Google DeepMind has unveiled its novel AI tool that can predict whether DNA mutations are likely to cause genetic diseases. The ambition is that it will facilitate faster diagnosis and development of life-saving treatments.
The new AI model, called AlphaMissense, makes predictions about missense mutations, which occur when a single letter is substituted in the DNA code. The letters correspond to four nucleotides: adenine (A), cytosine (C), guanine (G), and thymine (T). This results in a different amino acid within the protein, which can in turn affect its function.
On average, humans carry over 9,000 missense variants, which, in most cases, are harmless. However, in some cases, they can cause serious diseases such as cystic fibrosis, cancer, and sickle-cell anemia.
DeepMind used AlphaMissense on all 71 million possible missense variants. The model categorised 89% of them as either benign or pathogenic. The company says that’s in contrast to the 0.1% that have been identified by human experts.
The AI tool is based on DeepMind’s AlphaFold model, which last year predicted structures for nearly all proteins known to science. To train AlphaMissense, the company fine-tuned AlphaFold on labels distinguishing variants seen in humans and closely related primates.
The team behind AlphaMissense highlights that the results demonstrate the tool’s higher performance compared to other computational methods alongside its unprecedented accuracy when making predictions from the lab.
“Together, AlphaMissense predictions have the potential to accelerate our understanding of the molecular effects of variants on protein function, contribute to the discovery of disease-causing genes, and increase the diagnostic yield of rare genetic diseases,” the team writes in its paper published in Science.
Notably, DeepMind is sharing the findings with the broader scientific community, which can now have access to the catalogue of missense mutations, the AlphaMissense code, and the Ensembl Variant Effect Predictor plugin.
The team notes that the predictions are to be interpreted alongside other evidence and not designed for direct use in clinics. Indeed, the AI tool’s value lies in its potential to speed up processes that would otherwise require laborious, expensive, and time-consuming research.