Google DeepMind is moving its AlphaGenome system from predictions averaged across tissues toward the much finer question of what a particular DNA change does in a particular cell. Announced on November 11, AlphaGenome 2 is presented as a research tool accessible through AlphaGenome resources. Model predictions should not be treated as clinical diagnoses.

From one genome to many cellular outcomes

Almost every cell in a human body contains the same DNA, yet immune cells, neurons and heart cells follow different instructions. That difference is central to disease: a variant that barely changes activity in one tissue may alter gene expression in a specific cell population. The original AlphaGenome predicted molecular properties from DNA sequences across broad tissue measurements. The new model accepts information describing a cell type or state, allowing predictions to focus on one context rather than distributing the model's capacity across many.

DeepMind says it trained AlphaGenome 2 with approximately 107 million single-cell profiles drawn from human and mouse samples across 68 tissues. Besides predicted gene expression, the system can estimate aspects of chromatin accessibility, histone modifications and DNA methylation. DeepMind says researchers can use newly supplied single-cell RNA-seq data to make predictions for previously unseen cell states. That capability could narrow the set of laboratory experiments researchers need to run.

What the headline improvement actually measures

On a single-cell immune benchmark called TenK10K, DeepMind reports that AlphaGenome 2 identified 30% of relevant causal variants, compared with 16% for the benchmark comparison methods. It also reports identifying the direction of gene effects in 43% of cases, versus 23% for those baselines. Those are results from the company's evaluation, not independent proof that the tool will double discovery rates in live laboratories or improve patient outcomes.

One example concerns inflammatory bowel disease. DeepMind says the model computationally recaptured research linking a distant DNA regulatory switch to increased ETS2 activity in gut macrophages. That is informative because it shows why cell identity matters: the same genome change need not alter every neighboring cell. But reproducing a known biological association is not the same as prospectively discovering and validating a new disease mechanism.

Why scientists may care

Genetic association studies often find signals in non-coding DNA, far from the protein-coding genes most people associate with inherited traits. Finding a statistical connection between a variant and disease is only the first step. Researchers still need to ask which cells are affected, which genes change, and whether the predicted pathway survives controlled experiments. A context-specific model could help turn a large list of candidate variants into a smaller, better targeted experimental plan.

Access is another practical distinction. Researchers can explore the AlphaGenome research resources, including its API, but a computational benchmark does not establish clinical utility. Translating model predictions into patient-care decisions would require separate clinical assessment, validation and regulatory scrutiny. Independent replication, performance across populations and disease contexts, and wet-lab confirmation remain necessary before medical applications can be justified.

The immediate news, then, is a stronger research interface between genetic data and cell biology—not an AI cure or a replacement for experiments. Its value will be established by whether outside scientists can generate useful, reproducible hypotheses in cells and diseases the model's developers did not specifically highlight.

Sources