to sit in those regulatory regions, not in genes themselves.
That is what makes prediction so hard. A regulatory switch may be active in liver cells and silent in brain cells, so the same variant can be consequential in one tissue and irrelevant in another. Reading a genome in the abstract is not enough; researchers need to know what a variant does in the specific cells where a disease actually develops.
Why it matters
The practical problem in genetics is triage. Studies produce far more candidate variants than any laboratory can test, and each experiment costs time and money measured in months and years. A model that can rank variants by their likely effect — cell type by cell type — tells researchers where to spend those years, and which leads are probably dead ends.
AlphaGenome 2 also continues a recognisable DeepMind pattern: take one of biology's hardest, data-rich problems and turn it into a prediction task its models can be trained and judged on. Protein folding was the famous example; the genome's regulatory logic is a larger and messier one, and progress there feeds directly into how diseases are understood and, eventually, treated.
What happens next
The measure of AlphaGenome 2 will be adoption and accuracy: whether genetics laboratories build it into their pipelines, and whether its cell-level predictions hold up when experiments test them. Models in this field earn trust slowly, one verified prediction at a time.
Common questions
What does AlphaGenome 2 do?
AlphaGenome 2 is Google DeepMind's AI system for reading DNA, built for cell-specific research. It predicts how genetic variants behave in particular cell types — the places where diseases actually act — rather than giving one average answer for the whole body.
Why does cell-specific prediction matter?
Most disease-linked variants sit in regulatory DNA that switches genes on and off, and that control logic differs between tissues. A variant can matter in one cell type and be irrelevant in another, so predicting its effect cell by cell tells researchers which variants are worth laboratory study.




