DeepMind's AlphaGenome 2 Finds Nearly Twice the Gene-Altering DNA Variants

Every cell in your body carries the same DNA, yet a heart muscle cell and an immune cell fighting an infection read it in very different ways. AlphaGenome 2, which Google DeepMind released on November 11, is built around that difference: give it a million letters of DNA, name one cell type or state, and it predicts what a small change in that DNA will do inside that cell. Trained on more than 107 million individual cell profiles, it identifies nearly twice as many variants that change gene activity as the best model before it. For geneticists hunting the single letter behind a disease, that means sharper leads in the cells where the disease actually happens.

A Million Letters, Read One Cell Type at a Time

The first AlphaGenome, launched in June 2025, took a million-letter stretch of DNA and predicted thousands of experimental readouts across many tissues all at once. The parallel design was efficient, but DeepMind says it spread the model's capacity thin and blurred the line between closely related cell types. Disease often lives on that line: a variant can sit quietly in one cell and push a gene too hard in its neighbor.

AlphaGenome 2 is a conditional model. Alongside the DNA, about two novels' worth of letters, it takes one target condition, a tissue, a cell type or a cell state, and feeds it into every stage of its architecture. Trained on bulk and single-cell data from human and mouse samples across 68 tissues, it learns how the same stretch of DNA is read differently from one cell to the next.

The flexible part is new data. A scientist can upload a snapshot of gene activity from just a few dozen cells, and the model works on that unseen state without being retrained, matching or beating models trained directly on those datasets.

“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.”— Sarah Teichmann

Teichmann co-founded the Human Cell Atlas consortium, the global effort mapping gene activity across millions of cells. The question is whether the extra precision shows up in the results, and DeepMind put a number on it.

From 16% to 30% Across 28 Immune Cell Types

The test was TenK10K, a large single-cell map of 28 primary immune cell types. Among the many variants near each gene, the job is to pick out the few that actually change its activity, the first step in almost any genetic study of a disease.

The original AlphaGenome, even fine-tuned on single-cell data, and other specialized tools found about half as many. The gap was similar on the next question, whether a variant turns its gene up or down: 43% against 23%.

A real case shows what that precision buys. A recent landmark study traced inflammatory bowel disease to a variant in a distant regulatory switch that overactivates a gene called ETS2 in gut macrophages, a type of immune cell. Asked to predict across the immune cells of the gut, AlphaGenome 2 reproduced the finding on a computer: it singled out the causal variant and showed that it boosts ETS2 in gut macrophages while leaving other cell types and genes largely unaffected.

The model can also follow cells as they change. Given unseen data from a COVID-19 mRNA vaccine study, it told vaccine-activated T cells apart from resting ones, and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly once those cells are activated. It also adds DNA methylation, the chemical tags that help switch genes off, a readout DeepMind says researchers widely requested. So who gets to put all this to work, and how soon?

Open to Labs Today, and Headed for Drug Discovery

AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, with interactive Colab tutorials, and in private preview for commercial research. DeepMind has also precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions, for traits ranging from height to type 2 diabetes, so researchers can use them without running large computations of their own. Teams that share a use case by November 30 may get extra quota and credits, and AlphaGenome 2 skills for Google Antigravity, the company's agentic development platform, are coming soon.

Drug hunters have a stake too. Genome-wide association studies flag broad stretches of DNA linked to a disease, but rarely the exact variant, gene and cell type. Peter Donnelly, chief executive of Genomics Limited, calls AlphaGenome 2 an impressive advance in bridging that gap and says his company plans to build it into its Mystra platform. The Wellcome Sanger Institute is working with DeepMind to profile an even wider range of human cell types.

What is left to solve is the step from prediction to patient. DeepMind says AlphaGenome and AlphaGenome 2 are for research use only and have not been validated or approved for clinical or diagnostic purposes, so every prediction is a lead for a lab to confirm. What has changed is how precisely those leads point.

A genetic clue that once pointed at a broad stretch of DNA can now point at one variant, in one kind of cell.