Key Notes
- DeepMind released AlphaGenome Atlas, precomputed predictions for all 9 billion possible single-letter DNA variants across the human genome, built by running its AlphaGenome AI model genome-wide (a ~1-petabyte dataset, 30x AlphaFold Database's size).
- Each variant gets a unified "AVI score" (combining AlphaGenome and AlphaMissense predictions) plus a breakdown of which molecular processes, like splicing or gene expression, are affected.
- Free for non-commercial research via web portal, API and Google Antigravity, still not validated for clinical use, and independent researchers stress predictions still require lab follow-up, with one calling AlphaGenome itself "very slow and computationally intensive.".
Google DeepMind released AlphaGenome Atlas, an online database containing precomputed predictions for the molecular effects of all 9 billion possible single-letter changes to human DNA.
The reference human genome contains roughly 3 billion base pairs, or DNA “letters.” At each position, three possible substitutions can occur, producing the 9 billion total variants the Atlas now covers.
The Atlas is built on AlphaGenome, an AI model DeepMind first released last year and detailed in a January Nature paper. AlphaGenome can take a stretch of DNA up to one million letters long and predict thousands of molecular effects, including gene expression, RNA splicing and chromatin accessibility.
Previously, researchers had to select specific variants, write code and run the computationally demanding model themselves for each query. DeepMind has now precomputed that work across the entire genome, producing a roughly one-petabyte dataset, more than 30 times the size of its earlier AlphaFold protein-structure database.
Each variant in the Atlas is linked to an average of about 27,000 individual molecular predictions across hundreds of human and mouse cell and tissue types. A new unified metric, the AlphaGenome Variant Impact, or AVI, score, combines AlphaGenome’s predictions with those from AlphaMissense, DeepMind’s earlier model for protein-altering variants, giving researchers a single ranking number alongside a breakdown of which specific biological process, such as splicing or gene regulation, is predicted to be most disrupted.
DeepMind reports that in evaluations described in its January paper, the underlying AlphaGenome model matched or outperformed leading external models on 25 of 26 variant-effect prediction benchmarks. The company says the AVI score shows best-in-class performance on variant pathogenicity and rare-disease benchmarks specifically.
The Atlas is freely available for non-commercial research through a web portal, DeepMind’s existing API, and as a skill inside Google’s Antigravity tool, with commercial access planned through Google Cloud.
What It’s Actually Useful For
The practical value lies in filtering, not diagnosis. Testing every possible mutation experimentally in a lab is not feasible, so the Atlas lets researchers rapidly narrow millions of candidate variants down to the handful most likely to matter, which they can then verify with targeted lab experiments.
DeepMind’s genomics lead, Žiga Avsec, described the goal as accelerating work across fundamental biology, disease research and treatment development. In one early application, researcher Gareth Hawkes applied the Atlas to data from more than 54,000 UK Biobank participants and identified 22% more non-coding genetic associations than standard analysis methods, pinpointing 19 genetic regions linked to body mass index.
Most human trait-associated variants fall in the roughly 98% of the genome that doesn’t code directly for proteins, precisely the region where interpretation has historically been hardest, and where this kind of large-scale filtering tool has the most potential to help.
Real Limits Worth Naming
Independent researchers offered measured praise alongside genuine caveats. Sasha de Boer, a genomics researcher not affiliated with DeepMind who recently helped build a framework for comparing models like AlphaGenome, called it “the field’s leading model” but also “very slow and computationally intensive,” suggesting the precomputed Atlas mainly helps by giving people without high-end computing hardware access to results they couldn’t otherwise generate themselves.
Crucially, AlphaGenome is not validated for clinical use, and researchers involved with the release stress that its predictions require experimental, case-specific follow-up before being treated as established findings. That means the Atlas is best understood as a powerful hypothesis-generation and prioritization tool for scientists, not a diagnostic resource, and any specific variant prediction it produces still needs to be confirmed in the lab before it can inform actual medical understanding or treatment.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
