AlphaGenome Atlas: what DeepMind's map of 9 billion DNA variants does

What AlphaGenome Atlas actually is

AlphaGenome Atlas is a database, not a chatbot. DeepMind took its AlphaGenome model and ran it exhaustively across the human genome, then stored every result. The output is what they call "the most comprehensive catalogue of how genetic mutations affect molecular biology," and it's big: about 1 petabyte of predictions, more than 30 times the size of the AlphaFold Database.

The scale is the point. For each variant, the Atlas holds thousands of molecular-effect predictions spanning hundreds of human and mouse cell types and tissues. DeepMind's Demis Hassabis framed the ambition directly:

"With AlphaFold we mapped the protein universe - now with AlphaGenome Atlas we're charting the human genome..."

Demis Hassabis, X

Two people to attach names to: the work is credited to Pushmeet Kohli, DeepMind's VP of Science, and Žiga Avsec, who leads its genomics work.

The problem it's built for: the 98% we couldn't read

Here's the context that makes this matter. Only about 2% of the human genome codes for proteins. The other ~98% is non-coding DNA, the part that decides when, where, and how strongly genes switch on and off.

That's a problem because most disease-linked variants that genome-wide association studies turn up sit in exactly that non-coding space, where the effect was historically a mystery. A variant like that gets a frustrating label: a "variant of uncertain significance." You can see it, you just can't say what it does. Before the Atlas, resolving one meant running the model yourself or doing slow, expensive lab work.

The pitch is that you can now rank variants by predicted impact instead of testing them all. Julia Zeitlinger of the Stowers Institute described the effect well:

"gives us a searchable dictionary for non-coding DNA"

Julia Zeitlinger, Nature

The AVI score: one number for any variant

The thing you actually read off the Atlas is the AlphaGenome Variant Impact (AVI) score, a single number describing how much a given change is predicted to matter. Its clever move is that it combines two models: AlphaGenome for regulatory (non-coding) effects and AlphaMissense for protein-altering (coding) effects. That means a coding variant and a non-coding variant come back on the same scale, so you can compare them directly.

The scale is logarithmic, which is easy to misread, so here's the plain version:

And it isn't just a black-box number. Each AVI score breaks down into additive contributions from interpretable categories like chromatin accessibility, splicing, and conservation, so you can see which molecular process a variant is predicted to disrupt, not only that it disrupts something. DeepMind reports the score gives "best-in-class performance across many variant pathogenicity and rare-disease benchmarks."

Under the hood: how AlphaGenome reads DNA

The Atlas is a lookup table, but the thing that filled it is the AlphaGenome model, first released in June 2025. It's worth understanding how it reads a genome, because that's what the scores rest on.

The model takes a 1 million base-pair stretch of DNA at a time and runs it through a U-Net-style architecture with a central Transformer tower, an encoder-then-decoder shape borrowed in part from AlphaFold. The long context is what lets it connect a regulatory switch to a gene sitting far away on the strand, and it produces predictions down to single base-pair resolution. From that one pass it predicts across 11 functional genomic modalities, including gene expression, RNA splicing, chromatin accessibility, and genome folding. Alongside the variant scores, the Atlas also catalogues more than 2,500 recurrent DNA motifs with their genomic locations.

Atlas vs running the model yourself

If AlphaGenome already existed, why build the Atlas? Speed and access. Running a model this size live, one variant at a time, is fine for a handful of lookups and hopeless for genome-scale work. Testing all 9 billion possible changes that way simply isn't feasible.

So DeepMind precomputed the answers. That single decision is what turns a research model into something a biologist can use on a Tuesday afternoon. Jonathan Sebat, a psychiatric geneticist at UC San Diego, put the day-to-day impact plainly:

"Our own workflows in the lab can be streamlined quite a bit because we don't actually have to compute anything. We literally can just look up everything."

Jonathan Sebat, Scientific American

That look-it-up shift has already shown up in real work. A team at the Broad Institute, working with the GREGoR Consortium, used AVI scores to prioritize a non-coding variant as a candidate cause of severe epilepsy in a patient. In a related case, the prediction even pointed at the mechanism: a variant that created an incorrect splice site.

How to actually use it

There are three ways in, depending on who you are.

For most researchers, the portal is the story: type a variant into a website and read the AVI score plus the per-modality breakdown, no coding skills required. If you want to script it, the API is a Python package (dna_client, with predict_variant and score_variant calls) that's free for non-commercial use. Access clearly matters to the field already: DeepMind says around 9,000 researchers had used AlphaGenome through the API even before the Atlas made it point-and-click.

Where it stops

This is the part a lot of the launch-day coverage skated over, and it's the most important section for anyone deciding how much to trust a score.

DeepMind is upfront that the Atlas is not approved for any clinical use and is not a substitute for professional diagnosis. It's a prioritization tool. And it's a less accurate one than its famous sibling: experts note the Atlas is far less accurate than the AlphaFold database, even though it's much bigger, because predicting gene regulation is a noisier problem than predicting protein structure. Martin Kircher of the Max Delbrück Centre gave a fair read:

"This is a useful and generous way to scale up access to a strong model" ... it "won't replace experiments or, in the case of diagnosing disease, accounting for differences specific to individuals."

Martin Kircher, Nature

The sharpest scrutiny came from practitioners on Hacker News, and it's worth hearing. One commenter with an obvious biology background pushed back on the idea that per-variant scores translate cleanly into medical value:

"on their own, with the exception of a few diseases, individual SNP predictions are meaningless in terms of actual pathogenicity."

dekhn, Hacker News

Another raised a subtle methodological catch: if the benchmark labels lean on evolutionary conservation, a model can look good partly by re-reading a signal it was already given. And a recurring theme was that single-variant maps miss biology that only shows up when several changes interact, something a one-variant-at-a-time score can't capture. None of this makes the Atlas useless; it makes it what DeepMind says it is, a starting point rather than an answer.

What this means if you build with AI

I write this from the software side, not the lab, and the part that stuck with me isn't the genomics. It's the product decision underneath it.

AlphaGenome existed for over a year as an API. Powerful, and mostly reachable by people who could stand up compute and write Python. The Atlas didn't make the model smarter. It did the hard work ahead of time and wrapped it in something you can just use, so a biologist looks up an answer instead of engineering their way to one. That gap, between "here's a capable model" and "here's a thing that does the job," is where most AI actually gets adopted or abandoned.

That's the exact pattern we care about at eesel. A raw model is infrastructure; what a team hires is an employee. eesel is an AI teammate platform, and the teammates arrive ready to work: today that's an AI helpdesk teammate that joins your support queue already knowing your help center, and an AI blog writer. You don't wire up a model and hope; the hard setup is done so the teammate is useful on day one, the same way the Atlas is useful the moment you land on it.

Try eesel

If AlphaGenome Atlas is a lesson in making a powerful model genuinely usable, that's the whole idea behind eesel for support and content teams. Rather than handing you a model and a blank config, eesel gives you a teammate that already knows your context, and lets you prove it out before it touches a customer by simulating on your real past tickets.

And because plenty of teams now run their tools from a terminal or a coding agent, there's an eesel CLI: the same teammate and workspace you'd manage in the dashboard, driven from the command line. A person can use it directly, a script can automate it, and coding agents like Claude Code, Codex, or Cursor can operate it: inspect a teammate's instructions, send a test message, read back structured results, and propose a bounded change for a human to approve. It's the no-wiring, look-it-up philosophy, applied to support automation. You can try eesel free.