A 3-D image of the malaria protein Pfs48-45/ DeepMind. Researchers at the DeepMind Technologies artificial-intelligence lab said Thursday t...
 |
| A 3-D image of the malaria protein Pfs48-45/ DeepMind. |
The London-based lab, a subsidiary of Google parent Alphabet Inc., developed an algorithm called AlphaFold that predicts the three-dimensional structure of proteins, molecules that are found in all living organisms and play essential roles in the functioning of cells. The project was initiated in 2016.
Last July, DeepMind released an AlphaFold database with predicted structures for 350,000 proteins, including all of those in the human body, allowing researchers and labs around the world to make use of it for any purpose. It was expanded to 1 million proteins by December 2021.
On Thursday, DeepMind said it had expanded the database to include 214 million predicted proteins, or nearly all the proteins known to science. That includes proteins found in animals, plants, bacteria, and many other organisms. “When we launched the database last July, it was recognized as a pretty big leap forward for biology, and I think it also was a great demonstration of how AI can be used to advance scientific discovery. You can look up a 3-D structure of a protein almost as easily as doing a keyword Google search,” Dr. Hassabis said.
The effort to model proteins, underway for decades, has been accelerated greatly by AlphaFold, according to Ewan Birney, deputy director general of the European Molecular Biology Laboratory and director of EMBL’s European Bioinformatics Institute, which collaborated on the project. “This problem has been such a tough problem for so long,” Dr. Birney said.
During the past year, scientists at the University of Oxford used AlphaFold to advance their research into vaccines to stop the spread of malaria, which kills hundreds of thousands of people a year worldwide, according to Matthew Higgins, a professor of molecular parasitology. They used older methods to study a malaria protein called Pfs48/45. “We were never able to see in sufficient detail, despite many years of work, what this molecule looks like. We got a very fuzzy view of it,” Dr. Higgins said.
The postdoctoral researcher working on the problem was able to take the structure of the protein predicted by AlphaFold and compare it with the fuzzy view of the molecule derived from experimental methods. The two models fit together beautifully to create a sharp image of the molecule and how it works and how antibodies bind to it, according to Dr. Higgins. “So the use of AlphaFold was really really transformational, giving us a really sharp view of this malaria surface protein,” he said.
AlphaFold, a neural network system, was trained on known models of proteins and learned to predict proteins on its own. “This is an incredible milestone––both for science and for AI, exemplifying its role as a tool for scientific discovery and truly at a scale never attempted before,” said Shrikanth Narayanan, professor and Nikias chair in engineering at the University of Southern California.
Dr. Narayanan said sharing AlphaFold’s curated datasets has the potential to catalyze research and clinical translation worldwide. DeepMind is known for a number of pioneering AI models such as AlphaGo, which mastered the complex game of Go and in 2016 beat Lee Sedol, a top player.