Skin, Decoded
← Contents

Lesson 03 · The model

What AlphaGenome 2 actually does

6 minute read

AlphaGenome 2 was announced on 11 November 2026. It predicts how a DNA sequence — and changes to it — affect molecular activity in a specific human cell type.

You give it a stretch of DNA (up to about a million letters) and one condition: a tissue, a cell type, or a cell state. It returns predictions of the measurements a lab would otherwise make — gene expression, chromatin accessibility, histone modifications, splicing, and DNA methylation.

SEQUENCE ACGTTACGGCAT PREDICTED · ONE CELL TYPE EXPRESSIONCHROMATINSPLICINGMETHYLATION
One sequence, four predicted readoutsExpression · chromatin · splicing · methylation

What changed from the first version

AlphaGenome 1 predicted across thousands of experiments at once, which spread its capacity thin. Version 2 is conditional: it predicts for one cell type at a time and receives that condition as input at every stage of the network. In practice that means it can distinguish closely related cells — say, different immune cell types in the same tissue — which the first model tended to average together.

107 million+single-cell profiles used in training
68tissues represented
1 Mbmaximum input sequence (1,048,576 letters)
~2×more causal variants found than the previous model on one immune dataset

On the TenK10K dataset — a single-cell map of 28 primary immune cell types — the model identified 30% of causal variants, against 16% for the previous state of the art, and 43% of directional gene effects against 23%.

The case that makes it concrete

A 2024 Nature study found that a variant in a “gene desert” on chromosome 21 drives inflammatory bowel disease by over-activating a gene called ETS2 — but only in macrophages in the gut. The variant is far from the gene, and its effect is specific to one cell type. AlphaGenome 2 reproduced that finding computationally: it singled out the causal variant and showed the ETS2 increase in gut macrophages, with other cell types and genes largely unaffected.

What it is not

It is not a diagnostic. DeepMind states plainly that AlphaGenome and AlphaGenome 2 have not been validated or approved for clinical or diagnostic use. It does not tell anyone what to do about their skin. It produces a prediction about molecular activity, for a researcher to interpret.

The limits worth knowing

None of that makes it unimpressive. It makes it a scientific instrument with a stated operating range — which is exactly what you want a scientific instrument to have.

Check yourself

AlphaGenome 2 predicts for:

The ETS2 finding matters because it shows:

Can AlphaGenome 2 be used to diagnose a patient?