AlphaGenome 2: DeepMind's AI Reads DNA Cell by Cell

Every cell in your body carries essentially the same DNA. Yet a heart muscle cell, a neuron and an immune cell fighting an infection behave very differently. The difference lies in which genes each cell switches on or off. A tiny typo in the DNA may matter in one cell type and do nothing in another. That is one of the hardest puzzles in genetics, and it is the one Google DeepMind's new model, AlphaGenome 2, is built to attack.

DeepMind announced AlphaGenome 2 on November 11, 2026. It describes the model as a next-generation system that predicts how DNA sequences, and their variations, affect specific human cell types. Let's unpack what that means, what the numbers say, and what we should keep in mind before getting carried away.

The problem: same blueprint, many different cells

Think of DNA as a huge cookbook that every cell owns a copy of. A cell does not cook every recipe at once. A blood stem cell maturing into a red blood cell uses a different set of recipes than a gut macrophage. Which recipes get used is called gene expression, and it is a key driver of human traits and disease.

According to DeepMind, global efforts like the Human Cell Atlas are mapping gene expression across millions of cells. But predicting how subtle DNA variations change that expression inside each kind of cell is still, in their words, an unsolved challenge. That matters because most genetic variants linked to disease do not sit inside genes in an obvious way. Many sit in distant stretches of DNA that act as regulatory switches.

If you are a researcher staring at a list of variants linked to a disease, the question is not only "which variant?" but also "which gene does it affect?" and "in which cell?" Without that third answer, you may run expensive experiments in the wrong cells.

From AlphaGenome to AlphaGenome 2

This is a sequel. DeepMind says it launched the original AlphaGenome last year to predict how DNA variants affect molecular activity across broad tissues. It then expanded the reach with AlphaGenome Atlas, a platform with predictions for all 9 billion possible single-letter changes in the human genome. These tools help researchers prioritize and interpret variants, with applications in rare diseases, complex traits and basic biology.

The catch with the first model was granularity. It took a million-letter DNA sequence and predicted properties across thousands of experiments from many tissues at once. DeepMind explains that this parallel approach spread the model's capacity thin and made it hard to capture subtle differences between closely related cell types. It is a bit like asking one student to answer a thousand exam questions simultaneously: they get the broad strokes but miss the nuance.

AlphaGenome 2 changes the approach. Instead of predicting everything in parallel, it is a conditional model. You give it the DNA sequence plus a target biological condition, meaning a tissue, a cell type or a cell state, and it predicts the effects for that one condition. The condition is fed into every stage of the architecture, so the model can learn how the same stretch of DNA is read differently in different cells.

What it was trained on and what it predicts

DeepMind says the model was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. Single-cell data is the key ingredient. Rather than averaging a whole tissue sample, it measures activity cell by cell, which lets the model resolve, for example, distinct heart muscle cells or blood immune cells.

It predicts the same core readouts as the original: gene expression, histone modifications and chromatin accessibility. It now adds DNA methylation, which DeepMind calls a widely requested readout. According to the company, it also identifies nearly twice as many variants that change gene expression in cells as its previous state-of-the-art model.

One neat side effect of the conditional design is what DeepMind calls cross-modality imputation. Because the model separates a cell's biological state from its outputs, a scientist can input a single measurement, such as single-cell gene activity, and the model can predict unmeasured properties like chromatin accessibility. In practice, that could help researchers decide which experiments are worth doing next. That is a very practical benefit: lab time is expensive, and a good prior can save months.

The headline results

DeepMind highlights three kinds of improvement. First, better cell-type specificity: on genome regions outside its training data, the model explains much more of how gene expression differs across cell types within a tissue. That includes tracking continuous changes, like blood stem cells maturing into red blood cells and immune lineages.

Second, better variant effect prediction. On the TenK10K dataset, a large single-cell map of 28 primary immune cell types, DeepMind reports that AlphaGenome 2 reliably identified causal variants 30% of the time versus 16% for the original AlphaGenome fine-tuned on single-cell data and other specialized tools. For directional gene effects, meaning whether a variant turns a gene up or down, the figures were 43% versus 23%. These are the company's own reported numbers, so independent replication will be the real test.

