Google’s AlphaGenome 2 zooms in on the cellular roots of disease

Almost every cell in the body carries the same DNA, but heart cells, skin cells and immune cells switch genes on and off differently, and those patterns shift with what a cell is doing. So one genetic variant can be harmless in one cell and harmful in another. Google DeepMind's AlphaGenome 2 is built to predict those cell-specific effects.

That matters because, as DeepMind's genomics lead Žiga Avsec told The Verge, drugs target specific proteins in specific cell types. A model that links population-level genetic variants to what happens in a particular cell could shorten the path from 'this gene is linked to a disease' to 'here is where to intervene.'

This is the kind of AI story we're glad to lead with. No chatbot, no hype cycle, just a hard biology problem that a model may be able to help with. DeepMind says AlphaGenome 2 was trained on more than 100 million individual cell profiles across 68 tissues and finds nearly twice as many variants that change gene expression as its previous state-of-the-art model. It can also predict for cell types and states it wasn't trained on without retraining, and it can point to gaps in experimental data, which could help labs decide what to test next.

The usual caveats apply, and they're real. These are the company's own claims, reported through The Verge and a Google blog post. A model that predicts gene expression changes isn't a model that cures anything. The predictions still have to survive wet-lab validation, and the history of computational biology is full of signals that looked strong on screen and faded at the bench.

I'm also watching access. The model is free for non-commercial research through the AlphaGenome API, while commercial research gets only a limited private preview. That's a reasonable start, but it means academic labs can explore while the drug-discovery value runs through Google's terms. As these models become infrastructure for biology, who gets access on what terms will shape who benefits. And Avsec is upfront that AlphaGenome 2 is more expensive to run than its predecessor, which is why it didn't power the genome-wide Atlas DeepMind released in September. Finer resolution costs compute.

The longer arc is the one Avsec sketched: training across multiple species and feeding into virtual cells that simulate biological systems. That's ambitious and still distant. But it's the right ambition, and it puts AI to work on disease rather than on engagement.

AlphaGenome launched last year as DeepMind's genome-predicting model. AlphaGenome 2 is a successor, not a replacement. The release lands as Google, Anthropic and OpenAI all push into science and medicine, with DeepMind expected to open an automated materials science lab in the UK.