DeepMind says new AlphaGenome 2 predicts how single-letter DNA changes matter in hundreds of cell types

DeepMind has unveiled AlphaGenome 2, a next-generation model it says can predict how single-letter changes in DNA alter gene regulation in hundreds of human cell types. The company describes the tool as able to point researchers to which cells are likely affected by a genetic variant - an advance that could aid efforts to find the cellular causes of disease.

AlphaGenome 2 is a conditional model: along with a long stretch of DNA it takes a defined biological condition as input, such as a tissue, a specific cell type or a cell state. That condition can come from a reference catalogue or be encoded from a scientist's single-cell RNA-seq snapshot (single-cell RNA sequencing measures gene activity in individual cells). DeepMind says this lets the model adapt to new cell states without retraining and to predict effects that broader tissue-level models missed.

DeepMind says the model was trained on more than 100 million individual cell profiles across 68 tissues. The company also published an AlphaGenome Atlas, a precomputed resource with predictions for all 9 billion possible single-nucleotide changes in the human genome; DeepMind describes the Atlas as about 1 petabyte of data and offers the Atlas and an AlphaGenome API to researchers. The model is available for non-commercial research via the API, with commercial access in private preview.

On single-cell benchmarks DeepMind reports substantial gains. On the TenK10K immune dataset of 28 primary immune cell types, AlphaGenome 2 identified nearly twice as many causal variants (30% versus 16%) and more directional gene effects (43% versus 23%) than the earlier AlphaGenome fine-tuned on single-cell data. The company also says the model reproduced a published finding in inflammatory bowel disease, identifying a variant that raises ETS2 activity in gut macrophages.

The model can also compress a small single-cell experiment into a cell-state representation, then predict unmeasured molecular readouts such as chromatin accessibility or DNA methylation. DeepMind says this cross-modality ability helps scientists prioritise which experiments to do next and can reveal variant effects that show up only in particular cell states, for example during infection or after vaccination.

But the usefulness of any prediction tool depends on the data it learns from. Large genome-wide studies and many biobanks underrepresent people of African ancestry: a 2025 review in Frontiers in Genetics found that over 90% of genome-wide association studies (GWAS) published as of 2023 were on individuals of European descent, and warned this limits how well genetic findings translate to people of African ancestry. Some large GWAS remain overwhelmingly European - for example a 2022 blood-pressure meta-analysis drew on about 1,028,980 European individuals - while African and sub-Saharan African cohorts such as AWI-Gen appear in the literature but on a much smaller scale.

DeepMind's blog and public documentation say AlphaGenome 2 was trained on human and mouse data and list the scale of the cell-profile training set, but they do not specify how many genomes or single-cell profiles come from people of African ancestry or name the African datasets included, if any. The company's materials also do not detail where the Atlas and API data are hosted by region, which could affect access and costs for researchers in Africa.

The technical and licence details that matter to African researchers are partly described in DeepMind's developer pages and code repository: the AlphaGenome Atlas is available through a free-to-use website portal and the AlphaGenome API is offered free for non-commercial research, with precomputed Atlas lookups supporting higher query rates than live model predictions. The repository says commercial use is available via Google Cloud and that outputs are for research only and not to be used for clinical decision-making or for training other machine learning models. Query rates, rate limits and pricing for commercial use are not given in the public announcement.

Those gaps matter because genetic variant frequencies and patterns of linkage differ between populations. The Frontiers review emphasises that African ancestry populations harbour the greatest global genetic diversity and that this diversity is often poorly captured by current reference datasets; tools trained mainly on European-centred data can miss or mis-prioritise variants that are common or important in African populations. In practical terms this can make it harder for researchers in Africa to use model predictions to guide experiments, interpret GWAS signals, or build clinical or public-health tools that work across ancestries.

DeepMind's announcement and documentation give several clear facts - the Atlas covers 9 billion possible single-base variants, is roughly 1 petabyte, and AlphaGenome 2 was trained on more than 100 million single-cell profiles across 68 tissues - but they stop short of saying how African genomes and cell types are represented in the training or reference sets, and they do not publish the cloud-hosting regions, bandwidth or commercial licence details that affect access from Africa.

For researchers in Africa the model will become more useful as reference sets and single-cell maps include more samples of African ancestry, and as hosting, query limits and commercial terms are clear and affordable. Clearer documentation on which datasets contributed African samples, and whether the Atlas or API can be hosted or cached in regions closer to African users, would help labs plan how to use the resource and judge how predictions apply to local populations.

Newzroom dot Africa has reached out to DeepMind, Human Cell Atlas and H3Africa for comment and is awaiting a response. This story will be updated when they respond.

Source: Google DeepMind