Google DeepMind releases AlphaGenome 2, trained on 100M+ cell profiles across 68 tissues

Google DeepMind has released AlphaGenome 2, a model trained on more than 100 million individual cell profiles across 68 tissues to predict how genetic variants affect specific cell types and states. Google says it is available for non-commercial research through the AlphaGenome API from today, with a limited "private preview" for commercial research.

What AlphaGenome 2 does

AlphaGenome 2 is the successor to AlphaGenome, which DeepMind released last year, according to The Verge. Nearly every cell in the body carries the same DNA, but different cell types switch different genes on and off, and those patterns shift with what a cell is doing. The same variant can therefore have very different effects depending on the cell. AlphaGenome 2 is built to predict variant effects at the level of individual cell types and states.

Google said in a blog post published Wednesday that the model offers a much finer-grained view than broader models like AlphaGenome. DeepMind says it effectively turns static cell atlases, large repositories that catalog individual cells, "into a predictive engine for human genetics."

Claimed results

• According to the company, AlphaGenome 2 identifies nearly twice as many variants that can change gene expression in cells as its previous state-of-the-art model.
• The researchers found it can make predictions about new cell types and biological states without retraining.
• It can also identify gaps in experimental data, which could help scientists decide what to investigate next in the lab.

These results are reported by DeepMind. No benchmark scores, parameter count, context window or pricing were given. Pricing is not yet disclosed.

Compute cost and relationship to AlphaGenome

Žiga Avsec, DeepMind's genomics lead, told The Verge that AlphaGenome 2 is a successor but will not replace AlphaGenome any time soon. Its more detailed predictions are "more expensive" to produce. That cost would have made it impractical for the genome-wide Atlas covering "every possible DNA letter change in the human genome," which DeepMind released in September.

Why DeepMind says granularity matters

Avsec said cell-type specificity matters for treatment goals: "When you develop drugs, you're targeting specific proteins and specific cell types." Adam Gayoso, a DeepMind research scientist on the project, said the key advance is "tying human genetics to cell biology." He said the model lets researchers drill into variants identified in large populations, see their cellular effects, and potentially connect them to disease.

Avsec said building such a model has only recently become possible because new technologies capture individual cells and their states at high resolution and generate the datasets needed for training. DeepMind is working to generate more data, including for cell states that are poorly represented today. He said future versions could be trained on data "across multiple species." Longer term, he said models like AlphaGenome 2 could feed into efforts to build "virtual cells," digital simulations designed to mimic biological systems.

What this means

AlphaGenome 2 targets a specific limitation: variant-effect predictions that do not distinguish between cell types. DeepMind's claim of nearly twice as many expression-altering variants identified is the main performance figure so far, and it comes from the company rather than independent evaluation. Because the model is more expensive to run, DeepMind is positioning it alongside AlphaGenome, not as a replacement, and it was not used for the genome-wide Atlas. Access is currently limited to non-commercial research via the API, with commercial use restricted to a limited private preview.