The way predictions were made before Omicron’s spike protein experiment reflects the recent sea change in molecular biology brought about by artificial intelligence.The first software capable of accurately predicting protein structures was widely available only a few months before Omicron, thanks to Competitive Research Team At Alphabet’s AI lab in the UK DeepMind at the University of Washington.
Ford used both packages, but since neither has been designed or validated to predict small changes caused by mutations like Omicron, his results are more suggestive than definitive. Some researchers are skeptical of them. But the fact that he can easily experiment with powerful protein-predicting AI illustrates how recent breakthroughs have changed the way biologists work and think.
Subramaniam said he received four or five emails from people who provided predicted Omicron spike structures as they worked on his lab’s results. “A lot of people do it just for fun,” he said. Subramaniam said direct measurements of protein structure will remain the ultimate metric, but he expects AI predictions to become an increasingly important center of research — including for future disease outbreaks. “It’s transformative,” he said.
Because a protein’s shape determines how it behaves, understanding its structure can aid in a variety of biological studies, from evolutionary studies to disease research. In drug research, figuring out protein structures can help reveal potential targets for new treatments.
Determining the structure of a protein is far from simple. They are complex molecules assembled from instructions encoded in an organism’s genome and serve as enzymes, antibodies and many other machines of life. Proteins are made from strings of molecules called amino acids that can fold into complex shapes that behave in different ways.
Deciphering the structures of proteins has traditionally required painstaking laboratory work. Most of the roughly 200,000 known structures were drawn using a tricky process in which proteins form crystals and bombard them with X-rays. New technologies like the electron microscope used by Subramaniam could be faster, but the process is still far from easy.
In late 2020, after decades of slow progress, the long-held hope that computers could predict protein structures from amino acid sequences suddenly became a reality. DeepMind’s software, called AlphaFold, proved so accurate in a protein prediction competition that the challenge’s co-founder, University of Maryland professor John Moult, declared the problem solved. “Having personally researched this issue for so long,” DeepMind’s achievement was “a very special moment,” Moult said.
This moment also frustrated some scientists: DeepMind did not immediately announce the details of how AlphaFold works. “You are in this strange situation and you have made significant progress in your field, but you can’t build on that,” said David Baker, who works on protein structure prediction in a laboratory at the University of Washington. I told Wired last yearHis research group used cues from DeepMind to guide the design of open-source software called RoseTTAFold, which was similar to AlphaFold but not as powerful as AlphaFold, which was released in June. Both are based on machine learning algorithms that predict protein structures by training on ensembles of over 100,000 known structures. Next month, DeepMind Announcement details And released AlphaFold for anyone to use. Suddenly, the world had two ways to predict protein structures.
Minkyung Baek, a postdoctoral researcher in Baker’s lab who led the RoseTTAFold study, said she was surprised how quickly protein structure prediction has become the standard in biological research. Google Scholar reports that UW and DeepMind papers on their software have been cited in more than 1,200 scholarly articles in the short time since they were published.
While predictions have yet to prove critical to Covid-19 research, she believes they will become increasingly important in tackling future diseases. The answer to eradicating a pandemic won’t be shaped entirely by algorithms, but the structure of predictions can help scientists strategize. “The predicted structure can help you focus your experimental work on the questions that matter most,” Baek said. She is now trying to get RoseTTAFold to accurately predict the structure of antibodies and invading proteins when combined, which will make the software more useful for infectious disease projects.
Despite their impressive performance, protein predictors don’t reveal everything about a molecule. They spit out a single static structure for the protein and didn’t capture the bends and wiggles that occur as it interacts with other molecules. The algorithms are trained on databases of known structures that are more reflective of those structures that are easiest to map experimentally than the full diversity of nature. Kresten Lindorff-Larsen, a professor at the University of Copenhagen, predicts that these algorithms will be used more frequently and will be useful, but says: “When these methods fail, we also need to learn better.”