DeepMind launches AI model AlphaMissense for predicting genetic diseases

About 10 years ago, Žiga Avsec was a physics doctoral student who quickly learned about genomics through a university course on machine learning. He soon joined a lab researching rare diseases and became involved in a project aimed at identifying the exact gene mutations that cause a mitochondrial disease.

Avsec described the problem as "finding a needle in a haystack." In the genetic code, there are millions of potential suspect gene mutations that could have a significant impact on human biology. Of particular interest are the so-called missense mutations, where a single letter change in the genetic code leads to the production of a different amino acid in a protein. Amino acids are the building blocks of proteins, which in turn are the building blocks of various parts of the body, so even small changes can have significant and far-reaching effects.

There are 71 million possible missense mutations in the human genome, and the average person carries over 9,000 of them. Most of these mutations are harmless, but some have been associated with genetic diseases like sickle cell anemia and cystic fibrosis, as well as more complex conditions like type 2 diabetes, which may be caused by combinations of multiple small genetic changes. Avsec started asking his colleagues, "How do we know which mutations are actually dangerous?" The answer was, "We largely don't know."

Out of the 4 million known human missense mutations, only 2% have been classified as pathogenic or benign after years of painstaking and expensive research. Studying the impact of individual missense mutations can take months.

Now, Avsec, who is currently a research scientist at Google's DeepMind, and his team have developed a tool that can accelerate this process. AlphaMissense is a machine learning model that can analyze missense mutations and predict their likelihood of causing disease with 90% accuracy, surpassing existing tools.

AlphaMissense is built on AlphaFold, DeepMind's groundbreaking model that can predict the structure of billions of proteins based on their amino acid composition, but it works differently. Instead of predicting protein structures, AlphaMissense is more like large language models such as OpenAI's ChatGPT.

It has been trained on the language of human (and primate) biology, so it knows what a normal amino acid sequence in a protein should look like. When faced with an incorrect sequence, it can detect it, much like a mismatched word in a sentence. "It's a language model, but trained on protein sequences," said Jun Cheng, one of the co-authors of the paper published in Science today, alongside Avsec. "If we substitute a word in an English sentence, someone familiar with English can immediately tell if the substitution changes the meaning of the sentence."

Pushmeet Kohli, DeepMind's Research Vice President, used a recipe book as an analogy. If AlphaFold is concerned with how ingredients stick together, AlphaMissense predicts what happens if you use completely wrong ingredients.

The model assigns a "pathogenicity score" between 0 and 1 to the 71 million possible missense mutations, with higher scores indicating a greater likelihood of a specific mutation causing or being associated with a disease, based on its understanding of the effects of closely related variations. DeepMind's researchers collaborated with Genomics England, a government agency that collects a growing amount of genetic data, to validate the model's predictions against real research results on known missense mutations. The paper claims that AlphaMissense achieves an accuracy of 90% and classifies 89% of the variations.

Researchers attempting to determine whether a specific missense mutation is associated with a particular disease can now look up its predicted pathogenicity score in a table. The hope is that, just as AlphaFold has propelled advancements in various fields from drug discovery to cancer treatment, AlphaMissense will help researchers in multiple domains accelerate their study of genetic variations, enabling faster disease diagnosis and the discovery of new treatments. "I hope these predictions can give us more insights into which variations are causing diseases and have other applications in genomics and other fields," said Avsec.

The researchers emphasize that the predictions should not be used in isolation but applied to guide real-world research. AlphaMissense can help researchers prioritize the slow process of matching gene mutations to diseases and quickly rule out unlikely candidates. It can also help improve our understanding of overlooked areas in the genetic code: the model includes an "essentiality" indicator for each gene, measuring its importance for human survival (although the function of about one-fifth of human genes remains unclear).

Ewan Birney, Deputy Director-General of the European Molecular Biology Laboratory and Co-Director of the European Bioinformatics Institute, believes that AlphaMissense has particular applications in helping doctors quickly diagnose suspected genetic diseases in children. "We've always known that missense mutations have to be responsible for some undiagnosed cases, and this is a better way of prioritizing," he said, citing the example of the RPE65 gene, which causes blindness and can be treated by injecting gene therapy into the retina. AlphaMissense can help doctors quickly rule out other potential gene mutations (which could number in the thousands) in a patient's DNA to ensure they receive the correct treatment.

In addition to unraveling the effects of individual letter mutations, AlphaMissense also demonstrates the potential of AI models in biology. As it was not specifically trained to address missense mutations but rather trained broadly on protein biology, this model and similar ones could have applications far beyond individual mutations, advancing our understanding of the entire genome and its expression—from recipes to the entire restaurant. "The basic model framework comes from AlphaFold," said Kohli. "A lot of the intuitions carry over to some extent from AlphaFold, and we've been able to show that it works for this related but different task."