A single DNA-letter substitution can change an amino acid in a protein, with effects that range from negligible to disease-related. Google DeepMind’s AlphaMissense model has classified millions of these missense variants, offering researchers a way to sift through a vast set of possibilities.
A large catalog for a hard classification problem
Missense variants replace one letter in DNA, leading to a different amino acid in a protein. That change may alter how the protein functions. Some variants can contribute to diseases such as cystic fibrosis, sickle cell anemia, or cancer, while many have little or no effect.
According to DeepMind, the average person carries more than 9,000 missense variants. The challenge is identifying which of these many changes could be harmful. In a research report published in Science, DeepMind said AlphaMissense classified 89 percent of all 71 million potential missense variants as either likely pathogenic or likely benign. Human experts had validated 0.1 percent of those mutations so far, the report said.
The scale of the catalog could help researchers focus attention on variants that need closer examination. It gives them predictions across many more variants than experts have validated, while leaving room for further study and confirmation.
How AlphaMissense produces its predictions
AlphaMissense is based on a fine-tuned variant of DeepMind’s AlphaFold protein prediction model. AlphaFold predicts protein structure from an amino acid sequence. AlphaMissense uses a different approach to assess variants: it draws on databases of related protein sequences and on the structural context around each variant.
The model was trained on DNA from humans and related primate populations. It distinguishes common variants, which are therefore likely benign, from rare variants, which may be more likely to cause disease. It assigns a score between 0 and 1 to indicate the likelihood that a variant is pathogenic.
That score is not a direct prediction of how a mutation changes protein stability or structure. Instead, it provides a likelihood estimate based on the model’s inputs. Researchers can use it as one piece of information when deciding which variants merit attention.
What the results could mean for research
DeepMind reported that AlphaMissense achieved new state-of-the-art scores in benchmarks for predicting disease risk from genetic variants. It also said the model outperformed other computational methods on classifications from the public ClinVar archive and was the most accurate model for predicting laboratory test results.
Those benchmark results suggest the system may be useful for organizing research questions and supporting investigation into human genetics. DeepMind is working with Genomics England to explore how its predictions can support research into the genetics of rare diseases. The company says the work could contribute to faster disease diagnosis and, eventually, therapy development.
DeepMind is making AlphaMissense’s predictions available to the scientific community. Researchers can use the catalog alongside other tools to investigate how genetic variation relates to disease. The predictions do not, by themselves, establish that a particular variant causes illness.
From variant predictions to possible treatments
Classifying variants is one part of a broader effort to understand disease and develop treatments. More comprehensive predictions may help scientists identify promising areas for further research, while laboratory results and other evidence remain important to interpreting what a variant means.
Google and DeepMind have also created Isomorphic Labs for disease research and treatment, with one goal of speeding up drug development. AlphaMissense fits into this wider interest in applying AI tools to biological research, but the article describes its immediate contribution as a catalog of variant predictions rather than a treatment or diagnosis in itself.