Google might not be at the frontier of LLMs any more, but it’s coming up with impressive AI research all the same.
Google DeepMind has released AlphaGenome Atlas, a massive new database that predicts what would happen if you changed any single letter, anywhere, in human DNA. The human genome has around 9 billion possible single-letter changes (called single-nucleotide variants), and Atlas contains a prediction for the biological effect of every single one of them. It’s being made available for free to academic researchers through a website, an API, and as a skill inside Google’s agentic platform, Antigravity.
To understand why this matters, it helps to think of DNA as an instruction manual for the body, written using a four-letter alphabet. Sometimes a single letter in that manual gets swapped out — a “typo” — and scientists have long struggled to figure out which typos are harmless and which ones cause disease. Testing each mutation in a lab, one at a time, is far too slow and expensive to ever cover the whole genome. Atlas essentially runs all 9 billion of those tests virtually, using AI, and hands researchers the results on a plate.
Built on AlphaGenome, at planetary scale
The predictions come from AlphaGenome, an AI model Google DeepMind built to predict how a genetic mutation affects processes inside a cell. AlphaGenome was already useful for scientists checking one variant at a time, but Atlas turns it into something closer to a searchable, genome-wide map — DeepMind describes it as being like an atlas of physical geography, one that layers information such as altitude and location on top of a landscape, except here the landscape is the genome and the “altitude” is a mutation’s biological impact.
The scale is hard to overstate: Atlas is a 1-petabyte dataset, more than 30 times the size of DeepMind’s AlphaFold Database, the protein-structure database Google DeepMind has been steadily expanding since 2022. For comparison, AlphaFold — the Nobel Prize-winning work of DeepMind’s former lead scientist John Jumper, working under DeepMind CEO Demis Hassabis — went from around 190,000 experimentally known protein structures to more than 200 million AI-predicted ones, and became one of the most widely used tools in modern biology in the process. DeepMind is clearly hoping Atlas follows a similar trajectory for genetics.
A single score to rank mutations
Alongside the raw predictions, DeepMind is also releasing something called the AlphaGenome Variant Impact score, or AVI score. Instead of forcing researchers to wade through thousands of data points per mutation, the AVI score condenses everything down into one number that ranks how impactful a given DNA change is likely to be. It combines AlphaGenome’s predictions with those of AlphaMissense, an earlier DeepMind model focused specifically on mutations that alter proteins.
Crucially, the AVI score works across both the roughly 2% of the genome that directly codes for proteins, and the remaining 98% — the non-coding regions that don’t build proteins themselves but act like the switches and dials controlling when and how genes turn on. Most disease-linked genetic variation actually sits in this non-coding “dark matter” of the genome, which is exactly the part that’s been hardest to interpret.
Atlas also breaks each AVI score down into its underlying causes — was a mutation flagged because it disrupts how DNA is packaged, how genes are spliced, or something else — and separately catalogues over 2,500 recurring short DNA sequences, or “motifs,” which function something like the recurring words and phrases of the genetic language.
Already finding real disease-linked mutations
DeepMind says external collaborators have already put Atlas to work. Researchers at the Broad Institute used the AVI score to sift through thousands of candidate mutations in unsolved rare-disease cases and flagged a variant in a gene called DNM1, which is linked to a severe form of childhood epilepsy. Atlas’s predictions showed exactly how the mutation caused harm — it created a faulty splice site that led to an abnormally long, malfunctioning protein — and lab experiments later confirmed the prediction.
Separately, a researcher at the University of Exeter ran Atlas against genetic data from more than 54,000 people in the UK Biobank and uncovered 22% more meaningful genetic links than standard methods found, including new variants tied to aging and to the body’s oxygen-sensing systems. The same approach was used to narrow down genetic regions potentially linked to body mass index.
Free for now, paid version coming to Cloud
DeepMind is positioning Atlas as a step toward a broader, AI-driven toolkit for biologists rather than a one-off release, with plans to plug it into Antigravity for end-to-end research workflows. It’s available for non-commercial academic use for free starting today, with a commercial version coming to Google Cloud later. This mirrors the earlier playbook DeepMind followed with AlphaFold, and continues a research tradition — running from AlphaGo through AlphaFold to now AlphaGenome — that DeepMind has built its scientific reputation on, and which Hassabis has repeatedly pointed to as the payoff of the lab’s long-term, fundamental-research approach.