The pre-compute just got biological—DeepMind ran the experiment 9 billion times so researchers don't have to.
The Summary
- DeepMind released Atlas, a catalog of AI predictions for all 9 billion possible single-letter DNA mutations in the human genome
- The underlying AlphaGenome model could already predict how DNA changes affect gene regulation, but required researchers to code and run it themselves—now the predictions are pre-computed and accessible
- Atlas could help unlock causes of rare genetic diseases by mapping how regulatory DNA—not just genes—affects cellular behavior
The Signal
Most DNA doesn't code for proteins. The majority is regulatory—switches and dimmers that control when and how much genes express. These elements interact in ways that vary by cell type and tissue. Some regulatory regions affect genes located far away in the genome. Understanding how small changes in this regulatory machinery influence gene activity is fundamental to understanding most disease, according to Carl de Boer, a genomicist at the University of British Columbia.
AlphaGenome, DeepMind's model for predicting these effects, was announced in 2025, with a Nature paper and public release in January. The model compares an original DNA sequence with an altered version and predicts how that single-letter change ripples through gene expression and regulatory activity. But using it required expertise: researchers had to select variants, write code, and run the computationally expensive model themselves.
"Understanding how changes in DNA affect gene regulation is fundamental to understanding most disease."
Now DeepMind has run the model for every possible single-point mutation in a reference human genome. That's 9 billion predictions. The Atlas release on September 8 means researchers can skip the compute and go straight to the data. Query a variant, get a prediction. No coding required.
This matters for rare disease research in particular. When you sequence a patient's genome and find a novel mutation in a non-coding region, you often can't tell if it's causing the disease or just background noise. Atlas gives you a prior: here's what the model predicts this change does to regulatory function. It's not a diagnosis, but it's a filter. It helps researchers narrow the search space from millions of variants to dozens worth investigating.
Key implications for genomics researchers:
- Pre-computed predictions eliminate the barrier of running computationally intensive models
- Rare disease cases with mutations in regulatory DNA now have a reference point for prioritizing variants
- The catalog is available for non-commercial use, opening access beyond well-funded labs
The Implication
The pattern is familiar: take an expensive AI inference task, run it once at scale, and distribute the results. DeepMind did this with AlphaFold and protein structures. Now it's doing it with DNA regulatory predictions. The compute moat becomes an access layer.
Watch for this to accelerate research into diseases caused by non-coding mutations—conditions where the gene itself is fine, but the regulation is broken. That's a big chunk of genetic disease we've historically struggled to diagnose. If Atlas helps researchers connect even a fraction of those dots, the pre-compute was worth it. For researchers working on rare diseases, the next move is obvious: start querying your variants of unknown significance against Atlas and see what surfaces.