LONDON, Jan 28, 2026, 17:31 GMT
- DeepMind released the AlphaGenome API code as open source but still requires an API key to access the model itself
- The company reports that nearly 3,000 scientists have utilized AlphaGenome, generating roughly 1 million API calls daily
- According to research published in Nature, the model analyzes 1 million DNA bases and predicts multiple gene-regulation signals.
On Wednesday, Google DeepMind released the AlphaGenome API code on GitHub, expanding access to its AI model that predicts how DNA mutations impact gene regulation. Researchers can now integrate the system directly into their software, though the core model continues to operate on DeepMind’s servers.
Sequencing projects churn out vast lists of genetic variants, yet researchers struggle to pinpoint which changes in the non-coding genome truly have an impact. This non-coding DNA functions as a regulatory layer, controlling when genes turn on and off.
AlphaGenome tackles the challenge by forecasting gene regulation—the switches and dials controlling where, when, and how intensely genes are expressed. DeepMind’s team unveiled their findings in a Nature paper released Wednesday.
Pushmeet Kohli, DeepMind’s VP of science, revealed that almost 3,000 scientists across 160 countries have tapped into AlphaGenome since it dropped seven months ago. Usage has climbed to roughly 1 million API calls daily, even though early access was restricted to non-commercial researchers using DeepMind’s own servers.
The model processes up to 1 million DNA “letters” simultaneously, predicting shifts across 11 biological functions, including gene expression and RNA splicing—the step that edits RNA before protein formation. It leverages 5,930 human data tracks alongside 1,128 from mice, delivering predictions at single-base resolution for most outputs, according to Science News. Sciencenews
“It’s a significant step up in overall usefulness,” said Anshul Kundaje, a computational biologist at Stanford who works on genomics AI. According to him, the extended context allows the model to detect mutations that affect genes located far from the actual changed base.
Judit García González, a human geneticist at Mount Sinai in New York, noted that researchers used to “might need to use three different tools” to predict various functional outcomes. Peter Koo, a computational biologist at Cold Spring Harbor Laboratory, added that the strongest signal usually emerges when models agree, since the “consensus tends to be more reliable.”
DeepMind is pushing deeper into biology with a new model that goes beyond AlphaFold, their protein-structure system, and AlphaMissense, a 2023 tool targeting mutations in protein-coding regions. Kohli described AlphaGenome as “our solution to deciphering the complex regulatory code.” Lead author Žiga Avsec added it can predict “gene expression, DNA accessibility, histone modifications, transcription factor binding, and even folding structure of the genome.” Acs
DeepMind made the software package open source, yet you still need an API key to access AlphaGenome, with rate limits in place. According to the product docs, the service is free for non-commercial use only, and AlphaGenome’s outputs aren’t allowed for training other machine-learning models.
Researchers warn these predictions still require lab validation, especially when applying population data to individual cases. Kundaje noted that unpublished results from his lab indicate AlphaGenome has trouble predicting gene activity changes across different people, which is why it’s not yet used as a clinical diagnostic tool.
Natasha Latysheva from DeepMind said AlphaGenome aims to accelerate research into the genome’s “functional elements,” the segments that regulate gene activity. Carl de Boer, a University of British Columbia researcher not involved in the project, cautioned that the field still needs models that reliably replace lab experiments: “Achieving this goal will require continued work from the scientific community.” Theguardian
