AlphaMissense: an AI to sort the mutations in our genes

Why do most of our mutations remain a mystery?

Each of us carries thousands of small differences in our DNA. Some change nothing. Others, rarer, can cause disease. Among them, "missense" mutations change a single building block (an amino acid) in a protein, a bit like one letter changed in a word.

The trouble is that the vast majority of these mutations seen in the human genome have an unknown clinical meaning. When a genetic test finds one, doctors often cannot tell whether it is harmless or the cause of a disease. Testing each one in the lab would take far too long.

How did the team build AlphaMissense?

The researchers started from AlphaFold, an AI known for its work on the three-dimensional structure of proteins. They fine-tuned it, meaning they adjusted it, for a new job: estimating whether a missense mutation is likely to be harmful.

The idea relies on databases showing how often each variant turns up in humans and in primates. A variant that is common in healthy people has a good chance of being harmless. The model combines this with the structural context of the protein and with what evolution has kept unchanged from one species to another.

One important point: the model was not explicitly trained on disease data or laboratory measurements. Those data were used afterwards to test it.

What did the AI manage to do?

Across a wide range of genetic and experimental benchmarks, AlphaMissense reached state-of-the-art results, meaning among the best known at the time of the study. The model classes 89% of human missense mutations as likely benign or likely pathogenic.

Another result: the average harm score given to the mutations of a gene also helps estimate whether that gene is essential for the cell to survive. The AI can even spot short essential genes that existing statistical methods struggle to detect because they lack enough data.

Finally, the team released a database covering every possible single amino acid substitution in human proteins, open to every laboratory.

What does this change for patients?

In time, a catalogue like this could help geneticists decide which mutations to examine first, especially in rare diseases where the search is like looking for a needle in a haystack. It could also guide the search for new drug targets.

For now, it is a research tool. Nothing in the study shows a measured benefit for patients.

What are the limits of this study?

AlphaMissense scores are estimates. A variant labelled "likely pathogenic" is not a diagnosis, and a "likely benign" one is not a guarantee. Results need to be confirmed by experts and, often, by experiments or family history.

The model only looks at mutations that replace a single amino acid. It does not cover other kinds of DNA changes. And clinical studies still need to show how these scores really help doctors day to day.