At the core of human diversity and evolution, genetic mutations play a crucial role. These mutations, which are changes to our DNA sequence, arise when cells replicate during cell division. While mutations can lead to genetic disorders, the presence of a mutation does not necessarily signify an impending disorder.
In a bid to streamline the understanding of these mutations and their implications, researchers at DeepMind, Google’s AI subsidiary, have developed a machine learning model named AlphaMissense. This advanced model is equipped to discern which DNA variations in our genomes might result in diseases.
With this initiative, DeepMind is also launching a comprehensive catalog featuring 71 million potential variants. These variants can influence how human proteins function, and in specific scenarios, they can trigger diseases like cystic fibrosis, cancer, and sickle-cell anemia, as highlighted in DeepMind’s press release.
The AI community and medical practitioners alike are buzzing with excitement about the potential of AlphaMissense. DeepMind anticipates that this AI tool will enhance the diagnostic processes for rare genetic conditions and also pave the way for the discovery of novel genes responsible for diseases. The overarching vision is to empower doctors to gain deeper insights into illnesses, thereby enabling the development of groundbreaking treatments.
“AI’s capability to precisely predict the consequences of variants can supercharge research across a plethora of fields, from clinical genetics to molecular biology,” remarked the DeepMind research team.
However, while the enthusiasm around DeepMind’s innovation is palpable, not all responses have been uniformly positive. Alex Zhavoronkov, the founder of Insilico Medicine, an AI-driven pharmaceutical company, opined, “DeepMind is stellar at PR and their AI work is commendable, yet the tangible commercial prospects of AlphaMissense remain to be seen.”
According to data from Google, the AI model successfully identified 89% of the 71 million ‘missense’ variants – a single DNA letter tweak that modifies the protein a gene produces – as probable disease culprits. “Our goal,” says physicist Stephen Hsu of Michigan State University, “is to assess a protein alteration and determine if it’s detrimental to the human carrying it. Currently, many of these changes remain a mystery in terms of their health implications.”
Historically, determining the potential impact of a variant has been both time-consuming and costly. The integration of AI tools can drive research at a brisker pace. The DeepMind team stated, “With AI-driven predictions, scientists can preview outcomes for thousands of proteins simultaneously, facilitating the prioritization of resources and fast-tracking intricate studies.”
AlphaMissense is not an entirely new invention but draws heavily from DeepMind’s earlier model, AlphaFold. AlphaFold was designed to predict 3D models of protein structures. In contrast, AlphaMissense was tailored using AlphaFold, refining it with data from human and primate variant population frequency databases. Despite their different functionalities, AlphaMissense employs databases of related protein sequences and the structural context of variants to yield a score, gauging the potential of a variant to be harmful.
DeepMind’s team is optimistic that the insights provided by their latest model can answer long-standing questions central to genomics and expand the horizons of biological science.
For those interested in delving deeper, the findings of Google DeepMind’s research can be found in the reputable journal, Science. The study sheds light on the pathogenicity of a vast array of missense variants in the human genome and offers a valuable resource—a comprehensive database of predictions for all conceivable human single amino acid alterations.
