← Back to feed

Google DeepMind released AlphaGenome Atlas on September 8, a 1-petabyte database containing precomputed molecular effect predictions for all 9 billion possible single-nucleotide variants in the human genome. The dataset is available free for academic research through a web portal, API, and as a skill in Google Antigravity.

The database is more than 30 times larger than the AlphaFold Database. When DeepMind expanded AlphaFold in 2022, it grew from 190,000 experimental protein structures to 200 million predictions, covering nearly all catalogued proteins. AlphaGenome Atlas attempts a comparable expansion for genetic variant interpretation.

What Atlas Contains

For each of the 9 billion variants, Atlas provides thousands of molecular effect predictions spanning hundreds of human and mouse cell types and tissues. The predictions cover gene regulation, splicing, chromatin accessibility, and related processes. Critically, Atlas covers both the 2% of the genome that codes for proteins and the non-coding 98%, which houses most trait-associated variants identified in genome-wide association studies.

Each variant also receives an AlphaGenome Variant Impact (AVI) score — a single number combining predictions from AlphaGenome and AlphaMissense, DeepMind’s earlier model for protein-altering variants. The AVI score is accompanied by feature attributions that identify which molecular processes drive it. A separate layer maps over 2,500 recurrent DNA sequence motifs — the functional “words” of the genome — and their locations across the reference sequence.

DeepMind reports best-in-class performance on variant pathogenicity and rare disease benchmarks, though the comparison set and harness are not fully specified in the release materials.

Research Use

External collaborators used AlphaGenome Atlas before the public release. DeepMind states they identified and experimentally verified key variants in unsolved rare disease cases and found rare variants associated with common traits. No specific diseases or studies are named in the release.

The use case is practical: a researcher investigating a candidate variant in a non-coding region can look up its AVI score and feature attributions directly, without running AlphaGenome from scratch. Testing each of the 9 billion possible variants experimentally is practically impossible; precomputation makes the full space searchable.

Access and Scale

• Dataset size: 1 petabyte (30x the AlphaFold Database at launch)
• Variants covered: 9 billion single-nucleotide changes (every possible single-letter mutation in the human reference genome)
• Access: Free academic portal at alphagenome.google/atlas, AlphaGenome API on GitHub, Google Antigravity skill
• AVI score: Unified per-variant impact score combining AlphaGenome + AlphaMissense
• Cell types/tissues: Hundreds, spanning human and mouse

Commercial access is available separately; academic use requires only portal registration.

Context

AlphaFold changed structural biology by making protein structure prediction routine. AlphaGenome Atlas targets variant interpretation, a separate and in some ways harder problem — 3D structure is determined by sequence, but the phenotypic effect of a variant depends on cell type, tissue context, developmental stage, and interaction with other variants. Atlas addresses the first layer of that complexity: precomputed molecular-level predictions for every possible input.

The closest prior resource at this scale is the gnomAD database, which catalogs observed human variation from sequencing studies. gnomAD describes what variants exist in human populations and at what frequency; Atlas describes what each variant does to molecular biology, including variants that have never been observed in any sequenced person.