DeepMind's AlphaGenome 2 Predicts DNA Effects for Individual Cells: 30% vs 16%

Every cell in your body carries the same genome. But a heart muscle cell reads it very differently from an immune cell. The same mutation can be harmless in one tissue and devastating in another. Nobody has been able to predict this reliably in advance, and this is where Google DeepMind has taken a new step: it has released AlphaGenome 2.

What Was Wrong with the First Version

The first AlphaGenome took a DNA sequence about a million letters long and, in a single pass, predicted its properties across thousands of experiments spanning many tissues. According to DeepMind, this parallelism smeared the model's capacity. Subtle differences between closely related cell types were lost, like details in an over-compressed image.

The analogy for developers is simple. Imagine a single model that must output metrics for a thousand user segments at once and ends up giving an average answer for each. It is a familiar problem, and the solution is familiar too: add a condition.

How the Conditional Model Works

AlphaGenome 2 became a conditional model. Besides the DNA sequence, it takes as input a description of a biological state, that is, a specific cell type, and calculates effects only for it. It was trained on more than 107 million single-cell profiles from 68 human and mouse tissues.

The outputs have expanded too. The model predicts the strength of gene expression and how DNA is packed in the nucleus, that is, which regions of the genome are accessible. Now it also calculates DNA methylation, chemical marks that work as switches, turning genes on or silencing them.

There is also a practical trick: the model fills in missing data. If researchers measured only gene activity, it will predict DNA packing as well. This helps decide which lab experiments are worth running at all.

The key point from an engineering perspective: the model adapts to a new experiment without retraining. It is enough to provide gene activity data from roughly a few dozen cells, and it will calculate effects for a previously unseen state. According to DeepMind, it often outperforms specialized models trained specifically on such data.

What the Tests Showed

On the TenK10K benchmark with 28 types of primary immune cells, AlphaGenome 2 found about 30% of causal variants. In 43% of cases it correctly determined whether a variant strengthens or weakens gene activity. A fine-tuned AlphaGenome 1 managed only 16% and 23%.

Two more examples. Using pure computation, the model reproduced an already known mechanism of inflammatory bowel disease. It found a specific variant in a regulatory switch located far from the gene that selectively overactivates ETS2 in intestinal macrophages, while other cells and genes are barely affected. On data from a COVID-19 mRNA vaccine study, it distinguished activated T cells from resting ones and showed that a variant linked to severe disease disrupts cell function mainly in the activated state. In resting cells, this effect would have remained invisible.

DeepMind also precomputed predictions for hundreds of cell and tissue conditions. They are useful for analyzing genome-wide association studies (GWAS). Such studies link DNA regions to traits like height or type 2 diabetes, but point to them very coarsely. AlphaGenome 2 narrows the circle of likely causal variants and names the affected cell types.

Limitations and What to Do About Them

First, a sober look. The figures come from DeepMind itself. About 30% of causal variants found means most still slip away. The system is not approved for clinical diagnostics.

Access works like this: the model is open via API for non-commercial research, while commercial users get it through a closed preview. Biotech startups will have to negotiate separately.

Practical takeaways for those building products around scientific data. First, the "general model plus state condition" pattern works beyond biology, and the shift from averaged to conditional predictions is worth keeping in mind. Second, adaptation from a few dozen examples without fine-tuning greatly reduces the cost of experiments. Third, DeepMind plans to add AlphaGenome 2 features soon to the Google Antigravity platform together with Gemini 4 for automated research pipelines. If you build tools for labs or pharma, keep an eye on this direction now.