AlphaGenome

The human genome is a very long chain of four chemical building blocks, abbreviated A, C, G and T. This chain contains the blueprint for all the proteins that make up our body. Only a small part of the chain is blueprint, the far larger part controls when and where these instructions are actually used. It is precisely this controlling part that remains poorly understood to this day. AlphaGenome is a computer program by the company DeepMind that predicts, from any given segment of this chain, what effect it has in the body. It was introduced in June 2025 and is used free of charge by research groups.

The blind spot beyond the genes

About 98 percent of the human genome contains no direct blueprint for proteins. This region used to be dismissively called “junk DNA.” Today we know that switches and regulators sit there. They determine, for example, why a skin cell and a nerve cell behave completely differently even though both carry the same genome.

The problem: the vast majority of deviations found in sick individuals lie precisely in this region. A single swapped building block can render a switch unusable. Until now, no one could reliably say whether a detected deviation is harmless or not. Doctors therefore often faced a long list of anomalies without being able to classify them.

AlphaGenome provides a reasoned assessment for such sites. This does not replace a laboratory experiment, but it does sort the candidates. Researchers can then specifically examine the twenty most suspicious sites in the lab instead of twenty thousand. This saves years of work and a great deal of money.

From sequence of building blocks to prediction

AlphaGenome reads in up to one million building blocks at a stretch. This is important because switches are often located very far from the gene they control. Older programs could only look at short segments and missed such long-range effects. One can imagine this as the difference between a single sentence and an entire chapter: only in the chapter does the connection become apparent.

From this long segment, the model predicts several thousand measurement values simultaneously. This includes how strongly a gene is read out, where the genome is packaged accessibly, and which proteins dock there. It learned this from large public collections of real laboratory results on human and mouse tissues. So the model never read biology from a textbook, but derived patterns from measurement data.

The comparison mode is particularly useful. You enter the same segment twice, once in the healthy state and once with a deviation. AlphaGenome calculates both cases and shows the difference. This makes visible exactly what difference this one swapped building block makes.

From the lab to the headlines

In everyday life, you don’t encounter AlphaGenome directly. It is not a product to download, but an interface that research institutions access over the internet. Separate terms apply for commercial use. Typical users are university labs investigating rare genetic diseases, as well as groups researching the development of cancer.

In the news, the name usually appears in connection with DeepMind. The same company had previously caused a sensation with AlphaFold, a model that predicts the three-dimensional shape of proteins. This earned the 2024 Nobel Prize in Chemistry. AlphaGenome is often described as the next step, because it operates one level earlier, namely at the genome itself.

A common misconception is that AlphaGenome diagnoses diseases. It explicitly does not do this. It makes statistical predictions about biological measurement values, and these predictions can be wrong. DeepMind therefore rules out its use in patient care. As a tool for preselection in the lab, however, it is nonetheless a significant advance.