AI in 2021 Research institute DeepMind has announced the development of AlphaFold, its first digital biological neural network. The model can accurately predict the 3D structure of proteins, which determines the functions these molecules perform. “We’re just floating bags of water,” says Pushmeet Kohli, vice president of research at DeepMind. “What makes us special are the building blocks of life: proteins. How they interact with each other is what creates the magic of life.”
AlphaFold was deemed a 2021 breakthrough by Science magazine. In 2022, it became the most cited research paper in the field of AI. “People have been working on[protein structure]for decades and haven’t made much progress,” Kohli said. “Then came AI.” DeepMind also released the AlphaFold protein structure database, which contains protein structures for nearly every organism whose genome has been sequenced, and made it freely available to scientists around the world.
More than 1.7 million researchers in over 190 countries use the system for a variety of research, from designing plastic-eating enzymes to developing more effective malaria vaccines. A quarter of Alphafold research has been devoted to understanding cancer, COVID-19, and neurodegenerative diseases such as Parkinson’s and Alzheimer’s. DeepMind released the next generation of Alphafold last year, which extends the structure prediction algorithm to biomolecules such as nucleic acids and ligands.
“It’s democratized scientific research,” Kohli says. “Scientists working on neglected tropical diseases in developing countries who couldn’t get funding to calculate protein structures can now access the AlphaFold database and get these predictions for free with the click of a button.” For example, the Drugs for Neglected Diseases Initiative, one of DeepMind’s early partners, used AlphaFold to develop drugs to treat relatively understudied diseases that affect millions of people (such as sleeping sickness, Chagas disease, and leishmaniasis).
DeepMind’s latest breakthrough is called Alpha Missense. The model classifies so-called missense mutations, which are genetic changes that cause a different amino acid to be produced at a specific position in a protein. Such mutations can change the function of the protein itself, and Alpha Missense gives the mutation a likelihood score of whether it is pathogenic or benign. “Understanding and predicting these effects is essential for discovering rare genetic diseases,” Kohli says. The algorithm, released last year, has classified about 89% of missense mutations in humans. Until now, researchers have classified only 0.1% of possible mutations clinically.
“This is just the beginning,” Kohli says. Ultimately, he believes AI will enable us to create virtual cells that will dramatically accelerate biomedical research and allow biology to be studied in computers rather than in real-world labs. “With AI and machine learning, we finally have the tools to understand this incredibly sophisticated system we call life.”
This article appears in the July/August 2024 issue. WIRED UK magazine.
