Google DeepMind AlphaGenome Atlas Precomputes Predicted Molecular Effects for Nine Billion Human DNA Variants

Google DeepMind has released AlphaGenome Atlas, a large database of predicted molecular effects for possible single-letter changes in human DNA.

The company describes the September 8 release as a map covering about nine billion single-nucleotide variants.

Those variants represent one possible DNA letter change at each relevant position in the human genome.

ChatGPT is insanely powerful.

But most people waste 90% of its potential by using it like Google.

These 1,000+ proven ChatGPT prompts fix that and help you work 10X faster.

Sign up for Superhuman AI and get:

Claim your free prompts

The Atlas is built from precomputed outputs of the AlphaGenome model.

DeepMind says the resulting dataset is about one petabyte in size.

That is more than thirty times the size of the AlphaFold Database, according to the company.

The project changes the access pattern for a large genomics model.

Researchers do not need to generate every prediction during each analysis.

They can query a precomputed resource and rank variants before deciding which questions require laboratory work.

AlphaGenome was designed to process long DNA sequences while predicting many regulatory measurements at base-pair resolution.

The model can estimate signals related to gene expression, transcription, chromatin accessibility, transcription-factor binding, and RNA splicing.

The January Nature paper describes input sequences of up to one million DNA letters.

It reports leading benchmark performance on 22 of 24 genome-track tasks and 25 of 26 variant-effect tasks.

Those numbers describe benchmark evaluations, not proof that every prediction is correct in living cells.

The Atlas adds an AlphaGenome Variant Impact score.

The score combines predictions from AlphaGenome with protein-impact predictions from AlphaMissense.

Most developers babysit AI one prompt at a time. Top engineers don't, they build loops that plan, execute, and self-correct. The Code built The Ultimate Guide to Loop Engineering, giving you the exact system Silicon Valley engineers use. Sign up and get the guide free, plus a 5-minute daily newsletter.

Claim your Loop Engineering guide

DeepMind presents it as a way to rank variants across coding and non-coding regions.

The distinction matters because only about 2% of the genome directly codes for proteins.

Much of the remaining sequence participates in regulation, although the function of many regions remains uncertain.

The Atlas also links scores to feature attributions.

These attributions identify predicted contributors such as chromatin accessibility, splicing, conservation, and gene-expression effects.

It includes a catalogue of recurring DNA sequence motifs and their locations.

Together, these layers let a researcher move from a numerical ranking toward a possible molecular explanation.

The project has already been used in examples described by DeepMind.

The company says a Broad Institute team used the score to prioritize a variant in the DNM1 gene during rare-disease research.

It also says a University of Exeter researcher analyzed data from more than 54,000 UK Biobank participants.

DeepMind reports that grouping variants by predicted molecular effect revealed additional non-coding associations.

Those examples are company-reported uses, not independent clinical validation.

Nature’s reporting identifies the Atlas as a prediction system rather than a direct catalogue of experimentally observed effects.

That distinction should remain central when interpreting its scale.

Precomputed predictions can reduce the time needed to compare hypotheses.

They cannot establish that a variant causes a disease or that a predicted regulatory change occurs in a particular person.

The model also inherits limits from its training data, benchmark design, sequence context, cell types, and prediction targets.

The public release therefore expands the search space that scientists can inspect.

It does not remove the need for functional experiments, population analysis, replication, or specialist review.

DeepMind says the Atlas is accessible through a website and an API for academic research.

The company also describes later cloud access for commercial workflows.

The present research value is the combination of scale, query access, and interpretable score components.

The broader technical pattern is important beyond genomics.

A model can become a research instrument when its predictions are organized into a reusable, inspectable data layer.

That layer can make variant prioritization faster while keeping uncertainty visible.

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.