An experimental vaccination against COVID-19 was given the go-light by South Korean authorities in June, making it the first drug in history to be based on a human-created protein. Researchers developed the vaccine about a decade ago, via a laborious process of trial and error1 using spherical protein ‘nanoparticles.
Recently, a group headed by David Baker, a biochemist at the University of Washington (UW) in Seattle, claims in Science2,3 that it can create such molecules in seconds rather than months because to massive breakthroughs in artificial intelligence (AI). Efforts like these are part of a larger scientific shift, as life scientists increasingly use artificial intelligence technologies like DeepMind’s AlphaFold, which predicts the structures of proteins. As of July, DeepMind’s most recent version of AlphaFold correctly predicted the three-dimensional structures of all proteins. And in recent months, there has been a meteoric rise in the number of AI tools, some of which are based on AlphaFold, that can rapidly imagine whole new proteins. This used to be a time-consuming and sometimes fruitless endeavor.
Without much attention to what these molecules can accomplish, most efforts are concentrated on tools that can aid to design novel proteins, formed unlike anything in nature. In contrast, scientists and the increasing number of businesses using AI to protein design want to create proteins with practical applications like eliminating waste and curing illness.
For the last three decades, Baker’s lab has been developing novel proteins. His team began building Rosetta, a program that would automate this procedure, in the 1990s. Early on, scientists envisioned the three-dimensional form of a new protein, typically by piecing together fragments of known proteins, and the software then inferred the amino acid sequence that would produce this protein.
However, when synthesized in a lab, these “rough draft” proteins seldom folded into the intended shape and instead became stuck in a number of various conformations. Thus, it was necessary to take further measures to adjust the protein sequence so that it could fold into just the correct structure. This was a computationally intensive phase since it involves modeling every possible folding of each possible sequence.
