The Institute for Protein Design helped usher in an era for scientists to predict the structure of a protein using artificial intelligence tools, an accomplishment that helped it win the “Breakthrough of the Year” award from Science magazine last year for its RoseTTAFold software.
The University of Washington group now shows how AI can be used to create new proteins more quickly than previously possible. The advances, published in three papers in Science, should accelerate the development of new proteins for biomedicine, materials and other uses.
“Software for predicting protein structures is part of the solution but it cannot come up with anything new on its own,” said IPD postdoctoral fellow Justas Dauparas in a press release.
The IPD has long been at the forefront of making new proteins: its protein design software has been licensed to 30,000 academic groups and is the basis for multiple commercial endeavors, including IPD spinouts like Cyrus Biotechnology and Icosavax. The new studies go beyond IPD’s workhorse software to leverage AI for protein building.
In a study released in July, the researchers showed how AI can be generate new protein shapes by first “hallucinating” a shape based on a simple prompt, in the way DALL-E or other AI tools produce outputs. The structure is refined through a process dubbed “inpainting,” which the group compares to an autocomplete function.
In a second study released Thursday, the researchers sped up the process, showcasing their new AI-powered software tool, called ProteinMPNN. It completes the design task in one second, more than 200 times faster than previous tools. In a third study on Thursday, the researchers assessed their designs using AlphaFold, an AI tool developed by Alphabet’s DeepMind similar to RoseTTAFold that also won the “Breakthrough of the Year” award. They found that the designed proteins were likely to fold into the intended shapes.
“We found that proteins made using ProteinMPNN were much more likely to fold up as intended, and we could create very complex protein assemblies using these methods” said IPD postdoctoral fellow Basile Wicky.
“ProteinMPNN is to protein design what AlphaFold was to protein structure prediction,” said David Baker, IPD head and senior author on the three studies.
