Google DeepMind has introduced AlphaGenome, a new AI model designed to help researchers understand how DNA variants affect gene regulation.
The company says the model can analyze up to one million DNA letters at once and make predictions at the resolution of individual bases. That combination of long context and high resolution is central to the announcement, because many regulatory effects in the genome depend on interactions that span far beyond a single gene or mutation site.
AlphaGenome is aimed at what DeepMind describes as one of biology’s hardest problems: interpreting how changes in non-coding DNA affect molecular processes. The company says the model can predict thousands of properties linked to gene regulation, including where genes begin and end, how RNA is produced, how RNA is spliced and whether DNA is accessible or bound by proteins.
The model is also meant to score the effect of a variant by comparing mutated and unmutated DNA sequences. DeepMind says that makes it possible to evaluate how a change might alter regulatory behavior across multiple modalities in a single pass.
For researchers, the practical pitch is speed and breadth. Instead of using separate tools for different tasks, AlphaGenome is supposed to provide a more unified view of genome function. DeepMind says the model was trained on large public consortia datasets including ENCODE, GTEx, 4D Nucleome and FANTOM5, which gives it experimental grounding across a wide range of cell types and tissues.
The company also says AlphaGenome performs strongly on genomic prediction benchmarks, including tasks involving gene expression, chromatin contacts and splicing effects. In the announcement, DeepMind says it outperformed the best external models in many of the evaluated cases and was the only model able to jointly predict all of the assessed modalities.
One notable detail is that DeepMind is making AlphaGenome available in preview through an API for non-commercial research. That means the model is not being released as a general consumer product, but as a research tool meant to support scientific work while the company continues development.
The timing matters too. The announcement points out that AlphaGenome builds on prior work such as Enformer and complements AlphaMissense, which focuses on protein-coding variants. That leaves the much larger non-coding part of the genome as the key frontier the company wants to address.
DeepMind is also framing the model as a foundation for broader scientific discovery. If researchers can better estimate how a mutation changes regulatory behavior, they may be able to generate hypotheses faster and narrow down which variants matter most in disease biology.
The company notes that the research was later published in Nature, but the product announcement itself is about availability: AlphaGenome is now in preview via API. That is the immediate news. The longer-term question is whether researchers can use it to close some of the long-standing gaps between DNA sequence and biological function.
