Google DeepMind Open-Sources AlphaGenome for DNA Research

Making frontier AI tools free for researchers could accelerate understanding of genetic disease and drug development.

Illustration: Google DeepMind Open-Sources AlphaGenome for DNA Research
AI-generated illustration · Sylvaris

From API to Open Code

On January 28, 2026, Google DeepMind released the source code and model weights for AlphaGenome, an AI system that predicts the function of DNA sequences. The model had been available through an API since June 2025, but the release now allows researchers to run it on their own infrastructure.

According to DeepMind, more than 3,000 scientists from 160 countries have used the API, processing about 1 million requests daily. The open-source release removes access barriers and makes it easier for labs with limited budgets to experiment with the model.

What AlphaGenome Predicts

AlphaGenome analyzes DNA sequences up to 1 million base pairs long and predicts properties like gene expression, DNA accessibility, and where proteins bind. About 98 percent of the human genome is non-coding DNA that regulates genes rather than producing proteins directly.

The model outperformed competing tools across 25 of 26 internal benchmarks, and it runs on a single H100 GPU. Researchers have used it to identify mutations in cancer genomes that drive tumor growth, study neurodegenerative disorders, and explore infectious diseases.

Building on AlphaFold

AlphaGenome follows DeepMind's AlphaFold, which predicts protein structures and earned its creators half of the 2024 Nobel Prize in Chemistry. It complements AlphaMissense, which focuses on the 2 percent of the genome that codes for proteins.

The timing matters. Labs need tools that help make sense of genome-wide association studies and pinpoint which DNA variants cause disease. Publishing the model in Nature alongside the code release signals that DeepMind sees this as research infrastructure, not a proprietary product.

sources
more in Science & Frontier Tech
Meta AI models power Genesis Mission scientific research projects at Lawrence Berkeley Lab Meta's Segment Anything and DINOv2 models are being deployed in federally-funded scientific research infrastructure, demonstrating how general-purpose AI models scale to specialized scientific domains. Claude AI produces claimed counterexample to Jacobian Conjecture AI models are now generating novel mathematical results that challenge long-standing open problems, marking a shift in how frontier mathematics research may be conducted. Probabilistic Computer Uses Noise for Computation Instead of Suppressing It A shift from error-fighting to noise-embracing architectures could unlock new approaches to optimization and sampling problems traditional computers struggle with.