Introducing AlphaGenome 2
Announced on November 11, 2026, AlphaGenome 2 is an artificial intelligence model designed to tackle one of the most complex puzzles in biology: the human genome. It builds on the success of its predecessors, including the famed protein-structure predictor
AlphaFold. But while other tools have mapped vast landscapes of biological data, AlphaGenome 2 is built for precision. Its primary job is to predict how tiny changes, or variants, in our DNA sequence can alter the behavior of genes within our bodies.
From Broad Tissues to Specific Cells
The key innovation of AlphaGenome 2 is its ability to move from the general to the specific. While nearly every cell in your body contains the same DNA, different cells use different parts of that genetic blueprint. A genetic variant that affects a neuron in the brain might be harmless in a skin cell. Previous models looked at tissues as a whole, which could obscure these crucial differences. AlphaGenome 2, however, is a 'conditional' model; researchers can ask it to predict the effect of a DNA change within a very specific cell type — for instance, an immune cell in the gut or a T-cell activated by a vaccine. This is the difference between having a map of a country and having a detailed street view of a single neighbourhood.
The Search for Causation
This new level of detail is critical for moving beyond correlation to establish causation in disease research. For years, scientists have identified genetic variants that are more common in people with certain diseases, but proving that the variant causes the disease is a long, expensive process. AlphaGenome 2 can accelerate this by pinpointing exactly how a variant impacts gene activity in relevant cells. In one powerful example, the model was able to computationally replicate a 2024 laboratory discovery linking a specific gene, ETS2, to inflammatory bowel disease by showing how it became overactive in certain immune cells. By running these 'what-if' scenarios on a computer, researchers can focus their lab experiments on the most promising leads.
A Leap in Predictive Power
The model's power comes from both its sophisticated architecture and the massive amount of data it was trained on — over 107 million individual cell profiles from dozens of tissue types. This has resulted in a significant performance boost. According to DeepMind, AlphaGenome 2 correctly identifies the causal genetic variant behind a change in gene expression about 30% of the time, nearly doubling the 16% rate of the previous state-of-the-art model. It can also predict the direction of a gene's activity (whether it becomes more or less active) with 43% accuracy, up from 23%. These are major steps forward in a field where progress is often incremental.
A Tool for Research, Not a Doctor
Despite the breakthrough, Google is clear about the model's current status. AlphaGenome 2 is a tool for scientific research, not a clinical or diagnostic product. Its predictions must still be validated through traditional laboratory experiments before they can inform patient care. The increased precision also comes with a higher computational cost. For now, it is being made available to non_commercial researchers to help accelerate their own discovery processes, pointing them toward the most promising avenues for understanding the fundamental mechanics of disease and, eventually, developing new drugs and treatments.
