At 19:30 UTC on September 8, AlphaGenome Atlas had 82 points on Hacker News. Roughly nine hours later, the same submission showed 509 points, a 427-point climb that pushed it across this edition's breaking-news threshold. One question in the thread gets closer to the substance than the launch headlines: did Google DeepMind release a new biology model, or cache the outputs of one researchers could already query? The second description is closer. AlphaGenome Atlas precomputes predictions for 9 billion possible variants into a 1-petabyte catalog. It turns repeated model inference into a lookup, which could save researchers more time than another small gain on a benchmark.
Google DeepMind released the Atlas on September 8 as a free web portal for academic research, with programmatic access through the AlphaGenome API. Its entries cover every possible single-nucleotide substitution in the reference human genome. Each one carries predicted molecular effects, a summary AlphaGenome Variant Impact score, and feature attributions intended to explain which biological signals drove that score. These are model outputs at enormous scale. They are not observations of 9 billion mutations in people or cells.
The product is the precomputation
The arithmetic behind the headline is straightforward. The human genome has about 3 billion DNA positions, and each position can be changed to any of the other three nucleotide letters. That produces roughly 9 billion possible single-letter substitutions. DeepMind ran AlphaGenome across that space and stored the output; Nature reports that the Atlas also includes more than 100 million short insertions and deletions observed in human genomes.
AlphaGenome itself is not new. Its peer-reviewed Nature paper, published in January 2026, describes a sequence model that accepts up to 1 million DNA base pairs and predicts thousands of functional tracks, usually at single-base-pair resolution. The outputs span gene expression, splicing, chromatin accessibility, histone modifications, transcription-factor binding and three-dimensional chromatin contacts. In the paper, AlphaGenome matched or exceeded the strongest external models in 25 of 26 variant-effect evaluations.
Before the Atlas, researchers could send selected sequences and variants to the hosted model. Nature says about 9,000 researchers had used that API, which still required code and a decision about what to calculate. The portal removes the coding hurdle and most of the compute wait for common lookups because the predictions already exist. The difference resembles opening a search index instead of asking every visitor to crawl the web again.
The 1PB size is therefore meaningful as infrastructure, not as a count of new experiments. DeepMind says it is more than 30 times the size of the AlphaFold Database, whose 200 million entries contain predicted protein structures. AlphaGenome Atlas stores thousands of molecular predictions for each variant across many human and mouse tissues and cell types, plus more than 2,500 recurrent DNA sequence motifs. The comparison explains the storage bill; it does not make the two biological problems equivalent.
One score for triage
Thousands of tracks are useful when a researcher knows which mechanism to inspect. They are awkward when the first task is reducing a list of candidate variants. DeepMind's new AlphaGenome Variant Impact score, or AVI, combines AlphaGenome's regulatory predictions with AlphaMissense's estimates for protein-changing variants. A linked attribution breaks the score into signals such as altered splicing, gene expression, chromatin accessibility, conservation or predicted protein impact.
That summary can change the order of work in a rare-disease study. DeepMind says researchers at the Broad Institute used AVI to prioritize a previously overlooked variant in DNM1, a gene associated with epileptic encephalopathy. The underlying AlphaGenome tracks predicted that the change created an incorrect splice site and an abnormal protein extension. Experimental screens then validated the splice effect and found nearby variants with similar behavior. The model narrowed the search; a laboratory test supplied the evidence.
A second collaboration tested population-scale filtering. Gareth Hawkes at the University of Exeter applied the Atlas to whole-genome data from more than 54,000 UK Biobank participants. DeepMind reports that grouping rare variants by predicted molecular effect yielded 22% more non-coding associations with protein levels. In a separate body-mass-index analysis, restricting attention to the 1% of non-coding variants with the highest predicted impact identified 19 genetic regions for follow-up. Those results come from early collaborators and the accompanying research, not a clinical trial.
The 9 billion figure is easy to misread
An exhaustive list of possible substitutions is not an exhaustive model of human health. AlphaGenome predicts changes in molecular measurements derived from DNA sequence. It does not directly model a person's development, environment, medical history or interactions among distant variants. The model paper says complex traits sit outside that direct sequence-to-function scope because phenotypes depend on broader biological processes and gene-to-disease effects.
The paper names narrower technical gaps too. Effects from regulatory elements more than 100,000 bases away remain difficult. Tissue-specific predictions do not always transfer across cellular contexts, and condition-specific effects are still hard to predict. Training and evaluation focused heavily on protein-coding genes even though the Atlas is pitched partly as a guide to the non-coding 98% of the genome. Most plainly, the researchers had not benchmarked AlphaGenome for personal-genome prediction when the paper was published.
DeepMind puts a bright line around clinical use. The launch page says AlphaGenome has not been validated or approved for diagnosis or treatment. The API repository and terms summary likewise say predictions are for theoretical modeling and research and must not guide clinical decisions. A high AVI score is a reason to inspect a variant and design an experiment. It is not a medical result.
Free access comes with narrow permissions
Researchers can use the portal and API without charge for non-commercial work. The client library on GitHub carries an Apache 2.0 license, but that does not turn the Atlas into an unrestricted open dataset. The repository says Atlas outputs are generally limited to non-commercial use and may not be used to train another machine-learning model, except where the terms designate a downloadable artifact for broader use. DeepMind says commercial access will come through Google Cloud; the base AlphaGenome model is already available there.
The API documentation also draws a practical line between lookup and inference. Precomputed Atlas requests typically receive a higher query rate. Fresh AlphaGenome predictions suit analyses involving thousands of calls, while the hosted service is likely unsuitable for jobs requiring more than 1 million new predictions. For developers building research pipelines, the Atlas is the bulk path and the model API is the custom path. Neither is a license to mirror the entire petabyte.
The web interface broadens the audience beyond bioinformatics teams that can maintain an API client. A researcher can enter a variant, see its AVI score, examine the attributed molecular effects and decide which assay might test the proposed mechanism. Programmatic users can pull the same precomputed records into an existing prioritization pipeline. The release saves compute and setup time while leaving the hard scientific judgment in place.
Independent tests now matter more than scale
The evidence has two layers. AlphaGenome's underlying model has a peer-reviewed paper and extensive benchmarks. The Atlas, its AVI score and the first applications are described in a new companion preprint linked from DeepMind's announcement. Nature's launch coverage quotes bioinformatician Martin Kircher saying the resource can expand access to a strong model while still requiring experiments and case-specific evidence for diagnosis. That is a useful reading of the release: a faster way to choose questions, with no shortcut around validation.
Watch whether independent groups reproduce the DNM1 splice finding and the UK Biobank gains, then test AVI across other cohorts, tissues and conditions. Model versioning will matter as well: DeepMind says the map should improve with later AlphaGenome models, yet researchers will need stable records of which prediction produced a published result. The Atlas already answers the infrastructure question by making 9 billion model outputs searchable. Its scientific value will be measured one confirmed variant at a time.