Google DeepMind unveiled AlphaGenome 2 on Wednesday. The model predicts how DNA sequences and their variations affect specific human cell types. Its predecessor could do that for broad tissues, and that was exactly where it ran out of road.
Every cell in the body carries the same DNA blueprint, yet a heart muscle cell and an immune cell fighting an infection switch different genes on and off. The same genetic variant can therefore behave completely differently depending on the type and state of the cell. Predict tissue averages only, and that difference disappears.
What changed in the architecture
AlphaGenome 1 produced fixed parallel outputs for bulk tissues and cell lines from a one-million-letter DNA sequence. AlphaGenome 2 is a conditional model: along with the sequence it receives a biological condition. That condition either comes from a reference catalogue of cell types or is generated on the fly from a scientist’s own single-cell experiment, through what DeepMind calls a cell encoder.
The model was trained on more than 100 million individual cell profiles across 68 tissues. According to DeepMind it identifies nearly twice as many variants that change gene expression in cells as the company’s previous best model, and it also produces predictions for cell and disease states it never saw during training. DeepMind’s own framing: static cell atlases become a predictive engine for human genetics.
AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, with a private preview for commercial research.
Back in September we wrote about AlphaGenome Atlas and its nine billion variants, the platform holding predictions for every possible single-letter change in the human genome. AlphaGenome 2 zooms into the same question one level further down.
DeepMind also shipped Gemini Robotics 2.5 on Tuesday, a model for physical intelligence in the real world. Two model launches on two days, and neither has anything to do with chatbots.
Why 68 matters more than 100 million
The big number in the post is the 100 million cell profiles. The one doing the work is 68: that is how many tissues the training covers. Only that lets the model judge a variant inside a disease-relevant cell rather than averaged across a whole organ.
In practice this means researchers can triage candidates before they walk into the lab. Whether the predictions hold up gets settled by experiments nobody has run yet, and opening the API to non-commercial research is the fastest way to find out.
Sources
- Google DeepMind: AlphaGenome 2: Deciphering how cells read DNA to better understand disease
- The Verge: Google’s AlphaGenome 2 zooms in on the cellular roots of disease
- Google DeepMind: Gemini Robotics 2.5 accelerates real-world physical intelligence
- SiliconANGLE: Google DeepMind releases new advanced robotics AI model and next-generation gene prediction model