Google unveils AlphaGenome 2: model identifies nearly twice as many consequential DNA variants

Finding a DNA change associated with a disease is only the beginning. Researchers then need to understand where its effects appear: in heart, skin or immune cells. The answer may also depend on the cell's state. Looking too broadly can easily obscure an important effect. For a team developing a drug, that uncertainty makes choosing the next experiment harder.

Google DeepMind has unveiled AlphaGenome 2, a model that predicts the effects of genetic variants based on cell type and state. According to the company, it identifies nearly twice as many variants capable of changing gene activity as its previous best model. The practical benefit is a more specific hypothesis to test in the lab: which change to investigate, in which cells and under what conditions.

Why the model needs to distinguish between cells

Almost all human cells contain the same DNA, but they use it differently. A heart muscle cell and an immune cell need different sets of active genes. Even within a single cell type, the picture changes depending on what is happening in the body.

Imagine researchers have identified a genetic variant that is more common in people with a particular disease. That alone does not explain the disease mechanism. The variant might affect gene activity only in certain immune cells during an infection. In other cells, its effect might be weak or undetectable. This is a hypothetical example of a problem that calls for that level of detail.

AlphaGenome 2 is intended to help narrow the search. Instead of a general assessment of a DNA change, researchers get a prediction tied to its cellular context. This matters for drug development: a drug acts on specific biological targets, and the outcome depends in part on which cells those targets are active in.

What a hundred million cell profiles contributed

The model was trained on more than 100 million individual cell profiles from 68 tissues. These data can capture differences that disappear when tissue is examined at a broader level. DeepMind uses cell atlases—large collections of experimental observations—as the foundation for predicting the consequences of DNA changes.

According to the company, AlphaGenome 2 can also make predictions for new cell types and states without retraining. This is a promising capability for research teams: experimental data cover biology unevenly, and collecting detailed measurements for every case is impossible.

Another application is identifying gaps in existing data. If the model helps determine which observations are missing, a team can make more informed choices about its next laboratory studies. For a product built around such a system, the value could lie in helping plan experiments: which hypotheses to test first and where additional measurements are needed.

But the claimed nearly twofold advantage should not be interpreted as a universal measure of model accuracy or a promise to halve drug development time. The material presented refers to the number of identified variants that affect gene activity. There are not enough evaluation details to generalize that result to every research task.

What to consider before using the API

More detailed predictions require more computing resources. For that reason, AlphaGenome 2 does not yet replace the original AlphaGenome in every scenario. According to DeepMind, using the new model for its previously published atlas of all possible single-letter DNA substitutions in the human genome would be impractical.

This suggests a straightforward approach: assess the scale of the task first, then decide where cell-level detail is needed. A broad search and a detailed analysis of a small group of variants may call for different tools. The economics of that workflow will need to be assessed using your own data: the source material provides no request pricing or performance figures.

Google announced access to AlphaGenome 2 through the AlphaGenome API for noncommercial research. Commercial research is offered limited, restricted access. Founders of biotech products should factor this into integration planning from the outset: availability for research use does not in itself mean a commercial service is available.

Predictions also require experimental validation. AlphaGenome 2 helps connect a genetic variant to a possible cellular mechanism, but that does not establish a cause of disease or deliver a ready-to-use therapy. For biotech developers and product managers, the immediate task is to test whether this additional detail helps them choose more useful experiments.