Google’s Atlas of the human genome could pave the way for new treatments

DeepMind has published AlphaGenome Atlas, an AI-generated resource that provides predictions of how every single-letter change in the human genome could alter molecular processes, such as the quantity of a given protein. The company says the catalog spans roughly nine billion potential single-base substitutions and represents a genome-wide extension of earlier models that focused on protein-coding sequences.

The Atlas integrates predictions from DeepMind’s AlphaGenome model and is accompanied by a Variant Impact Score (AVI) that ranks variants by likely molecular effect, enabling researchers to prioritize which mutations merit experimental follow-up. DeepMind also offers multiple access paths to the resource, including a web portal, an interface in its AlphaGenome environment, and a skill for its Antigravity agentic platform.

AlphaGenome was trained on public human and mouse genomic databases, and converting model outputs into a precomputed, searchable catalog required substantial computation and analysis, DeepMind’s genomics lead Ziga Avsec said in a briefing. The finished dataset is substantial — DeepMind estimates it occupies about one petabyte of storage — and the company is making it available for noncommercial research use now, with commercial access planned on Google Cloud in the near future.

The Atlas follows DeepMind’s earlier genomics tools, including AlphaMissense and last year’s AlphaGenome model, and is part of a broader push by Google-owned DeepMind to apply AI to scientific problems. The company’s high-profile AlphaFold protein-structure work and ongoing efforts in drug discovery and other scientific domains provide context for the new release; DeepMind cofounder Demis Hassabis has recently shifted focus toward scientific research and related ventures such as Isomorphic Labs.