Level 1 — Absolute Beginner

Google DeepMind is a company that builds AI. AI means artificial intelligence. This week, Google DeepMind made a new tool. The tool is called AlphaGenome Atlas.

Every person has DNA. DNA is like an instruction book inside our cells. The book tells our body how to grow and work. Sometimes one letter in the book changes. This is called a mutation.

There are 9 billion possible one letter changes in human DNA. The new tool uses AI to guess what each change might do. It gives each change a score. A high score means the change might be a big problem. A low score means it is probably safe.

The tool holds more than 1 petabyte of data. That is a huge amount, about 30 times bigger than an older Google tool called AlphaFold. Scientists all over the world can use AlphaGenome Atlas for free. It can help them learn more about diseases.

DNA

the material inside our cells that has instructions for how our body works

genome

all of the DNA instructions in a person's cells

mutation

a change in one letter of DNA

AI

artificial intelligence, a kind of computer program that can learn and predict things

database

a big collection of information stored so people can search it

scientist

a person who studies and learns about the world

disease

an illness that makes a person's body not work well

gene

a small part of DNA that tells the body how to make one thing, like eye color

Level 2 — Elementary

Google DeepMind, the AI research lab known for breakthroughs like AlphaFold, has launched a powerful new tool called the AlphaGenome Atlas. It is a searchable, AI powered database that anyone can use online.

The human genome is the complete set of DNA instructions found in every human cell. It is written using a code of chemical letters, and there are about 9 billion different single letter changes, called mutations, that could happen anywhere in that code. The AlphaGenome Atlas maps the predicted effect of every single one of these possible changes.

The database is enormous, holding more than 1 petabyte of information, roughly 30 times larger than DeepMind's earlier AlphaFold database, which predicted the 3D shapes of proteins. To build it, DeepMind combined its AlphaGenome model with another AI tool called AlphaMissense and several other features into one overall score. This score, called the AlphaGenome Variant Impact score, ranks each possible DNA change from low impact to high impact.

Before this tool existed, testing the effect of a single DNA change usually meant running lab experiments or slow computer models one variant at a time, a process that could have taken many human lifetimes to cover all 9 billion possibilities. The Atlas is free for researchers everywhere, including academic labs, universities, and public health scientists, and DeepMind hopes it will help scientists better understand genetic diseases by revealing which little studied DNA changes are likely to disrupt important genes.

genome

the complete set of DNA instructions found in a living thing's cells

mutation

a change in one letter of a DNA sequence

database

an organized collection of information that can be searched by computer

petabyte

a huge unit of digital information, equal to about one million gigabytes

protein

a molecule built from instructions in DNA that does jobs inside the body

impact score

a number that ranks how much effect something is predicted to have

researcher

a person who studies a subject carefully to learn new facts about it

disrupt

to interrupt or damage the normal way something works

Level 3 — Intermediate

Google DeepMind, the artificial intelligence lab that previously transformed structural biology with AlphaFold, has released the AlphaGenome Atlas, a searchable, AI powered database designed to map the predicted impact of every possible single letter change across the human genome. The genome, the complete set of roughly 3 billion DNA letters that make up the instructions for building and running human cells, can undergo mutations at any of those positions, and each position can change into one of three other letters, producing about 9 billion possible single letter variants in total.

To assess the likely effect of each variant, DeepMind combined its AlphaGenome model, which predicts how DNA sequences influence gene activity, with AlphaMissense, an earlier tool built to evaluate protein altering mutations, along with several additional features, folding them into a single measure called the AlphaGenome Variant Impact score. Each of the 9 billion possible changes receives a ranking on this scale, from low predicted impact to high predicted impact, giving researchers an instant estimate of how disruptive a given mutation might be.

The scale of the resulting database is striking. It holds more than 1 petabyte of data, about 30 times the size of AlphaFold's protein structure database, reflecting the sheer number of variants being scored across the entire genome rather than a smaller set of known proteins. Before the Atlas existed, answering the same question for even a single variant typically required a laboratory experiment or a computer model run one variant at a time, a workflow that would have taken many human lifetimes to work through all 9 billion possibilities.

Access is free for academic labs, universities, and public health scientists anywhere in the world, a decision that could widen who is able to do this kind of genetic research rather than limiting it to well funded institutions. DeepMind's stated hope is that the Atlas will accelerate the study of genetic disease by drawing attention to little studied DNA changes that are likely to disrupt the genes controlling particular health conditions, turning a search that once took a research career into one that takes seconds.

genome

the full set of DNA, roughly 3 billion letters long, that encodes the instructions for a human life

variant

one specific version of a DNA sequence that differs from the standard one at a given position

protein altering mutation

a DNA change that affects the protein a gene produces, potentially changing how it functions

disrupt

to interfere with or break the normal function of something, such as a gene

impact score

a single numerical ranking that summarizes how severe a predicted effect is likely to be

public health

the field concerned with protecting and improving the health of whole populations

Level 4 — Advanced

Google DeepMind has released the AlphaGenome Atlas, an AI powered, searchable database that assigns a predicted impact score to all 9 billion possible single letter substitutions across the human genome, the complete sequence of roughly 3 billion DNA letters that encodes the instructions for human life. The launch extends the company's earlier foray into structural biology, AlphaFold, into the more sprawling territory of genetic variation itself.

The Atlas works by fusing DeepMind's AlphaGenome model, which forecasts how a stretch of DNA shapes gene activity, with AlphaMissense, a tool built to gauge the severity of mutations that alter proteins, alongside several other predictive features. The result is distilled into a single figure, the AlphaGenome Variant Impact score, which orders every conceivable one letter change on a spectrum from largely inconsequential to potentially severe.

What distinguishes the project is scale. The database exceeds 1 petabyte, roughly 30 times the footprint of AlphaFold's catalogue of protein structures, a difference that reflects the sheer combinatorics of scoring every possible mutation across an entire genome rather than the finite set of known proteins. Historically, characterizing even one such variant required either a bench experiment or a computational model run in isolation, a variant by variant approach that, applied to all 9 billion possibilities, would have consumed the working lives of many researchers in succession.

DeepMind has made the tool freely available to academic laboratories, universities, and public health researchers worldwide, a choice that could democratize access to a resource otherwise within reach only of well resourced institutions. The company's ambition is that surfacing the predicted consequences of obscure, previously unexamined variants will sharpen understanding of the genetic underpinnings of disease, converting what was once a painstaking, case by case inquiry into something closer to an instant lookup.

genome

the entirety of an organism's DNA sequence, encompassing every gene and the regions between them

variant

a version of a DNA sequence that departs from a reference sequence at one or more positions

protein altering mutation

a genetic change with the potential to alter the structure or function of the protein it encodes

distill

to reduce something complex into a simpler, more concentrated form

combinatorics

the branch of mathematics concerned with counting and arranging possible combinations

democratize

to make something previously restricted to a few accessible to a much wider group

underpinnings

the fundamental basis or foundation of something

predictive model

a computational system trained to forecast an outcome from patterns in data