Google DeepMind has introduced a groundbreaking AI tool called AlphaGenome Atlas, which researchers say could revolutionize our understanding of the human genome and potentially lead to new disease treatments. The platform provides a "predictive map" of every possible single-letter change in the human genome, offering insights into how these changes might affect biological processes. The human genome is composed of about 3 billion pairs of chemical "letters" — A, C, G, and T — that determine how genes function, when they are activated, and how proteins are produced. Small changes to these letters, known as mutations, can be harmless, contribute to individual differences, or lead to diseases. However, identifying which mutations are significant is a major challenge, given the vast number of possible changes — around 9 billion. AlphaGenome Atlas addresses this by predicting how each of these 9 billion potential mutations could impact the body at the molecular level. For example, it can estimate how much a specific protein might be produced if a mutation occurs. This makes it the most comprehensive catalog of how genetic changes affect biology, according to the researchers. Scientists can access these predictions through a web portal, via DeepMind's agentic development platform called Antigravity, or through the AlphaGenome interface. To help researchers prioritize the most relevant mutations, Google has also introduced the Variant Impact Score (AVI), a tool that ranks mutations based on their potential impact, using insights from other AI models that predict the effects of DNA changes. The project builds on previous AI models like AlphaGenome and AlphaMissence, which were developed to identify genetic drivers of disease and predict how small mutations might alter proteins. AlphaGenome Atlas expands on these efforts by making predictions across the entire genome, including regions that do not directly code for proteins but play a role in regulating gene activity. In a press briefing, Ziga Avsec, DeepMind’s genomics lead, explained that while the underlying AlphaGenome model was already available, creating a comprehensive catalog of all possible variants was a complex and time-consuming process. The model was trained using public databases of human and mouse genomes, allowing it to learn patterns between DNA changes and biological functions. The resulting dataset, which is about 1 petabyte in size, will be available to researchers for non-commercial use starting today, with commercial access to be offered on Google Cloud soon. AlphaGenome Atlas is the latest in Google’s efforts to use AI to solve fundamental challenges in science and medicine. This comes at a time when DeepMind co-founder Demis Hassabis is stepping back from managing the AI lab to focus on scientific research, including his work at Isomorphic Labs, a drug-discovery startup. The company’s most well-known AI achievement is AlphaFold, a model that predicted protein structures with remarkable accuracy and earned Hassabis and John Jumper the 2024 Nobel Prize in Chemistry. Google has also used AI to predict weather patterns, optimize computing solutions, and assist researchers through an AI "co-scientist" that helps with complex problem-solving.