Google DeepMind publishes AI-powered predictions for the effect of all 9 billion possible single-point mutations in the human genome

Google DeepMind said Tuesday that it has used artificial intelligence to predict the biological consequences of all 9 billion possible single-letter changes to human DNA, and is making the resulting database available free to academic researchers worldwide.

AlphaGenome Atlas, as DeepMind calls the database, is a precomputed catalogue of what each substitution of a single DNA base is likely to do to the machinery that switches genes on and off. Until now researchers had to run such a model one variant at a time or had to test variants in the laboratory, a process that was painstakingly slow. It would have taken many human lifetimes to discover the consequences of all 9 billion possible single-letter mutations.

The Atlas promises to make the job of biologists and medical researchers considerably easier, potentially speeding up the understanding of genetic diseases and the hunt for possible cures.

Pushmeet Kohli, DeepMind’s vice president for research and head of its AI for science team, told reporters on a briefing call that this was the first time any researcher in the world could reach a comprehensive map of human genetic variation “by simply opening a browser.”

Kohli also framed the release as helping to complete the unfinished business of the Human Genome Project, which in 2003 succeeded in mapping the entire human DNA sequence. “As the saying goes, we bought the book,” he said, “but we did not understand how to read it.”

Atlas is available for non-commercial use from today through a website Google DeepMind has set up for it. The company said it would be available for commercial use through a licensing arrangement through Google Cloud “soon.” Kohli said that Google DeepMind’s sister company, Isomorphic Labs, which is using AI for drug discovery, would have access to Atlas but that it would also require a commercial license for access. He did not specify exactly what the terms would be for commercial licensing. A paper describing the Atlas and how it was created is being released on bioRxiv, a repository for biomedical preprint academic papers.

Helping understand mutations in DNA’s vast ‘non-coding’ segments

DNA provides the recipe for the proteins a living cell can make. It consists of two chains of nucleotides, or molecules containing nitrogen, that coil around one another to produce a double helix. The nucleotides in DNA are formed of one of four different base components, cytosine (C), guanine (G), adenine (A), or thymine (T), as well as a sugar and a phosphate group. The bases form pairs, with guanine always binding with cytosine and adenine always binding with thymine. But sometimes, a single one of these letters will mutate, swapping an A for a G, for instance. This will ultimately switch the entire base pair when the DNA is copied. This base pair substitution can, in some cases, radically change the shape of the protein the DNA instructs a cell to produce. Those changes, in turn, can cause diseases.

DeepMind built Atlas by running AlphaGenome—an AI model DeepMind released last year that predicts the effects of single-letter genetic mutations—across a reference sample of the human genome, and then comparing each reference base against each of the three possible alternatives.

According to DeepMind’s research paper on Atlas, each variant is linked to an average of about 27,000 individual predictions about how the mutation will affect everything from gene expression to how it will alter the way in which the DNA sequence is transcribed into specific instructions for protein manufacture. It makes these predictions across hundreds of cell types and tissue types from both humans and mice. The team also scored more than 100 million insertions and deletions observed in population databases, including the U.K. Biobank and All of Us, a large database run by the U.S. National Institutes of Health that collects genetic, medical, and lifestyle data from Americans.

To make all of these predictions more usable, DeepMind is also releasing a summary metric, which it calls the AlphaGenome Variant Impact (AVI) score. That score folds AlphaGenome’s predictions about the effect of mutations on gene regulation together with predictions from AlphaMissense— an earlier model that DeepMind built that looks specifically at protein-altering mutations. An AVI score of 10 puts a variant among the 10% most impactful in the genome, while an AVI score of 30 places it among the strongest one in a thousand. Each score is broken down into the processes driving it, showing whether a variant is flagged for splicing, gene expression or protein change.

That breakdown matters because the protein-coding part of DNA accounts for about 2% of the genome; the other 98% governs when and where genes are switched on. Mutations to this “non-coding” portion of DNA have been far harder for scientists to interpret so far. “AlphaMissense looks at proteins,” said Žiga Avsec, DeepMind’s genomics lead. “With AlphaGenome, we are focusing on the regulatory part of the genome.”

Atlas also includes a catalogue of more than 2,500 recurring short DNA sequences, or motifs—the segments transcription factors bind to—mapped across the genome.

Early testers report promising results

DeepMind gave several scientists access to Atlas to beta test it prior to today’s release. Laura Covill and Anne O’Donnell-Luria of the Broad Institute worked with the GREGoR Consortium, which works on unexplained rare genetic disorders, to use the AVI score to re-examine several unsolved cases. In a patient with epileptic encephalopathy, Atlas pointed the scientists to a variant in DNM1, a gene important to synaptic function in brain cells. Sixty-nine percent of its score came from splicing: the model predicted the variant creates a spurious splice site in a version of the gene found only in the brain, lengthening the resulting protein by 13 amino acids. Because that segment is barely expressed in blood, earlier RNA sequencing of blood samples had been inconclusive. Laboratory experiments confirmed the prediction, and the variant was reclassified as likely pathogenic.

In a retrospective test on previously solved GREGoR cases, the paper reports, AVI placed the known causal variant among a patient’s top 50 candidates 29.5% of the time, against 12.5% for CADD, an existing ranking method.

Gareth Hawkes, a Medical Research Council fellow at the University of Exeter, applied Atlas to whole-genome data from more than 54,000 U.K. Biobank participants, hunting for rare non-coding variants affecting levels of proteins circulating in the blood. Filtering candidates by their predicted molecular effect yielded 22% more associations than the same analysis run without Atlas, and in one case narrowed a region from 526 candidates to four. “The human genome is a massive search space,” Hawkes said in a statement supplied by DeepMind. “We can use it to shrink the haystack.”

Julia Zeitlinger, an investigator at the Stowers Institute for Medical Research, used the motif maps to sort transcription factors by what they do in different cell types—separating, for instance, repressors that leave DNA accessible but still block a gene from switching on. Mapping thousands of such sites without Atlas “would not have been possible,” she said, because doing it experimentally is laborious. Four decades of that work, she added, has validated only a tiny share of the motifs the model predicts.

Ewan Birney, director of EMBL’s European Bioinformatics Institute, said his organization is working to integrate the AVI score into Ensembl’s Variant Effect Predictor, a widely used annotation tool that the EMBL hosts. “These tools reach their full value when they’re open and plugged into the wider data ecosystem,” he said in a statement.

Atlas predictions don’t replace the need for lab experiments

DeepMind acknowledged that the Atlas predictions are not a substitute for experimental evidence. Avsec said AlphaGenome works well for some classes of variant, such as those affecting splicing or promoters, but can miss others, particularly in enhancers. He said the Atlas predictions are not, overall, as accurate as what the DeepMind AI model AlphaFold was able to achieve for protein structure prediction. The predictions are “accurate enough to really point us in the right direction with downstream studies,” he said, but researchers should not treat them as “the universal truth.”

DeepMind’s paper on Atlas describes it as a research tool that can form only part of the evidence chain behind a clinical diagnosis, and notes gaps in its training data and a limited ability to capture effects that act indirectly, through changes in the levels of regulatory proteins.

This story was originally featured on Fortune.com