DeepMind’s AlphaGenome 2 Reads DNA Cell by Cell to Hunt Disease-Causing Variants

Google DeepMind has unveiled AlphaGenome 2, a new artificial intelligence model that predicts how changes in DNA alter gene activity inside individual human cell types — a shift the company says could help scientists pinpoint the cellular roots of disease.

DeepMind announced AlphaGenome 2 on Wednesday 11 November, a little over a year after the first AlphaGenome arrived in June 2025. Where the original model predicted DNA activity across broad tissues, the new version zooms in on specific cells and cell states, the level at which many genetic diseases actually take hold.

From tissue averages to single cells

The problem AlphaGenome 2 tackles is one of resolution. A genetic variant can matter enormously in one specialised cell type — a gut macrophage, say, or an activated T cell — while its signal is diluted to nothing when averaged across a whole tissue. Tissue-wide models can therefore miss the very effects that drive disease.

AlphaGenome 2 is built differently. It is a conditional model: it takes a DNA sequence of up to a million letters together with a chosen condition — a tissue, a cell type or a cell state — and predicts the molecular readouts for that condition alone. Those readouts include gene expression, RNA splicing, chromatin accessibility and histone marks, with DNA methylation added in this version.

The model was trained on more than 107 million individual cell profiles drawn from human and mouse samples across 68 tissues. A built-in cell encoder also lets researchers feed in data from their own single-cell experiments, turning a small sample into a new condition without retraining the model.

Nearly twice the variant recall

On DeepMind’s own benchmarks, the gains are substantial. Tested on TenK10K, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 recovered 30 per cent of putatively causal gene-expression variants at 90 per cent precision, against 16 per cent for the first AlphaGenome fine-tuned on single-cell data. It called the direction of gene effects correctly 43 per cent of the time, against 23 per cent before.

“AlphaGenome 2 represents a step-change improvement in variant predictive performance,” DeepMind said in its announcement, adding that the model identifies nearly twice as many variants that change gene expression in cells as its predecessor.

The company also gave a disease example. In an analysis of inflammatory bowel disease, the model flagged a variant known as rs2836882 from a seven-variant credible set, linked it to a gene called ETS2 roughly 290 kilobases away, and localised the predicted effect to gut monocytes and macrophages — the immune cells where the disease mechanism is thought to operate.

What it can — and cannot — do yet

The results come with clear limits, and DeepMind states them plainly. Even at 90 per cent precision, seven in ten causal variants in the TenK10K test still went unnoticed. All of the figures are the company’s own; there has been no independent replication yet. The model can also underestimate effects that sit outside accessible chromatin, act through very distant regulatory elements, or work after transcription. It is not yet clear whether it can predict responses to gene knockouts, overexpression or dynamic changes in cell state.

Crucially, both AlphaGenome models are for research use only. DeepMind states they “have not been validated or approved for clinical or diagnostic purposes” — so no doctor or patient should treat its predictions as medical advice. The new model is also more computationally expensive than the first, which is why DeepMind says a genome-wide atlas like the one built for the original model, covering all nine billion possible single-letter changes in the human genome, is impractical for version 2.

That caution is part of a wider pattern in science AI, where powerful new instruments arrive alongside careful warnings about their limits — a balance also visible in recent work such as Harvard’s concussion-detecting headset, studies of amino acids found on the asteroid Bennu, and the FDA’s fast-tracked sleep drug for children.

Available now for research

AlphaGenome 2 is available from today for non-commercial research through the AlphaGenome API, with interactive Colab tutorials for expression conditioning. Commercial research access is in private preview. DeepMind has also published precomputed scores for trait-associated variants across hundreds of human cell types, for researchers who would rather use the results than run the model themselves.

If the benchmarks hold up under independent scrutiny, the model’s real value may be less in any single prediction than in the questions it lets scientists ask: not just which DNA letter changed, but in which cell, in which state, and to what effect. For a field that has long read the genome in tissue-wide averages, that is a finer lens — and, potentially, a faster route from a genetic clue to a disease mechanism.