DeepMind's AlphaGenome 2 Narrows DNA Predictions From Whole Tissues to Single Cell Types
Google DeepMind announced AlphaGenome 2 on November 11, a DNA sequence model that predicts how genetic variants affect gene activity in specific human cell types and states, trained on more than 100 million single-cell profiles across 68 tissues. DeepMind says it identifies nearly twice as many expression-changing variants as its predecessor on one immune-cell dataset and can adapt to new single-cell data without retraining. It is available free for non-commercial research via the AlphaGenome API, in private preview for commercial research, and is not validated for clinical use.
DeepMind's genome model now answers a narrower and more useful question. The first AlphaGenome predicted how DNA variants affect molecular activity averaged across broad tissues; AlphaGenome 2 predicts it for a specific cell type or state, which is where drug developers and geneticists usually need the answer. It is also more expensive to run, which is why DeepMind says it is not replacing the original.
What DeepMind announced#
Google DeepMind released AlphaGenome 2 on November 11. According to DeepMind:
- It is trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues.
- On the TenK10K single-cell dataset of 28 immune cell types, it identified 30% of causal variants against 16% for the original model fine-tuned on single-cell data, and 43% of directional gene effects against 23%.
- It can be pointed at a scientist's own new single-cell data from a small sample of cells and make predictions for that cell state without retraining.
- It is available for non-commercial research via the AlphaGenome API, in private preview for commercial research, and DeepMind says it is not validated or approved for clinical or diagnostic use.
What the outside view adds#
The Verge reports that DeepMind's genomics lead Žiga Avsec described AlphaGenome 2 as more expensive to run than its predecessor, which made it impractical for the genome-wide atlas of every possible single-letter change that DeepMind released in September. Avsec told The Verge that cell-type specificity matters because drugs target specific proteins in specific cell types. DeepMind has already been working that protein-and-drug side of the problem, through the AlphaFold database of predicted protein complexes and through Isomorphic Labs, its drug-design spinout. DeepMind's post quotes Peter Donnelly, chief executive of Genomics Limited, saying his company plans to use the model in its platform.
Limits#
- The performance figures are DeepMind's own comparisons on datasets it selected; we found no independent benchmark at launch.
- Predictions are computational. They help prioritise which variants and cell types to study, but are not experimental confirmation.
- Commercial access is limited to a private preview with no published pricing.