DeepMind Launches AlphaGenome 2, Cell by Cell, via API

DeepMind has made AlphaGenome 2 available for non-commercial research, with commercial access in a closed preview and no clinical endorsement. The model provides predictions at the cellular type level, shows quantified performance on TenK10K, and is set to join Google Antigravity alongside tools like Gemini 4.

No clinical use and limited access announced

AlphaGenome 2 is not approved for clinical diagnostics. The model is accessible for non-commercial research via an API, while commercial uses go through a closed preview. DeepMind plans to soon deploy dedicated features on Google Antigravity, where AlphaGenome 2 will be combined with tools like Gemini 4 to automate research workflows. Meanwhile, DeepMind indicates it has precalculated predictions for hundreds of cellular and tissue conditions, particularly to support AI-driven drug development efforts.

Quantified results on TenK10K and validation cases

On the TenK10K dataset, which includes 28 categories of primary immune cells, AlphaGenome 2 was able to identify nearly 30% of causal variants and correctly anticipated the direction of effect in 43% of cases. A modified version of AlphaGenome 1 only exceeded 16% and 23% for these two indicators. The system also takes into account continuous trajectories, such as the differentiation of blood stem cells into red blood cells, achieving significantly higher accuracy for this. When applied to data from an mRNA vaccine against COVID-19, the model was able to differentiate activated T cells from resting cells and showed that a variant associated with severe forms primarily affects the activated state, an impact that does not manifest at rest. Purely computationally, it reproduced a mechanism of inflammatory bowel disease and isolated the variant located in a distant regulatory switch selectively activating ETS2 in intestinal macrophages, with other cell types and genes being largely unaffected. In more general tests, the new model found nearly twice as many relevant variants as its predecessor.

What the model can predict at the cellular type level

AlphaGenome 2 takes into account individual cellular states and calculates gene activity, the physical organization of DNA, and chemical switches. It predicts the structure of chromatin in the nucleus and the fraction of the genome that is accessible, and includes predictions of DNA methylation. The model can also infer missing modalities from a single measurement: if only genetic activity is available, it provides estimates for additional properties such as DNA packaging.

Immediate applications in genetics and screening

The precalculated predictions of AlphaGenome 2 are used to analyze genome-wide association studies (GWAS). These GWAS link regions of DNA to traits such as height or type 2 diabetes but often localize signals imprecisely. The model narrows down the set of likely causal variants and identifies the affected cell types, such as astrocytes or immune subpopulations. When its confidence is sufficient, it indicates whether a mutation increases or decreases the expression of a gene, information that is useful for drug development. In the lab, the ability to complete measurements also helps in selecting subsequent experiments, while already calculated predictions cover hundreds of cellular and tissue states.

Technical choices and break from AlphaGenome 1

AlphaGenome 1 evaluated sequences of about one million letters and projected their properties across thousands of experiments and tissues simultaneously. DeepMind reports that this parallel approach diluted the model's capacity and complicated the detection of fine differences between closely related cell types. AlphaGenome 2 opts for a conditional model: it takes a specific biological state as input and calculates the effects for a unique cell type. Designed to predict the impact of sequences and DNA variations on human cell types, it has been trained, according to DeepMind, on over 107 million profiles from 68 tissues, with data covering humans and mice. The stated goal is to address changes at the cellular type level and help identify cellular drivers of diseases.