Dreaming up new proteins, AI churns out possible medicines and vaccines
Instead of predicting the shapes of naturally occurring molecules, software designs original ones
On July 21, scientists from institutions including the University of Washington published in the journal Science a new artificial intelligence (AI) software capable of mapping the structures of proteins that do not yet exist in nature. Even more exciting, scientists have used this software to create the original compounds that could potentially be used in industrial responses, cancer treatments, and even vaccine candidates to prevent respiratory syncytial virus (RSV) infection.
Computational biologist Jue Wang and his colleagues’ successful development of this artificial intelligence (AI) software is based on a series of recent advances in “using computers to predict the 3D structure of native proteins from amino acid base sequences.” Last year, an artificial intelligence (AI) program called AlphaFold developed by DeepMind predicted the structures of tens of thousands of monomeric human proteins. In addition, AlphaFold and an artificial intelligence (AI) software package called RoseTTAFold predicted the structures of thousands of protein-protein complexes. Based on these milestones, the protein structure prediction software was named “Science’s 2021 Breakthrough of the Year” by Science.
Predicting how native proteins fold is one thing; designing new proteins from scratch is another. In 2017, a team led by University of Washington protein designer David Baker showed that they could use Rosetta, a previously developed software program for protein structure prediction that did not involve AI, to design potential protein-based drugs that could bind to the flu virus. and molecular targets of bacterial toxins and their inactivation. Team members start by feeding the software the known fragment they want, a small piece of protein structure called a binding motif that can bind to a target. They then asked Rosetta to scan a database of protein structures they had previously designed and find an existing scaffold that might hold the active site in the correct shape. The software then combines the two parts and makes adjustments and improvements.
While this approach could aid drug discovery, it only works if Rosetta can find a suitable scaffold. “In the past, we could only passively expect to find a good match, but that’s not the case now,” says David Baker.
David Baker, Jue Wang, and their colleagues have adapted the AI-driven RoseTTAfold to design its own proteins from scratch using two different strategies.
The first strategy, called “inpainting,” begins similarly to the previous one, giving artificial intelligence (AI) a starting point, such as an active site or another key feature of a desired protein. Just as a word processor’s autocomplete feature completes a word after you type a few characters, artificial intelligence (AI) uses its understanding of how proteins fold to fill in the rest of the protein around the central feature.
The second strategy, called “constrained hallucination”, is more open than the first strategy. It gives the software a protein to target, such as binding to a metal. The program then generates a virtual protein consisting of a random sequence of amino acids and changes that sequence iteratively, evaluating the effect of each change on the protein’s likely shape and function. Artificial intelligence (AI) will keep the parts it thinks work and mutate others, steadily evolving toward the goal.
In both cases, the final predicted protein was able to be made and tested in the lab. Happily, both strategies worked. David Baker and his colleagues have created new proteins that bind to receptors on cancer cells, grab metals in solution, and bind carbon dioxide. Finally, to identify potential RSV vaccines, the team’s artificial intelligence (AI) generated 37 novel proteins designed to present the F protein site V, a key part of the virus, to the immune system. Three of the 37 bound to known RSV neutralizing antibodies, suggesting that they may be effective.
While the results weren’t always perfect, by imagining different proteins, the software came up with structures hitherto unseen in nature. “Researchers can use this as a starting point to evolve improved proteins in the lab,” said Joe Watson, who was involved in the study.
“This is a perfect use of artificial intelligence (AI),” said Yang Zhang, a protein designer at the University of Michigan. “Although researchers have used computers and other means to design new proteins for decades, AI like the one proposed in this new study has been around for decades. The (AI) approach is likely to increase the success rate.”
