> ## Content Index
> Fetch the complete content index at: https://www.frontiermodels.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# AlphaGenome Atlas opens lookup for nine billion DNA changes
- URL: https://www.frontiermodels.com/news/alphagenome-atlas-opens-lookup-for-nine-billion-dna-changes/
- Published: 2026-09-11T01:39:46.000Z
- Updated: 2026-09-11T04:15:02.000Z
- Description: Researchers can retrieve precomputed molecular predictions through a free portal and API. Access varies by dataset, and impact rankings cannot establish a diagnosis.
- Author: Frontier Models Desk
- Tags: News, google, deepmind, alphagenome

Google DeepMind launched [AlphaGenome Atlas](https://alphagenome.google/atlas?ref=frontiermodels.com) on September 8, giving researchers precomputed predictions for roughly nine billion single-letter DNA substitutions. The practical change is that scientists can look up a variant’s predicted molecular effects instead of running AlphaGenome for each candidate, making large-scale screening more accessible through a free research portal and an API.

\~9 billionSingle-letter substitutions covered

1 petabyteDataset size reported by DeepMind

\~27,000Predictions per variant, on average

DeepMind describes a one-petabyte resource linking molecular predictions to a new AlphaGenome Variant Impact (AVI) ranking score and maps of recurring regulatory DNA sequences. Its accompanying preprint reports about 27,000 experiment-specific predictions per variant across hundreds of biological samples. About 15,000 remain when excluding “active allele” scores, which describe regulatory activity rather than the change caused by a substitution. These are predicted laboratory readouts, such as changes in gene activity or RNA processing.

What “every possible DNA letter change” covers 

At each scorable position in the GRCh38 human reference genome, Atlas considers the three other DNA letters: an A can become C, G or T. Positions marked N, meaning an unknown base, are excluded; the count does not cover structural changes, repeat expansions or combinations of variants in a person’s genome. AVI ranks predicted impact: a PHRED score of 20 means the top 1% of single-letter substitution scores, not a 99% probability of disease.

![The Revolocity DNA sequencer from BGI](https://storage.ghost.io/c/2c/b3/2cb3c4ca-c510-4af5-a585-007c769e0b39/content/images/2026/09/commons-64624405.jpg)

Revolocity DNA sequencer. Photo: Scotted400 / [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Revolocity%5FDNA%5Fsequencer.jpg?ref=frontiermodels.com) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/?ref=frontiermodels.com) · unmodified.

That reference-genome boundary matters for anyone building a lookup pipeline. Queries must use the alternative letter relative to GRCh38; reversing reference and alternative letters produces a lookup miss. If the reference is A, the database contains A-to-C and A-to-T predictions, but no precomputed C-to-T substitution at that position. The study separately scored more than 100 million observed insertions and deletions, known as indels. Its data-availability statement says the public genome-wide release currently covers single-nucleotide variants only.

The workflow benefit is credible without assuming every prediction is correct. [Scientific American’s reporting](https://www.scientificamerican.com/article/new-google-deepmind-alphagenome-atlas-could-transform-our-understanding-of-genetic-diseases/?ref=frontiermodels.com) quotes UC San Diego geneticist Jonathan Sebat on how precomputation could streamline laboratory work. That supports the value of replacing repeated computation with lookup, although the reporting packet contains no measured time savings across representative laboratories or a functional test of the portal.

AVI gives researchers a way to order those candidates. It combines AlphaGenome’s regulatory predictions with AlphaMissense’s protein-impact predictions, evolutionary conservation and other annotations, using 18 inputs. The training labels distinguish rare from more common population variants as proxies for impact. They are not experimentally established disease labels. Feature attributions show which inputs drove the score, allowing researchers to investigate, for example, a predicted splicing disruption rather than accepting the rank alone.

The authors’ validation gives that ranking a substantive basis. Across ten held-out genome-editing screens, AVI had the strongest correlation with measured effects in eight. On ClinVar variants within introns, the intervening portions of genes, AVI scored 0.76 against the next comparator’s 0.44 on AUPRC, a measure of how well a model retrieves pathogenic variants while limiting false positives.

The advantage was uneven. AVI trailed GPN-Star-M on the TraitGym Mendelian benchmark, 0.76 to 0.77\. These are evaluations run by Atlas’s authors and collaborators in a preprint; several authors have filed AlphaGenome-related patent applications. The results support testing AVI for prioritization, but provide no single error rate for the entire nine-billion-variant catalogue.

A high score also cannot establish what happens in a patient. [Nature’s coverage](https://www.nature.com/articles/d41586-026-02835-4?ref=frontiermodels.com) emphasizes the continuing need for experiments and individual-case information. DeepMind says AlphaGenome has neither been validated nor approved for clinical use. [ClinGen’s computational-evidence guidance](https://www.clinicalgenome.org/docs/calibration-of-computational-tools-for-missense-variant-pathogenicity-classification-and-clingen-recommendations-for-pp3-bp4-cri/?ref=frontiermodels.com) does not supply a shortcut: its cited calibrations concern established missense predictors, not AVI’s non-coding scores.

For researchers starting now, access depends on the data they need. The [portal](https://alphagenome.google/atlas?ref=frontiermodels.com) and [Python SDK and API](https://github.com/google-deepmind/alphagenome?ref=frontiermodels.com) provide non-commercial access. According to the paper, [static AVI downloads](https://alphagenome.google/downloads?ref=frontiermodels.com) permit commercial and non-commercial use; splicing scores and AVI feature attributions have non-commercial downloads and API access, while raw Atlas features are API-only and non-commercial. Motif bulk downloads and broader commercial Atlas access through Google Cloud are forthcoming, without a firm commercial date or price.

We would use Atlas to decide which variants deserve closer investigation. Researchers should check whether the relevant cell type and assay are represented, since coverage is uneven, and remember that a reference-based prediction does not capture every personal genetic context. Performance across diverse ancestries still needs evaluation. The useful next step after a promising score is a testable molecular explanation and corroborating evidence, not a diagnosis.