Third, a case study. A landmark study had shown that a variant in a distant regulatory switch drives inflammatory bowel disease (IBD) by overactivating the ETS2 gene in gut macrophages. DeepMind says AlphaGenome 2 reproduced this finding computationally. It singled out the causal variant and showed that it selectively boosts ETS2 in gut macrophages, leaving other cell types and genes largely unaffected. Rediscovering a known result is not the same as discovering a new one, but it is a sensible sanity check.

New cell states without retraining

Perhaps the most interesting feature is that the model can handle cell states it has never seen. Scientists can upload a single-cell RNA-seq snapshot from just a few dozen cells. That gene activity is compressed into a representation of the cell's state, which then conditions the model. DeepMind says this can match or outperform models trained or fine-tuned directly on those datasets.

Their example is a COVID-19 mRNA vaccine study. Given unseen single-cell data, the model distinguished vaccine-activated T cells from resting ones. It also indicated that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state, an effect that would be missed in resting cells. Professor Sarah Teichmann, co-founder of the Human Cell Atlas consortium, is quoted by DeepMind praising how the model adapts to new single-cell data without retraining.

I find this idea compelling. Disease is rarely a static condition. Cells change state when they fight infection, when they inflame, when they respond to a drug. A model that can follow those shifts is closer to biology as it actually happens.

Linking genetics to cell biology

Genome-wide association studies (GWAS) have found thousands of variants associated with traits and diseases, but they usually flag broad stretches of DNA rather than one culprit. DeepMind says it has precomputed AlphaGenome 2 predictions for trait-associated variants across hundreds of human cell and tissue conditions so researchers do not need to run massive computations themselves.

Two challenges are targeted. One is mapping enhancer-gene links in real human cell types. Experimental CRISPR screens mostly use stable, lab-grown cell lines, so scientists often lack these wiring diagrams for specific cells in the body. The model extends the mapping to primary human cells, such as brain astrocytes. The other is pinpointing causal variants and the cells where they act, across traits ranging from height to type 2 diabetes. When confident, DeepMind says the model also accurately predicts whether a variant raises or lowers gene activity.

Peter Donnelly, CEO of Genomics Limited, said in a statement included by DeepMind that human genetics is one of the most powerful tools in drug discovery, and that sequence-prediction models help bridge the gap from an association signal to what a drug programme needs. His company plans to incorporate AlphaGenome 2 into its Mystra platform. That is a hint at where commercial interest lies.

Access, limits and my take

AlphaGenome 2 is available today for non-commercial research through the AlphaGenome API, and in private preview for commercial research. DeepMind also says it will soon release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform.

Now the fine print, which DeepMind states plainly: AlphaGenome and AlphaGenome 2 are for research use only and have not been validated or approved for clinical or diagnostic purposes. So this is not a tool your doctor will use to read your genome next month. It is a way to help scientists prioritize hypotheses. A prediction is a lead, not a verdict, and it still needs experimental confirmation.

There are other reasonable questions. The benchmark gains come from the developers' own evaluations. A 30% rate of identifying causal variants is a big improvement over 16%, but it also means most calls are not captured. And models learn from the data they are given, so the cell types and populations that are well profiled will be better served than those that are not. I would like to see how performance holds up for under-studied cell types and for people from diverse ancestries, though the post doesn't address that.

Still, the direction feels right. AI in medicine often gets attention for flashy diagnosis claims, like Microsoft's AI beating doctors on hard cases, and those deserve scrutiny. Tools like AlphaGenome 2 work further upstream, in the unglamorous but crucial job of understanding why disease happens. That kind of foundational work is what eventually makes personalized medicine and better help for rare diseases possible. It also sits alongside other biology-focused AI efforts, like AI-designed proteases, where models start to act as lab partners rather than just chatbots.

What it means for the rest of us

Most of us will never run a variant-effect prediction. But the consequences may reach us indirectly: faster identification of drug targets, smarter experiments, maybe earlier answers for families living with conditions that have no explanation. The honest version is that this will take years, and much will depend on lab validation, open access and who gets to use these tools.

So here is a question to take with you. If an AI can tell scientists which of your cells a genetic quirk affects, who should have access to that knowledge, and how do we make sure it benefits patients and not only pipelines? How do you feel about AI helping to read the code of life?