Google DeepMind has open-sourced AlphaGenome, an artificial intelligence model designed to help scientists study DNA and the biological processes that underlie disease. The model was built to accelerate DNA-focused medical research by showing how genetic instructions relate to health and illness.
AlphaGenome can analyze stretches of DNA up to one million base pairs long and predict a wide range of functional properties, giving researchers a detailed view of how the genome works. The tool is already being used in noncommercial research and is now available more broadly in open-source form.
New AI tool for DNA research
AlphaGenome is a machine-learning model that predicts what a given DNA sequence does, not just what letters it contains. Researchers at Google DeepMind describe it as a way to connect DNA to outcomes such as gene activity, protein production, and other molecular events in cells.
The model can forecast gene expression, DNA accessibility, histone modifications, transcription factor binding, and even aspects of how the genome folds in three dimensions, with high accuracy over sequences up to a million base pairs long. According to DeepMind, AlphaGenome also generates higher-resolution predictions of molecular properties than earlier approaches, which can provide more precise data for medical studies.
AlphaGenome’s architecture combines three major components, each based on a different kind of artificial neural network. A convolutional neural network first scans the DNA base pairs, transformer models refine those initial results, and a final set of artificial neurons turns the processed data into molecular property predictions scientists can analyze.
In a paper in the journal Nature, DeepMind reported that AlphaGenome outperformed competing models on 25 of the 26 internal benchmarks the team ran. The group also says the system can reach this performance using relatively modest hardware, since it can be run on a single H100 graphics processing unit.
From limited API to open-source release
The algorithm first appeared in June as an application programming interface that was restricted to noncommercial research use. DeepMind says that more than 3,000 scientists adopted the API, which now handles about one million requests every day.
AlphaGenome has been publicly available for noncommercial use since June 2025, when a preprint describing the model was posted, and the latest version is now hosted on GitHub. During this early-access period, the developers actively collected feedback from scientists about how well the system worked in practice.
According to Google DeepMind researchers, AlphaGenome has already been used to identify specific mutations in cancer genomes that drive the proliferation of tumors. They note that this is one example among several uses that have emerged as the tool has circulated in the research community.
Decoding the genome’s regulatory code
Much of the genome consists of regulatory regions that control when and where genes are turned on, and these regions can be difficult to interpret. Pushmeet Kohli, vice president of research at Google DeepMind, describes AlphaGenome as “our solution to deciphering the complex regulatory code” that makes up most of the genome.
DNA contains many instructions for producing proteins, but cells use only a small fraction of those instructions at any given time. DeepMind says AlphaGenome can help determine which instructions a cell is using in a particular situation, and how DNA changes might alter that pattern.
Because it links DNA sequences to features such as gene expression and DNA accessibility, AlphaGenome gives researchers a way to trace how changes in the genome may influence cellular behavior and, eventually, health. The model’s ability to handle very long sequences means it can analyze large genomic regions where multiple regulatory elements interact.
Building on AlphaFold and AlphaMissense
AlphaGenome follows earlier tools from the same lab, including AlphaMissense, which predicts the effects of mutations in protein-coding parts of the genome. The new model extends this idea to the wider genome by focusing on when and how DNA is active, rather than only on protein sequences.
DeepMind previously developed AlphaFold, which can automatically predict a protein’s three-dimensional structure, a task that once took months of manual work. One account reports that AlphaFold’s co-creators received half of the 2025 Nobel Prize in Chemistry for this achievement, while another report states that the creators of AlphaFold were awarded the Nobel Prize in Chemistry in 2024.
AlphaGenome is presented as the next step in applying artificial intelligence to core problems in biology and genomics. Scientists writing about the model note that only time will show how much influence it will ultimately have on genome research and related areas.
