On September 8, Google DeepMind released the AlphaGenome Atlas, a database that predicts the effect of every possible single-letter change in the human genome: all 9 billion of them. It is a 1-petabyte dataset behind a website that, in Google's words, requires zero coding skills.

For anyone who has stared at a variant of uncertain significance, that sounds like the answer. We checked what the human record of variants looks like next to it.

How much of the genome has anyone classified?

The human genome has about 3.1 billion positions, and each can change to three other letters. That is roughly 9 billion possible single-nucleotide variants. ClinVar, the NCBI archive where laboratories and clinicians submit variants and their interpretations, holds a very different number.

Source: Google DeepMind, AlphaGenome Atlas (September 8, 2026); NCBI ClinVar via E-utilities, queried October 4, 2026. ClinVar counts are records of type single nucleotide variant. The 9 billion figure comes from Google.

View data

set	variants
Possible single-letter changes (AlphaGenome Atlas)	9000000000
ClinVar single-nucleotide variant records	4243418

ClinVar has 4.24 million single-nucleotide variant records, out of 4.65 million records in total. Against 9 billion predictions, the observed and submitted record is about 2,100 times smaller. That is the gap an atlas could fill.

What is in that record?

The classification mix matters more than the total.

Source: NCBI ClinVar via E-utilities (db=clinvar), queried October 4, 2026. These five classes cover 93% of all records. The rest are conflicting or have no assertion.

View data

class	records
Uncertain significance (VUS)	2400109
Likely benign	1220370
Benign	281460
Pathogenic	257857
Likely pathogenic	178278

2.40 million records are variants of uncertain significance, 51.6% of everything in ClinVar. Pathogenic and likely pathogenic variants together are 436,135 records, 9.4%. Benign and likely benign add up to 1.50 million, 32.3%.

The uncertain group is what a clinician cannot act on. It is also the group where a good prediction would help most.

What does AlphaGenome add?

The model behind the atlas was published in Nature in January 2026. It takes 1 megabase of DNA sequence and predicts thousands of functional tracks, including gene expression, chromatin accessibility, transcription factor binding and splicing. The authors report state-of-the-art results on 22 of 24 genome-track tasks and 25 of 26 variant-effect tasks, and score a variant in about a second.

The atlas applies that model to every possible change and puts the results in a web portal. The size gives a sense of how much it stores: 1 petabyte for 9 billion variants is roughly 111 kilobytes per variant, so each variant carries a lot more than a single score.

What is still missing?

Three things, and the first is the one we care about most.

A statement of how well it works. The announcement post we read gives no accuracy figures and lists no limitations. The paper reports benchmarks, but the atlas post does not repeat them or say how they translate to the variants a clinic sees. We cannot tell readers how reliable any single prediction is.

A clinical standard. The American College of Medical Genetics and Genomics guidelines treat computational predictions as supporting evidence, not as a classification. Later ClinGen work showed that some predictors can count for more, but only after being calibrated against labelled variants. Without that calibration, a prediction for each of 9 billion variants does not shrink the uncertain pile by itself.

A way to know the model is wrong. Predictions are most useful where labels exist to check them, and ClinVar's labels are the ones we just counted. Regions and variant types where there are few labels are where the model is hardest to test.

Limits

We did not run AlphaGenome Atlas, and our view of it comes from Google's announcement and the model's papers. The ClinVar counts depend on the search terms we used: they count records, not unique genomic positions, and are a snapshot from October 4, 2026. The 9 billion figure is Google's, and our arithmetic of 3.1 billion positions times three alternate letters is consistent with it. We are not claiming the atlas is wrong. We are saying that it comes without the information needed to judge it.

Sources

1. Google DeepMind, AlphaGenome Atlas, September 8, 2026: all 9 billion single-nucleotide variants, 1-petabyte dataset, web portal.
2. NCBI, ClinVar, queried through the E-utilities API on October 4, 2026: 4,647,675 records; 4,243,418 single-nucleotide variants; 2,400,109 uncertain significance; 257,857 pathogenic; 178,278 likely pathogenic; 1,220,370 likely benign; 281,460 benign.
3. AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model (bioRxiv, 2025) and the Nature Structural and Molecular Biology commentary on the 2026 Nature publication.
4. Richards et al., Standards and guidelines for the interpretation of sequence variants, Genetics in Medicine, 2015; Pejaver et al., Calibration of computational tools for missense variant pathogenicity classification and ClinGen recommendations for PP3/BP4 criteria, American Journal of Human Genetics, 2022.
5. Related: How AI agents score on variant interpretation and what a correct variant call costs.