Google DeepMind's AlphaGenome: AI model predicts the effects of DNA mutations

On June 25, 2025, Google DeepMind unveiled AlphaGenome, an AI model that takes a DNA sequence of up to 1 million base pairs and predicts thousands of molecular effects of single-nucleotide variants and mutations in regulatory regions of the genome. In January 2026, the research was published in the journal Nature, and the model itself became available via a separate link.

What the model predicts

AlphaGenome takes a long DNA sequence as input, up to 1 million base pairs, and outputs predictions for thousands of molecular properties that describe its regulatory activity. To score the effect of a specific variant, the model compares its predictions for the mutated and unmutated versions of the same sequence.

• Input: a DNA sequence of up to 1 000 000 base pairs
• Training data comes from the ENCODE, GTEx, 4D Nucleome, and FANTOM5 consortia
• The model covers hundreds of human and mouse cell and tissue types
• It predicts where genes start and end, splice sites, how much RNA is produced, DNA accessibility, and DNA binding to proteins
• The model's authors are Ziga Avsec and Natasha Latysheva of Google DeepMind

How AlphaGenome differs from AlphaMissense

AlphaGenome fills a gap that DeepMind's previous model, AlphaMissense, doesn't cover. AlphaMissense specializes in variants in protein-coding regions, which make up only 2% of the human genome. The remaining 98%, the non-coding regions, control gene activity and contain many disease-linked variants, yet these are exactly the regions where mutation effects have been hardest to predict accurately.

AlphaGenome also builds on Enformer, an earlier DeepMind genomics model, adding support for longer sequences and higher-resolution predictions. The architecture combines convolutional layers that detect short patterns in DNA, transformers that pass information between all positions in the sequence, and final layers that turn the detected patterns into predictions across different molecular modalities. During training, the computation for a single sequence is split across multiple interconnected Tensor Processing Units (TPUs).

How to access the model

Google DeepMind has opened access to AlphaGenome through its own API as a preview, for now only for non-commercial research, with plans to release the model more broadly in the future. According to the company's blog, the AlphaGenome work was published in the journal Nature in January 2026, after which the full paper and model access became available via separate links.

"We believe

AlphaGenome can be a valuable resource for the scientific community, helping scientists better understand genome function, disease biology, and ultimately drive new discoveries and the development of new treatments," Google DeepMind wrote in its blog.

What it means

AlphaGenome gives researchers a single tool to quickly assess how a specific mutation affects the genome's regulatory activity, replacing a collection of narrowly specialized models built for individual tasks. For disease genetics, that means faster initial screening of disease-linked variants before any expensive lab experiments.

FAQ

Who can access AlphaGenome right now?

As of the June 25, 2025 announcement, access is open through an API in preview mode for non-commercial research purposes only; Google DeepMind plans a broader release of the model later.

How does AlphaGenome differ from AlphaMissense?

AlphaMissense scores variants in protein-coding regions of the genome, which make up about 2% of all DNA. AlphaGenome scores variants in non-coding regulatory regions, which cover the remaining 98% of the genome and contain many disease-linked variants.