Every single-letter genetic change across human DNA receives a predictive map

Precomputing nine billion theoretical mutations in human genetic code allows researchers to scan regulatory mechanisms without running costly lab tests or writing custom code.

The human genome contains roughly three billion base pairs of genetic code, yet only about 2 percent of those instructions direct the creation of proteins.¹² The remaining 98 percent acts as a regulatory switchboard that dictates when genes activate, where they function, and how much cellular material they generate.¹³ A single altered letter in these regulatory stretches can disrupt human health just as severely as a flaw in a protein-coding gene, but identifying which mutations matter has remained deeply difficult.

In a living cell, genetic instructions follow an orderly chain of events before producing an effect. For a gene to produce a working protein, cellular machinery must bind to docking locations on the DNA strand, open the local chromatin structure, read the sequence into messenger RNA, and splice together the valid protein-coding segments while discarding intervening material. If an altered letter creates an accidental docking motif, the wrong regulatory proteins latch onto the strand. If the alteration ruins an existing boundary or introduces a false splice site, the cell splices the RNA message improperly, yielding a defective protein or halting production entirely.

To map those consequences across the entire genetic code, researchers at Google DeepMind released AlphaGenome Atlas on September 8, 2026.⁴ The project is an artificial intelligence database that precomputes the molecular consequences of all nine billion possible single-letter alterations in the human genome, corresponding to three theoretical changes at each of the roughly three billion positions.³⁵ Described in an 83-page preprint led by Jun Cheng and colleagues, the resource provides open browser access without requiring users to write code, alongside an automated programming interface for computational researchers.⁶⁷³

Why does non-coding DNA pose such a challenge?

Most genetic differences tied to human traits and common diseases reside in non-coding sequences that regulate cellular activity rather than inside protein blueprints.⁸⁶ Testing each theoretical change experimentally is impossible because human DNA contains roughly nine billion single-letter possibilities.³⁹ When researchers examine non-coding regions, standard laboratory tools struggle to determine whether a rare change is a harmless background variation or the trigger for an inherited illness.

AlphaGenome Atlas organizes these consequences through a summary metric called the AlphaGenome Variant Impact score.³¹⁰ The score combines non-coding predictions from the AlphaGenome sequence model, protein disruption scores from the earlier AlphaMissense tool, and evolutionary conservation data into a single ranking from 10 to 20 on a standard logarithmic scale.³ Each score is split into 18 distinct components that explain whether a given change affects RNA splicing, alters chromatin accessibility, or damages protein shape.⁶³

What does the precomputed atlas provide?

The new database catalogues the predicted functional consequences of roughly nine billion single-letter changes and more than 100 million observed short insertions and deletions across a one-petabyte repository.⁶¹⁰ That dataset is more than 30 times larger than Google DeepMind's AlphaFold protein database.³¹¹ By calculating these effects in advance on a computing cluster, the project eliminates the need for individual biology laboratories to run specialized machine-learning hardware.

Early trials by independent collaborators have begun testing the resource on real patient data. Researchers Laura E. Covill and Anne O'Donnell-Luria at the Broad Institute used the score to evaluate an unresolved case of epileptic encephalopathy, finding that a non-coding variation in the DNM1 gene created an incorrect splice site that led to an abnormally elongated protein.³⁶ At the University of Exeter, Gareth Hawkes applied the database to whole-genome sequencing from more than 54,000 UK Biobank participants, uncovering 22 percent more non-coding genetic associations than standard methods and identifying 19 genomic regions linked to body mass index.

Where does the predictive model face limits?

The database provides computational predictions rather than clinical diagnoses, and it cannot replace laboratory experiments or case-specific clinical reviews.¹⁰³ AlphaGenome Atlas is not approved for clinical use, and its scores serve as research hypotheses regarding molecular mechanics.³⁶ Like many sequence models, the underlying system is less dependable when interpreting regulatory interactions that occur across distances greater than 100,000 base pairs.¹²² Furthermore, while the model leads existing tools across most non-coding categories, competing tools such as GPN-Star-V achieve higher accuracy on variants in the initial untranslated segments of genes.⁶

Broader adoption will test how well these predictions generalize across diverse human ancestries, particularly because major cohorts such as the UK Biobank skew toward European populations. Academic investigators can access the platform without cost, while commercial developers must license the data.⁹ Google DeepMind genomics lead Žiga Avsec noted that precomputing the dataset removes computational friction so that researchers can quickly isolate plausible variants before designing focused laboratory experiments.¹⁰

This piece was prepared from the preprint and public records; the authors have not been interviewed.

References

This article is based on 15 sources, listed in the order they are cited.

1. 1 TG The Guardian third party · 28 Jan 2026 Google DeepMind launches AI tool to help identify genetic drivers of disease See the source
2. 2 BN BBC News third party · 28 Jan 2026 AI model from Google DeepMind reads recipe for life in our DNA See the source
3. 3 D deepmind.google third party · 8 Sep 2026 AlphaGenome Atlas: Molecular predictions for 9 Billion human DNA variants — Google DeepMind See the source
4. 4 C cryptobriefing.com third party · 8 Sep 2026 Google DeepMind unveils AlphaGenome Atlas mapping nearly 9 billion DNA mutations in the human genome See the source
5. 5 N nowosci.ai third party · 8 Sep 2026 Google DeepMind Releases AlphaGenome Atlas With Predictions for 9 Billion DNA Variants See the source
6. 6 A agentpedia.codes AlphaGenome Atlas: Complete Guide to DeepMind's DNA Map See the source
7. 7 H https://x.com/sundarpichai announcement · 8 Sep 2026 Google DeepMind Launches AlphaGenome Atlas to Map 9 Billion DNA Variants See the source
8. 8 N nature.com Advancing regulatory variant effect prediction with AlphaGenome - Nature See the source
9. 9 A aiweekly.co third party · 8 Sep 2026 Google DeepMind Ships 1PB AlphaGenome Atlas of 9B DNA Variants | AI Weekly See the source
10. 10 N nature.com DeepMind’s new genome ‘atlas’ charts effects of all 9 billion human gene mutations See the source
11. 11 X x.com Google DeepMind (@GoogleDeepMind) on X See the source
12. 12 D deepmind.google third party · 25 Jun 2025 AlphaGenome: AI for better understanding the genome — Google DeepMind See the source
13. 13 G Google third party · 8 Sep 2026 AlphaGenome Atlas: a high-resolution map of human DNA See the source
14. 14 SA Scientific American third party · 8 Sep 2026 New Google DeepMind atlas could transform our understanding of genetic diseases See the source
15. 15 D docs.cloud.google.com AlphaGenome | Gemini Enterprise Agent Platform | Google Cloud Documentation See the source