DeepMind AlphaGenome 2 Maps DNA Variants Cell by Cell
Google DeepMind just released AlphaGenome 2, an AI model that predicts how single DNA mutations affect specific cell types in minutes.
Models · Source: Google DeepMind
What happened
Google DeepMind just launched AlphaGenome 2. It is a next-generation conditional AI model that predicts how DNA sequences and their variations impact specific human and mouse cell types. The model was trained on more than 100 million individual cell profiles across 68 tissues. This allows it to learn how the exact same stretch of DNA is read differently depending on the specific cell type.
The previous version processed bulk tissue averages in parallel. That approach spread the model capacity thin and missed subtle differences between closely related cells. AlphaGenome 2 changes the architecture entirely. It takes a million-letter DNA sequence and a target biological condition as input at every stage. It predicts core experimental readouts like gene expression, histone modifications, chromatin accessibility, and DNA methylation.
The real-world results are staggering. Researchers at the Francis Crick Institute recently spent a year and a half using sophisticated gene-editing to find a genetic variant driving inflammatory bowel disease. The variant overactivated the ETS2 gene in gut immune macrophages. AlphaGenome 2 replicated this exact finding computationally in just 30 minutes without lifting a single pipette.
Key facts
- 100 million — Number of individual cell profiles used to train AlphaGenome 2
- 68 — Number of tissues the training data spans
- 30% — Causal variants identified by AlphaGenome 2 on the TenK10K dataset, compared to 16% previously
- 30 minutes — Time it took the AI to replicate an inflammatory bowel disease discovery that took human researchers 1.5 years
- 98% — Percentage of the human genome comprised of non-coding DNA sequences that the model helps decipher
Why it matters
We are moving from blurry tissue averages to high-definition single-cell biology. Non-coding DNA sequences make up 98 percent of the human genome. They are notoriously hard to decipher. Until now, models could flag a problematic DNA mutation but could not tell you if it acted in a heart muscle cell, a blood immune cell, or a pancreatic beta-cell. AlphaGenome 2 solves this cell-type specificity problem. It adapts directly to new single-cell data without needing retraining.
This fundamentally changes the economics of genetic research and drug discovery. Finding causal variants used to require massive, expensive wet-lab experiments. Now, researchers can input single-cell gene activity data from a disease state and generate predictions instantly. The model tracks continuous changes, like when human blood stem cells mature into red blood cells. It turns static cell atlases into dynamic predictive engines for human genetics.
For builders
Free API for non-commercial research
DeepMind opened the AlphaGenome API today for non-commercial research. Academic builders and bioinformatics students can start testing variant effects immediately. You can input your own single-cell experiments and bypass months of manual lab work.
Commercial access is in private preview
Biotech founders and pharma companies must apply for private preview access. Those who secure early access will have a massive speed advantage. You can prioritize disease-relevant variants for drug pipelines faster than competitors relying on traditional wet labs.
Cross-modality imputation unlocks missing data
The model separates biological state from outputs. Builders can input just one measurement, like single-cell gene activity. The model will then predict unmeasured properties like chromatin accessibility. This helps teams prioritize exactly which expensive experiments to run next.
My take
Biology is rapidly becoming a pure data problem. When an AI can do 18 months of complex wet-lab work in 30 minutes, the bottleneck shifts from physical experiments to compute power. Biotech founders who do not integrate these predictive models will simply be priced out of discovery. The future belongs to those who simulate first and pipette second.
Original reporting: Google DeepMind. This is my rewrite and opinion.