Last year, Google DeepMind introduced AlphaGenome to predict how DNA variants influence molecular activity across broad tissues. The company’s new version, announced Wednesday, narrows that frame to individual cell types and states—an important distinction when identical DNA can be read very differently by a heart cell, a skin cell or an activated immune cell.¹

DeepMind says AlphaGenome 2 was trained on more than 100 million individual cell profiles spanning 68 tissues. Rather than issuing broad, parallel tissue predictions, it takes a biological condition alongside a DNA sequence, aiming to identify variant effects that tissue-level averages can obscure. The company says it found nearly twice as many variants affecting cellular gene expression as its earlier leading model, including signals tied to disease-relevant cells.¹

Sundar Pichai framed the launch as a tool for “uncover[ing] the cellular drivers of disease,” with the ability to predict how variants affect specific cell types.Sundar Pichai (sundarpichai)² DeepMind chief Demis Hassabis similarly called the cell-level predictions “crucial information for understanding complex diseases.”Demis Hassabis (demishassabis)³

The potential use case is straightforward but consequential: researchers could take variants flagged in large genetic studies and investigate their likely effects in the precise cells where illness emerges. Google’s researchers told The Verge that this could help connect human genetics to cell biology and, eventually, drug development, where targets are often specific proteins in specific cells.⁴

But the upgrade does not simply replace AlphaGenome. Žiga Avsec, DeepMind’s genomics lead, said the model’s more granular predictions are “more expensive” to produce—too demanding, for now, to generate an atlas covering every possible single-letter change across the genome.⁴ AlphaGenome 2 is available through an API for non-commercial research, while commercial access remains in private preview.¹