When AlphaFold solved the protein-folding problem in 2020, it showed that artificial intelligence could crack one of biology’s deepest mysteries: how a string of amino acids folds itself into a working molecular machine.
The team at Google DeepMind behind that Nobel Prize-winning platform then turned their lens from from the structure of proteins to how these molecules function in the body. Applying similar machine-learning methods, they first developed AlphaMissense, an AI tool for predicting which changes in protein structure are likely to cause disease. AlphaProteo, a system for designing proteins that bind to specific molecular targets, came next.
Now the architects of the Alpha platform are pushing beyond proteins into genomics, seeking to decipher how the vast regulatory regions of DNA shape when, where, and how genes are turned on and off.
Enter AlphaGenome. Described as a “Swiss Army knife for exploring non-coding DNA,” the deep-learning tool offers a way to systematically interpret the 98 percent of the genome that does not encode instructions for making proteins, but instead orchestrates how those genetic instructions are used inside the cell.
“This allows us to model intricate processes… with unprecedented precision,” Žiga Avsec, head of genomics at Google DeepMind, said in a press conference unveiling the new tool.
Narrowing the Genomic Search Space
AlphaGenome has its limitations. For instance, the tool’s training data draw largely from bulk tissue datasets, curbing its reliability in rare cell types or specific developmental stages, notes Christina Leslie, a computational biologist at Memorial Sloan Kettering Cancer Center. “Generalization to new cell types is a huge limitation,” she says.
It also struggles to capture distant effects when regulatory regions are hundreds of thousands to millions of DNA letters away from their target genes, Leslie pointed out.
Even so, the model is helping scientists to prioritize which genetic variants are most likely to matter, narrowing the search from across the genome to a manageable set of testable hypotheses. “It is the state of the art right now,” Leslie says.
According to DeepMind, thousands of scientists around the world are already using AlphaGenome, which is freely available on GitHub for academic research purposes. It is being put to work across a range of applications, including pinpointing genetic drivers of cancer and rare diseases, discovering new drug targets, and designing synthetic strands of DNA with tailored regulatory functions.
“It’s exciting to have things like AlphaGenome come out and perform much better than all the other dedicated algorithms that are exploring various aspects of genome biology,” says Richard Young, a biologist at the Whitehead Institute for Biomedical Research who has collaborated with Google DeepMind on its AI co-scientist platform but was not involved in AlphaGenome. “It’s a huge accelerator.”
High Resolution at Large Genomic Scale
The arrival of AlphaGenome marks another step in AI’s steady advance into some of biology’s most stubborn and consequential challenges.
For DeepMind, there is also a clear industrial logic at work. The company’s growing stable of biological models—spanning protein structure, mutation, and generation, and now genomic regulation—amounts to a vertically integrated platform for molecular prediction. That platform, in turn, should help unlock new diagnostic capabilities and therapeutic strategies, according to Pushmeet Kohli, vice president of science and strategic initiatives at Google DeepMind.
“All these different models are solving key problems that are relevant for understanding biology,” Kohli says.
AlphaGenome is the latest—and most expansive—piece of that strategy. Trained on raw DNA, the model predicts 11 types of biological signals that help determine how genes are used inside cells. These include whether a gene is turned on or off, where gene activity begins, how genetic messages are edited, how tightly DNA is packed, which regulatory proteins bind to it, and how distant regions of the genome interact with one another.
Many of these features already have their own specialty AI tools—SpliceAI for splice site prediction, ChromBPNet for local chromatin accessibility, Orca for three-dimensional genome architecture. But such tools are typically used in isolation, requiring researchers to stitch together results from multiple sources.
“AlphaGenome replaces this fragmentation with a more unified framework, which is more convenient and user-friendly—and we hope this will accelerate scientists’ workflows,” says Natasha Latysheva, a computational geneticist at Google DeepMind.
And while there have been attempts to capture all manner of regulatory effects in a single model, earlier architectures such as Borzoi and Enformer typically traded fine-scale resolution for breadth of biological coverage.
AlphaGenome tries to escape that trade-off. The model can ingest up to one million DNA letters at a time, preserving long-range regulatory context, while still making predictions at single-base-pair resolution. In practical terms, that means it can ask how a change in one nucleotide might reverberate across a vast swathe of the genome.
Connecting DNA Changes to Disease Biology
The new paper presents several demonstrations of this capability.
In one case, AlphaGenome correctly predicted how a tiny deletion disrupts a splice site in a gene involved in blood vessel biology, leading to reduced RNA output. In another, it captured how mutations near a cancer-linked gene boost its activity, helping to drive an aggressive form of leukemia.
Whether this predictive power generalizes beyond well-studied genes remains an open question, though.
“This is obviously a potentially valuable tool—but it’s a tool,” says Charles Mullighan, deputy director of the St. Jude Children’s Research Comprehensive Cancer Center. “It’s not a final point of discovery, but it’s going to be a very important tool for giving insights that then might guide further functional analyses and experiments.”
One “quirk” of the system, notes Latysheva, is its bias toward false negatives over false positives, meaning it is more likely to miss a genuinely important DNA variant rather than incorrectly flag a harmless one. “But the flip side of that is if it does predict a strong effect, it’s actually very accurate,” she says. So, when the model serves up a strong prediction, “you can have a decent amount of confidence that it knows what it’s doing.”
That confidence proved useful for Y-h. Taguchi and Kenta Kobayashi from Chuo University in Japan when they set out to stress-test a data-driven link between sleep deprivation and specific neuronal cell types. Early adopters of AlphaGenome, the bioinformatics researchers used the AI tool as an independent cross-check, confirming that genes implicated by sleep loss were especially active in their neurons of interest—just as their earlier analysis of gene-expression data from brain tissue had predicted.
“AlphaGenome succeeded in the cross validation,” says Takuchi, who published the results January 1 in the journal Genes.
That sort of validation underscores AlphaGenome’s role. Like AlphaFold before it, the system is not meant to explain biology in full, but to make its most opaque regions easier to explore.
