Google's AlphaGenome Atlas predicts effects of every possible single-base change

Google announced AlphaGenome Atlas on Tuesday, a resource that attempts to predict the consequences of every possible single-base variant in the human genome. Because the human genome is roughly 3 billion bases long, testing the three DNA bases that do not appear in the reference genome at each position means sending a total of 9 billion bases through the AlphaGenome software.

AlphaGenome is built to identify potential functions of non-coding DNA, the portion of the genome that does not encode proteins but makes up the vast majority of it. Some of that non-coding DNA is essential: it tells the cell where and when to make messenger RNAs, how to process them into mature protein-coding forms, and similar regulatory jobs. But much of it appears to be little more than the remains of viruses and other molecular parasites.

The protein-coding share of the human genome is less than 3 percent. The rest is a mix of centromeres, which help ensure chromosomes are divided evenly between cells; caps that protect chromosome ends; the regulatory DNA that controls gene activity; the signals that determine what should and should not be included in mature messenger RNAs; and sequences that control how DNA is packaged inside the cell. Most of the non-coding DNA, however, is junk: the remains of ancient viral infections, DNA-level parasites, and genes that mutation has inactivated.

Distinguishing the useful non-coding DNA from the junk has been a long-running challenge for biologists. The proteins that interact with DNA are not especially picky about the sequences they stick to, and can bind at random throughout the genome while tolerating a degree of mutation. Many of those proteins are cell-type specific, with different populations of DNA-binding proteins in liver cells, nerve cells, immune cells, and others. In many cases, having many different protein binding sites packed into a compact space matters more than the presence of any single one of them.

Biologists have built various software tools that identify individual sites of interest in non-coding DNA, but the value of AlphaGenome Atlas is having all of the analysis done by a single software package across the entire genome. The article notes that most one-base changes to the human genome do nothing, while a few are significant, which is the pattern the tool is meant to sort out.

What AlphaGenome actually offers beyond what could be extracted from its training data will not be clear until biologists start using it heavily, assuming they do. The announcement positions the resource as a comprehensive reference, but its real-world utility depends on whether the research community adopts it and whether its predictions hold up against experimental evidence.

The open question is validation. A prediction tool is only as good as the checks that confirm its calls against what cells actually do, and that testing has not yet happened at scale. Whether AlphaGenome Atlas becomes a standard reference for interpreting non-coding variants, or joins the list of impressive models that researchers rarely consult, will depend on the results of that scrutiny.