LONDON, Jan 28, 2026, 17:31 GMT
- DeepMind published AlphaGenome API code under an open-source license, while keeping model access behind an API key
- Company says nearly 3,000 scientists have used AlphaGenome, with about 1 million API calls a day
- Research published in Nature says the model can analyze 1 million DNA bases and predict multiple gene-regulation signals
Google DeepMind on Wednesday published code for AlphaGenome’s application programming interface, or API, on GitHub, widening access to its AI model for predicting how DNA changes affect gene regulation. The move lets researchers wire the system into their own software, while the underlying model still runs on DeepMind’s servers. Github
Sequencing projects are producing long lists of genetic variants, but researchers often cannot tell which changes in the non-coding genome actually matter. That non-coding DNA acts as a control layer, helping decide when genes switch on and off.
AlphaGenome targets that problem by predicting gene regulation — the switches and dials that shape where, when and how strongly genes are expressed. DeepMind’s researchers reported results in a Nature paper published on Wednesday.
Pushmeet Kohli, DeepMind’s vice president of science, said nearly 3,000 scientists from 160 countries have used AlphaGenome since its launch seven months ago. He said usage is running at about 1 million API calls a day, after early access was limited to non-commercial researchers connecting to DeepMind-run servers. Statnews
The model can analyze up to 1 million DNA “letters” at a time and forecast changes across 11 biological activities, from gene expression to RNA splicing, the editing step that shapes RNA before it becomes a protein. It draws on 5,930 human data tracks and 1,128 mouse tracks, and can make predictions down to single-base resolution for most outputs, Science News reported. Sciencenews
“It’s quite a leap forward in its overall utility,” said Anshul Kundaje, a Stanford computational biologist who develops genomics AI. The longer context, he said, helps catch mutations that influence genes far from the altered base.
Judit García González, a human geneticist at Mount Sinai in New York, said researchers previously “might need to use three different tools” to predict different functional consequences. Peter Koo, a computational biologist at Cold Spring Harbor Laboratory, said the best signal often comes from looking at agreement across models because the “consensus tends to be more reliable.”
The model extends DeepMind’s push into biology beyond AlphaFold, its protein-structure system, and AlphaMissense, a 2023 tool focused on mutations in protein-coding regions. Kohli called AlphaGenome “our solution to deciphering the complex regulatory code,” while lead author Žiga Avsec said it can predict “gene expression, DNA accessibility, histone modifications, transcription factor binding, and even folding structure of the genome.” Acs
DeepMind has released the software package under an open-source license, but access to AlphaGenome itself still goes through an API key and rate limits. Product documentation says the service is free for non-commercial use and that AlphaGenome outputs should not be used to train other machine-learning models. Alphagenomedocs
But researchers caution the predictions still need lab checks, particularly when moving from population data to an individual patient. Kundaje said unpublished tests in his lab suggest AlphaGenome can struggle to predict how gene activity shifts between people, one reason it is not a clinical diagnostic tool today.
Natasha Latysheva, a DeepMind researcher, said the company hopes AlphaGenome will speed work on the genome’s “functional elements” — the pieces that control gene behavior. Carl de Boer, a researcher at the University of British Columbia who was not involved, said the field is still working toward models that can stand in for experiments: “Achieving this goal will require continued work from the scientific community.” Theguardian
