Google unveils DNA analysis model AlphaGenome 2, which precisely predicts cell-specific variants
- Input
- 2026-11-12 07:23:46
- Updated
- 2026-11-12 07:23:46

[Financial News] DeepMind Technologies Limited, Google's artificial intelligence (AI) division, has unveiled AlphaGenome 2, a successor to AlphaGenome, its model for analyzing and predicting gene activity.
DeepMind Technologies Limited unveiled AlphaGenome 2 on the 11th (local time), a new tool that allows scientists to identify the cell-level causes of disease.
Unlike its predecessor, unveiled last year, which used a parallel approach to predict molecular activity across multiple tissues at once, AlphaGenome 2 adopts a “conditional model” structure that takes the specific cell type and state as inputs each time.
DeepMind Technologies Limited says this overcomes the limitations of its predecessor, which had been limited in its ability to detect subtle differences between individual cells.
The model can identify subtle effects that vary depending on which cell a given DNA variant occurs in, even when the variant is the same.
AlphaGenome 2... Autonomous analysis of new single-cell experimental data
The training data and variant-detection performance have also improved dramatically.
AlphaGenome 2 was trained on data from more than 107 million cells across 68 tissues, and its rate of detecting causal genetic variants nearly doubled, rising from 16% for the first generation to 30%.
It can also autonomously analyze new single-cell experimental data without separate retraining.
DeepMind Technologies Limited said that when single-cell data from COVID-19 messenger RNA (mRNA) vaccine research was entered, AlphaGenome 2 clearly distinguished immune T cells activated by vaccination from T cells in a resting state, despite never having encountered the data during training.
In an interview with The Verge, a U.S. technology publication, Ziga Avsec, DeepMind Technologies Limited's head of genomics, said, “When developing new drugs, you target specific proteins and specific cell types,” and noted that “this level of precision is essential.”
AlphaGenome 2 to be deployed intensively to identify cell-level causes of disease and targets for new drug development
However, because AlphaGenome 2 performs precise cell-level inference, its computational costs are higher than those of its predecessor.
Avsec predicted that the models would be used in a division of labor: the first-generation model to build large databases predicting all 9 billion variants across the genome, and AlphaGenome 2 to focus on identifying the causes of disease at the level of specific cells and finding targets for new drug development.
DeepMind Technologies Limited is making AlphaGenome 2 freely available for non-commercial research starting today, while offering it for commercial research as a private preview.
AlphaGenome is a follow-up project to AlphaFold, a protein-prediction program that won the 2024 Nobel Prize in Chemistry.
[email protected] Lee Seok-woo, International Affairs Correspondent Reporter