AlphaGenome Atlas Release: 9B Variants, Roughly 1 Petabyte

A model becomes a lookup

AlphaGenome Atlas does not introduce a new prediction model. It takes AlphaGenome, which compares a reference DNA sequence with an altered one, and precomputes its outputs across the human reference genome. Researchers can retrieve those predictions through a web portal or API instead of choosing variants, writing an inference pipeline, and running the model for each comparison. [1][2]

The nine billion figure comes from roughly three billion reference positions multiplied by the three other DNA letters that could replace the letter at each position. Nature also reports more than 100 million observed short insertions and deletions in the resource. This is therefore an exhaustive catalogue of single-letter substitutions, plus a large observed indel set, not a prediction for every possible rearrangement, repeated segment, or combination of variants in a person. [2]

Precomputation changes who can use the model more than what the model knows. The roughly one-petabyte Atlas removes repeated inference work and exposes the same predicted molecular effects as a searchable shared resource. [1][2]

One score, with the reasons behind it

Each variant can have thousands of predicted effects across gene expression, splicing, chromatin accessibility, protein binding, and other measurements in hundreds of human and mouse cell types and tissues. The new AlphaGenome Variant Impact score, or AVI, compresses signals from AlphaGenome and the protein-focused AlphaMissense model into one ranking value. [1][4]

The score is not the only output. Atlas also exposes feature attributions that break an AVI result into contributions such as splicing, gene expression, chromatin accessibility, conservation, and predicted protein impact, alongside more than 2,500 recurring DNA motifs and their locations. A researcher can rank a long candidate list, then inspect why the leading variants received their scores. [1]

A prioritization result, not a diagnosis

In the launch case study, Broad Institute researchers used AVI to elevate a non-coding variant near DNM1 in an unsolved severe-epilepsy case. AlphaGenome predicted that the change created an incorrect splice site and an abnormal protein extension. Experimental screens then supported that mechanism and found nearby variants with similar effects. [1][2]

That result demonstrates a useful sequence: rank candidates, inspect a predicted mechanism, and test it in the lab. It does not turn AVI into a clinical diagnosis, and it does not establish prospective accuracy across patients or diseases. [2][4]

Where the map stops

AlphaGenome reads one million DNA letters at a time, but the peer-reviewed model paper says effects from distant regulatory elements, tissue-specific and condition-specific behavior, non-coding genes, and personal genomes remain difficult or unevaluated. Complex traits also depend on development, environment, and gene-to-disease pathways beyond the molecular effects the model predicts. [4]

The Atlas makes a capable prediction system cheap to query; it does not make its predictions experimental measurements. Its best use at launch is as a filter for deciding which variants and mechanisms deserve scarce laboratory time. [2][4]

Access is free for non-commercial use through the portal and API, subject to AlphaGenome's terms. The documentation also says outputs should not be used to train other machine-learning models. Free access is therefore not the same as an unrestricted data license or general commercial availability. [5]