Google DeepMind, the AI research lab behind the Nobel Prize-winning AlphaFold protein structure prediction system, has developed a powerful new tool that tackles an even more complex challenge: deciphering the mysterious non-coding regions of DNA.
AlphaGenome, released in June 2025, represents a major breakthrough in genomic research by predicting how genetic variants affect gene regulation across the entire genome. While only 2% of human DNA directly codes for proteins, the remaining 98% – once dismissed as 'junk DNA' – plays a crucial role in controlling when and how genes are activated.
The model's architecture combines convolutional neural networks to detect short DNA patterns with transformer modules that capture long-range interactions between distant genomic elements. This hybrid approach enables AlphaGenome to process sequences up to one million base pairs long while maintaining single-letter resolution – a significant improvement over previous models that had to sacrifice either context length or precision.
"We have, for the first time, created a single model that unifies many different challenges that come with understanding the genome," said Pushmeet Kohli, vice president for research at DeepMind. The system was trained on public databases from consortia including ENCODE, GTEx, and FANTOM5, which experimentally measured regulatory properties across hundreds of human and mouse cell types.
AlphaGenome has already demonstrated its capabilities in real-world applications. In a June 2025 preprint study, researchers used the model to accurately simulate how specific mutations trigger gene overexpression in T-cell acute lymphoblastic leukemia, replicating known disease mechanisms without laboratory experiments.
The technology has significant implications for disease research, potentially helping scientists identify the genetic roots of disorders by tracing how mutations impact gene regulation. It could also accelerate synthetic biology by guiding the design of DNA with specific regulatory functions. DeepMind has made AlphaGenome available via API for non-commercial research, with plans to release the full model in the future.