知Yizhi Daily
← Back to this edition

AI

DeepMind Releases AlphaGenome 2, Refining Predictions of Genetic Mutations’ Cellular Effects

Edition 2026-11-12 · 08:00 Beijing time

Google DeepMind has launched the genomic model AlphaGenome 2 to predict the differing effects of the same genetic variant across cell types and states. Trained on more than 100000000 cell profiles from 68 tissues, the model can help researchers connect population genetics findings with cellular mechanisms, disease biology, and drug targets.

According to The Verge, Google DeepMind has released AlphaGenome 2. It is the successor to AlphaGenome, introduced last year, and focuses on the problem that nearly every cell in the human body carries the same DNA, but different cells turn different genes on and off. Heart muscle cells, skin cells, and infection-fighting immune cells have different gene activity patterns, and the same variant can also have different consequences depending on cell type and state.

DeepMind says AlphaGenome 2 can provide more fine-grained predictions than the previous generation, identifying effects in specific cells that are linked to disease but may be overlooked by broader models. The training data covers 68 tissues and more than 100000000 single-cell profiles; the company describes it as turning a static cell atlas into a prediction engine for human genetics. The number of variants it identifies that can alter cellular gene expression is nearly twice that of previous advanced models.

The model can also make predictions for new cell types and biological states without retraining, and identify gaps in experimental data, providing leads for subsequent laboratory research. Adam Gayoso, a DeepMind research scientist, said the key breakthrough is connecting human genetics with cell biology; Žiga Avsec, head of genomics, noted that drug development targets specific proteins and cell types, making this level of granularity important for understanding and treating disease.

Original sources

The Verge ↗
Back to this editionOpen interactive reader