AlphaGenome 2 nearly doubles variant detection in human cell types

Google has announced AlphaGenome 2, a DNA sequence model that predicts how genetic variants affect gene expression in specific human cell types — and it identifies nearly twice as many variants that change gene expression in cells as its predecessor. The company trained the model on more than 100 million individual cell profiles across 68 tissues, and released it today for non-commercial research via the AlphaGenome API, with commercial research access in private preview.

The launch matters because the hardest problems in human genetics sit at the intersection of two massive datasets. Genome-wide association studies have flagged thousands of disease-linked variants, while efforts like the Human Cell Atlas are mapping gene expression across millions of cells. Nobody has had a reliable computational bridge between the two. AlphaGenome 2 is Google's attempt to build it, and Google says it turns static cell atlases into a predictive engine for human genetics.

What changed from AlphaGenome to AlphaGenome 2?

The original AlphaGenome, launched last year, predicted how DNA variants affect molecular activity across broad tissues. Google later expanded it with AlphaGenome Atlas, a platform holding predictions for all 9 billion possible single-letter changes in the human genome. Those tools helped researchers prioritize and interpret variant effects in rare diseases, complex traits, and fundamental biology.

But the first model had a structural weakness. It took a million-letter DNA sequence and predicted properties across thousands of experiments from many tissues simultaneously. That parallel approach spread the model's capacity thin, Google says, and made it hard to capture subtle differences between closely related cell types.

AlphaGenome 2 fixes this with a conditional architecture. Instead of predicting everything at once, it predicts DNA effects for one condition — a tissue, cell type, or state — at a time, taking that biological condition as input at every stage alongside the DNA sequence. The result: it learns how the same stretch of DNA gets read differently across cell types. The takeaway is simple — the model stops averaging and starts specializing.

The conditional design also enables cross-modality imputation. The model separates a cell's biological state from its outputs, so a single measurement like single-cell gene activity lets AlphaGenome 2 predict unmeasured properties such as chromatin accessibility. That helps researchers decide which experiments to run next. Google trained the model on bulk and single-cell data from human and mouse samples, over 107 million cell profiles in total. It predicts the same core readouts as before — gene expression, histone modifications, chromatin accessibility — plus DNA methylation, a widely requested addition.

How much better are the predictions?

Google reports three concrete gains.

• Cell type specificity. On genome regions outside its training data, AlphaGenome 2 explains significantly more of how gene expression differs across cell types within a tissue, including continuous changes such as human blood stem cells maturing into red blood cells and immune lineages.
• Variant effect prediction. On the TenK10K dataset, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 reliably identified nearly twice as many causal variants (30% vs. 16%) and directional gene effects (43% vs. 23%) as the original AlphaGenome fine-tuned on single-cell data and other specialized tools.
• Disease-relevant variants. A recent landmark study showed that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene in gut immune cells called macrophages. AlphaGenome 2 reproduced this finding computationally, singling out the causal variant and showing it selectively boosts ETS2 in gut macrophages while leaving other cell types and genes largely unaffected.

The IBD result is the most telling detail. It shows the model can find variants that act in one cell type and stay silent elsewhere — exactly the kind of effect that broad-tissue models miss.

Can it handle cell states it has never seen?

Yes, without retraining. Scientists can upload a single-cell RNA-seq snapshot of gene activity from just a few dozen cells. The model compresses that activity into a cell state representation and uses it to condition predictions on unseen states. Google says AlphaGenome 2 matches or outperforms models trained or fine-tuned directly on those datasets.

That matters for disease and vaccination states, which can't all be anticipated at training time. Google gives one example: given unseen single-cell data from a COVID-19 mRNA vaccine study, AlphaGenome 2 distinguished vaccine-activated immune T cells from resting ones. It also showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state — an effect invisible in unactivated cells. In short, the model adapts to new biology on the fly.

How does it connect genetics to cell biology?

Google precomputed AlphaGenome 2 predictions for trait-associated variants across hundreds of human cell and tissue conditions, so researchers can query them without running large-scale computations themselves. This tackles two long-standing problems.

• Mapping enhancer–gene links in real human cell types. Many disease-linked variants sit in distant regulatory switches called enhancers, and experimental CRISPR screens mostly use stable lab-grown cell lines, leaving gaps in wiring diagrams for cells in the human body. AlphaGenome 2 extends enhancer-gene mapping to primary human cell types such as brain astrocytes, which support and protect neurons.
• Pinpointing causal GWAS variants. Association studies flag broad stretches of DNA linked to disease but don't identify the exact variant, target gene, or cell type. Across complex traits ranging from height to type 2 diabetes, the precomputed scores help pinpoint the most likely variants and cell types — and, when confident, whether a variant turns a gene's activity up or down.

What comes next?

Google says it is collaborating with researchers at the Wellcome Sanger Institute to profile a wider spectrum of human cell types, developmental lineages, and perturbed cellular states. It will soon release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform, pairing the specialized model with agentic science tools like Gemini 4 — a combination Google says will help scientists accelerate research.

The company is soliciting community use cases via a form with a 30 November deadline, and selected projects may receive additional quota and credits. Google also invites feedback through its community forum to shape the roadmap.

Two caveats anchor the story. AlphaGenome 2 remains research-only: Google states clearly that it has not been validated or approved for clinical or diagnostic purposes. And the performance claims are Google's own, reported against its earlier models and other specialized tools. The real test will come as independent labs use the API to hunt variants in their own disease data — starting now, since non-commercial access is open today.