Google DeepMind has unveiled AlphaGenome 2, a new model that predicts how DNA variants change gene activity inside individual cell types rather than across broad tissues. Announced on 11 November, the model finds nearly twice as many variants that alter gene expression as its predecessor, and it is available now for non-commercial research through the AlphaGenome API. DeepMind says the advance could help scientists pinpoint the cells where disease begins, though the model is for research use only and has not been approved for clinical or diagnostic work.
What AlphaGenome 2 does differently
The original AlphaGenome, launched last year, took a million-letter DNA sequence and predicted molecular activity across thousands of experiments from many tissues at once. That parallel approach spread the model’s capacity thin and blurred the differences between closely related cell types. AlphaGenome 2 rebuilds the design around a single condition at a time.
Alongside the DNA sequence, the model now takes a chosen biological condition — a tissue, a cell type or a state — as input throughout its architecture. That lets it learn how the same stretch of DNA is read differently in different cells.
The conditional design also lets the model fill gaps in the data. Given one readout, such as single-cell gene activity, it can predict unmeasured properties like chromatin accessibility, helping researchers decide which experiments are worth running next in the lab.
The results DeepMind is claiming
Trained on more than 107 million individual cell profiles across 68 tissues, AlphaGenome 2 explains far more of how gene expression differs between cell types in the same tissue. On TenK10K, a large single-cell map of 28 primary immune cell types, it identified 30% of causal variants against 16% for the original model.
It also got the direction of gene effects right 43% of the time, against 23% for the original model fine-tuned on single-cell data. That precision helps researchers find overlooked candidate variants for new discoveries, rather than re-testing the variants every other tool already flags.
The model reproduced a landmark finding computationally. A Nature study showed that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene in gut immune cells. AlphaGenome 2 singled out the same causal variant and showed it boosts ETS2 in gut macrophages, leaving other cell types largely unaffected.
AlphaGenome 2 predicts the same core readouts as its predecessor — gene expression, histone modifications and chromatin accessibility — and adds DNA methylation, a widely requested readout. It can generalise to new cell states without retraining: given data from a COVID-19 vaccine study, it distinguished vaccine-activated immune T cells from resting ones and flagged a variant linked to severe COVID-19.
Our Read
The step that matters here is resolution. Tissue averages hide the cell where a variant actually acts, and most disease-linked variants sit in regulatory switches far from the genes they control, which is why so many association studies end in a region rather than an answer.
A model that scores a variant in one cell type at a time turns a genome-wide association from a broad region into a specific cell, gene and direction — the three facts a drug programme needs before it can start.
The caveats are real and DeepMind states them plainly. The performance figures are the company’s own, from a technical paper that has not been peer reviewed, and even at its best the model recovers 30% of causal variants — a minority, not a census.
The model is also research-only, so no clinical claim follows from these numbers yet. But the ETS2 reproduction is the kind of independent check that matters: the model found, from sequence alone, a result that took a laboratory a year and a half to establish.
DeepMind’s drug-discovery spinout Isomorphic Labs is reportedly chasing $5 billion at a $50 billion valuation, and AlphaGenome 2 is the same science programme from the other end — better variant interpretation feeding better targets. DeepMind is quietly building the full stack from DNA prediction to drug discovery, and this release is the middle layer getting sharper.
Why It Matters
For genetics, the bottleneck has never been finding variants — genome-wide studies flag thousands of them. The bottleneck is knowing which variant, in which cell, does what. AlphaGenome 2’s precomputed scores for trait-associated variants across hundreds of cell conditions give researchers that mapping without large computations of their own, from height to type 2 diabetes.
AI agents have already run a DNA lab end to end, and AlphaGenome 2 points the same way: prediction first, then the experiment. If researchers can trust the model’s shortlist, they can spend their lab time on the variants that matter instead of screening thousands that do not. That is a real change in how genetic research gets done, not just a better score on a benchmark.
Background
AlphaGenome 2 follows the original AlphaGenome, launched last year, and the AlphaGenome Atlas, which holds predictions for all 9 billion possible single-letter changes in the human genome. The new model is available today for non-commercial research via the AlphaGenome API, with Colab tutorials, and in private preview for commercial research. DeepMind also plans AlphaGenome 2 skills for Google Antigravity.
DeepMind has spent the year proving its models outside the lab, including a formal verification of Signal’s encryption protocol. AlphaGenome 2 is the same pattern applied to biology: a research model released with an API, a paper and a community, judged on whether scientists use it to find something new.
Source: DeepMind’s announcement and The Verge’s report.
0 Comments