If you have an ACS member number, please enter it here so we can link this account to your membership. (optional)

ACS values your privacy. By submitting your information, you are gaining access to C&EN and subscribing to our weekly newsletter. We use the information you provide to make your reading experience better, and we will never sell your data to third party members.


Structural Biology

Accurate protein structure prediction AI made openly available

Structural biologists have free access to the newly developed tool

by Laura Howes
July 15, 2021 | A version of this story appeared in Volume 99, Issue 26


Ribbon structures of two proteins nestling together.
Credit: Ian Haydon/Institute for Protein Design
The RoseTTAFold software predicted hundreds of new protein structures, including this model of human interleukin-12 bound to its receptor.

Structural biologists can now solve protein structures that have been puzzles for years. This week, Minkyung Baek at the University of Washington and coworkers report results from their neural network, dubbed RoseTTAFold; it has been available to researchers for over a month as a tool to predict the three-dimensional folded structure of proteins from their sequences (Science 2021, DOI: 10.1126/science.abj8754).

Predicting the structure of a protein from its sequence alone was a seemingly impossible goal until recently. Then in November 2020, the firm DeepMind announced that their AlphaFold software program had solved the protein structure prediction problem. At the time, DeepMind did not release details of its algorithm beyond a 10 min presentation nor make it freely available for others to use.

David Baker, Baek’s postdoc advisor, has worked on the structure prediction problem for years. He remembers being excited. “DeepMind made really remarkable advances,” he says, “I found it really stimulating, and it gave us a whole bunch of ideas to think about.”

Using the AlphaFold presentation as a starting point, Baek drove the development of this new tool. The three-track neural network simultaneously considers the protein sequence, how amino acids interact at close range, and possible 3D conformations. Baek trained the network on hundreds of known protein structures; the result can solve new structures in minutes.

John Burke, a structural biologist at the University of Victoria and a coauthor on the new paper, was an early beneficiary. He says his team contacted the Baker lab a while ago while trying to solve a protein domain based on low-resolution cryo-electron microscopy data. His team had “thrown everything in our arsenal” at this portion of the protein and still couldn’t get a structure. When the data was fed into RoseTTAFold, it “snapped into place straight away.”

Having validated the software, the team uploaded a description of the work as a preprint in June, before peer review (bioRxiv 2021, DOI: 10.1101/2021.06.14.448402v1). The same day, Savvas Savvides at Ghent University began trying out the tool. “We jumped on it,” he says, adding that the software was able to solve a bottleneck he had been working to solve for the past 3 years.

While the AlphaFold algorithm solves some structures more accurately than RoseTTAFold, RoseTTAFold can also model assemblies of different proteins. More importantly to the researchers C&EN spoke with, RoseTTAFold is freely accessible, and the code can be analyzed, downloaded, and modified.

At the same time as the RoseTTAFold paper was released, DeepMind published an article outlining details of AlphaFold and released open-source code (Nature 2021, DOI: 10.1038/s41586-021-03819-2). In an emailed statement, DeepMind CEO Demis Hassabis says, “We pledged to share our methods and provide broad, free access to the scientific community. Today we take the first step towards delivering on that commitment.”


This story was updated on July 20, 2021, to correct the link to the full RoseTTAFold code.


This article has been sent to the following recipient:

Chemistry matters. Join us to get the news you need.