AlphaGenome 2 brings cell-specific predictions to human genetics research

Google DeepMind released AlphaGenome 2 on November 11. The model predicts how a DNA variant could affect molecular activity in particular human cell types. Researchers can now investigate a question that a tissue-wide average can obscure: which cells might account for a signal? DeepMind’s release announcement

What changes when researchers can look at individual cell types?

A tissue contains several kinds of cells. A change that matters in one relatively rare population can look small when its signal is averaged with those of its neighbours. The original AlphaGenome made predictions across broad tissues. Its successor draws on more than 100 million cell profiles from 68 tissues to make finer distinctions.

DeepMind reports that AlphaGenome 2 identified nearly twice as many variants affecting gene expression in cells as its previous leading model. That comparison concerns a particular research task. It tells us little by itself about how accurately the model will answer a new lab’s question. For the evaluation methods and scope of the result, see the technical report.

What access is available?

DeepMind offers the model for non-commercial research through the AlphaGenome API. Commercial research access is in private preview. A team evaluating a commercial project should establish whether it can obtain that access before committing to a delivery date. The official documentation is the starting point for the research interface.

How should a research team interpret a prediction?

My advice is to decide in advance what independent evidence would change the team’s conclusion. If a prediction points toward a particular cell population, it can help researchers decide where to spend scarce follow-up resources. The prediction alone cannot establish that a variant caused a person’s illness.

For an internal review, record the model version, the question the team asked and which conclusions still depend on external evidence. This makes it easier to revisit a decision as measurements or model versions change. DeepMind explicitly limits AlphaGenome 2 to research use and says it has not been validated or approved for clinical or diagnostic purposes. DeepMind’s model overview

For teams evaluating a research API, my AI consulting work covers access, integration and the evidence needed to judge a pilot. The AI Web Awards project is a separate example of my platform work.