Google DeepMind open-sources AlphaGenome medical research model – News

Google DeepMind today open-sourced AlphaGenome, an artificial intelligence model that researchers can use to study biological processes.

The Alphabet Inc. unit first debuted the algorithm in June. Until now, it was accessible only through an application programming interface limited to noncommercial research use cases. According to DeepMind, the API has been adopted by more than 3,000 scientists and processes about 1 million requests per day.

The Alphabet unit created AlphaGenome to accelerate DNA-focused medical research projects. According to the company, the model can help scientists better understand the role of DNA in biological processes and study diseases.

DNA contains instructions that cells use to produce proteins. Proteins, the basic building blocks of life, facilitate interactions between cells and power numerous other biological processes. Researchers can use AlphaGenome to understand how changes to protein production instructions impact health.

The model also lends itself to studying certain related phenomena. DNA contains numerous protein production instructions, but cells use only a tiny percentage of them on a day-to-day basis. AlphaGenome makes it easier to determine which instructions are used by a cell in a given scenario.

A DNA molecule is made up of segments called base pairs that are arranged in a double helix. Those base pairs are the instructions that cells use to make proteins. Each base pair comprises two nitrogen-based chemical compounds, or nucleobases, that are linked together by hydrogen atoms.

According to DeepMind, AlphaGenome can map out the molecular properties of DNA sequences with up to 1 million base pairs. That’s significantly more than the context window of earlier models. AlphaGenome also generates higher-resolution molecular property predictions, which provides researchers with more accurate medical data.

The model comprises three modules based on different AI architectures. Each one performs a different set of molecular calculations. 

The first module is a convolutional neural network, a type of AI that is mainly used for image and video analysis tasks. It’s responsible for carrying out the initial set of tasks involved in analyzing base pairs. From there, the processing results are refined by transformers. A third set of artificial neurons turns the data into molecular property predictions that scientists can use in their research.

In a Nature paper released today, DeepMind detailed that AlphaGenome outperformed competing models across 25 of the 26 evaluations it ran internally. Furthermore, the model can provide that performance using a relatively limited amount of hardware. Researchers can run AlphaGenome on only a single H100 graphics processing unit.

The model’s release comes five years after DeepMind introduced its seminal AlphaFold neural network. The latter algorithm can automatically predict a protein’s shape, a task that historically took months of manual work. AlphaFold’s co-creators won half of the 2025 Nobel Prize in chemistry. 

Image: DeepMind