- Author
- Greg Uyeno, freelance science journalist, in IEEE Spectrum
- Published
- 8 September 2026
- Kind
- Short news article with quotes from DeepMind staff and one outside researcher
- Original
- spectrum.ieee.org/alphagenome-atlas
- Subject
- AI put to use
What it is
Greg Uyeno, a freelance science journalist, reports for IEEE Spectrum on a DeepMind release. The Atlas stores predictions of how each single-letter genome change affects gene regulation. The article explains the release and gives expert reactions.
What it says
- Uyeno reports DeepMind released precomputed predictions for all 9 billion single-letter DNA changes.
- The Atlas needs no code and gives a single-number impact score for each variant.
- DeepMind's Avsec says the team needed an 80-fold speedup to build the Atlas.
- The article lists the speedup methods as distillation, GPU kernel optimization and removing redundant work.
- The article says the model sees 1 million base pairs, so distant enhancers are hard.
- Outside researcher Carl de Boer calls the model leading but slow. The score may be misread, de Boer says.
- The article says the dataset is around 1 petabyte and free for noncommercial research.
What evidence it gives
The article quotes DeepMind's Pushmeet Kohli and Žiga Avsec. It also quotes Carl de Boer of the University of British Columbia. It gives figures for the number of variants, the speedup and the data size. It cites no test of the predictions' accuracy.
Interests and limits
The article rests mainly on DeepMind's own announcement and staff. Only one source is outside DeepMind. It does not test the predictions or compare them with lab results. It is a short news piece, so technical detail is thin. The article's own caveat is that the impact score simplifies many aspects.