AlphaGenome 2 Predicts Variant Effects by Cell Type

Summary: On November 11, 2026, Google DeepMind announced AlphaGenome 2, a redesigned DNA sequence model that predicts how genetic variants affect specific human cell types rather than broad tissue averages. Trained on over 107 million individual cell profiles across 68 tissues, it roughly doubles the number of causal variants researchers can recover on a key benchmark. For anyone building in biotech or health tooling, it also ships as a gated API product — non-commercial research today, commercial access in private preview — which shapes what you can build with it right now.

Why the old parallel design hit a ceiling

Every cell in your body carries the same DNA, but each cell type switches different genes on and off. That regulation layer — the instructions that orchestrate gene activity, as DeepMind’s AlphaGenome science page puts it — is where most disease-linked variants act, and it is hard to model.

The first AlphaGenome took a million-letter DNA sequence and predicted molecular properties across thousands of experiments from many tissues at once. Efficient, but as the DeepMind team (Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec) writes, that parallel approach spread the model’s capacity thin and made it difficult to capture subtle differences between closely related cell types. A variant that matters only in gut macrophages gets averaged away when the model is predicting a whole tissue panel simultaneously.

AlphaGenome 2 attacks that problem structurally. It is a conditional model: alongside the DNA sequence, it takes a target biological condition — a tissue, cell type, or cell state — as input at every stage of the architecture. The model predicts one condition at a time, so it can learn how the same stretch of DNA is read differently across cell types.

The mechanism: conditioning plus a cell encoder

Two design choices do the heavy lifting:

Condition inputs from a reference catalog or your own data. Researchers can pick a condition from a catalog of cell types, or generate one on the fly. The model includes a cell encoder that compresses a single-cell RNA-seq snapshot from as few as a few dozen cells into a cell state representation, then conditions predictions on that unseen state — no retraining required.
Cross-modality imputation. Because the model separates a cell’s biological state from its outputs, inputting just one measurement (say, single-cell gene activity) lets it predict unmeasured properties like chromatin accessibility. That helps researchers decide which experiments to run next instead of running all of them.

The training data also changed. AlphaGenome 2 was trained on both bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles across 68 tissues. It predicts the same core readouts as the original — gene expression, histone modifications, chromatin accessibility — plus DNA methylation, which DeepMind notes was a widely requested addition.

The conditioning approach is what drew praise from outside the project. Professor Sarah Teichmann, co-founder of the Human Cell Atlas consortium, said: “What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining, predicting how DNA variants act in newly profiled cell states from even a small sample of cells.”

What the numbers show — and where they come from

On the TenK10K dataset, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 identified nearly twice as many causal variants (30% vs. 16%) and directional gene effects (43% vs. 23%) compared to the original AlphaGenome fine-tuned on single-cell data alongside other specialized tools. DeepMind also reports the model tracks continuous changes in gene expression — for example, as human blood stem cells mature into red blood cells and immune lineages — on genome regions outside its training data.

Two worked examples show what that buys in practice:

IBD and ETS2. A recent landmark study showed that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene in gut macrophages. Running predictions across gut immune cell types, AlphaGenome 2 reproduced this finding computationally, singling out the causal variant and showing it selectively boosts ETS2 in macrophages while leaving other cells and genes largely unaffected.
COVID-19 vaccination. Given unseen single-cell data from an mRNA vaccine study, the model distinguished vaccine-activated T cells from resting ones, and revealed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state — an effect invisible in unactivated cells.

One caveat worth keeping in mind: these benchmark gains come from DeepMind’s own evaluation in its announcement and accompanying paper, not from independent validation I can point to in the sources.

Precomputed scores and platform moves

Beyond the model itself, DeepMind precomputed AlphaGenome 2 predictions for GWAS trait-associated variants across hundreds of human cell and tissue conditions. The goal is to help researchers connect broad association signals to specific causal variants, target genes, and cell types — including enhancer–gene mapping in primary human cell types like brain astrocytes, where lab-grown CRISPR screen data is thin. This builds directly on the precomputed-variant-map approach of the earlier AlphaGenome Atlas, which I covered in AlphaGenome Atlas: A Precomputed Map of 9 Billion DNA Variants Changes How We Prioritize Genetic Research.

There is also a platform signal. DeepMind says AlphaGenome 2 Skills are coming to Google Antigravity, its agentic development platform, pairing specialized scientific models with Gemini 4 as a research agent. The excerpt of the announcement I have access to cuts off partway through the acknowledgements section, so I can’t confirm further release details from the page itself.

Peter Donnelly, CEO of Genomics Limited, indicated his company looks to incorporate the model into its Mystra platform — an early sign of commercial interest even while commercial access remains gated.

Access, limits, and what to check next

Availability is the practical constraint. AlphaGenome 2 is available now for non-commercial research via the AlphaGenome API, with interactive Colab tutorials for expression conditioning; commercial research requires a private preview. DeepMind is also collecting community use cases via a form through November 30, with selected projects eligible for additional quota and credits.

The stated limitation is firm: both AlphaGenome and AlphaGenome 2 “are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.” Nothing in the sources establishes pricing, commercial availability timelines, or clinical validity.

If you’re building in this space, the concrete next steps are: apply for the private preview and pin down its cost and quota terms before planning a commercial workflow; try the expression-conditioning Colab with a small single-cell sample to see whether the cell encoder handles your data’s cell states well; and if you have a compelling non-commercial use case, submit it before November 30 for the quota and credits offer. The model’s conditioning architecture is genuinely new — but the access terms, not the science, will decide what you can ship.

Sources

Original post

View the original post on X