The lab of David Baker at the University of Washington’s Institute for Protein Design has released its application to solve one of the toughest problems in the life sciences: how to quickly and accurately predict the folding of a protein computationally.
The findings, building on work performed by the Google-owned company DeepMind last fall, were published today in the journal Science, on the same day DeepMind released its approach in the journal Nature.
While DNA provides the instructions, proteins are the building blocks of the body. The applications of both teams should provide an accelerant for research of all stripes across the life sciences, from basic science to drug development.
With 20 amino acid building blocks, the options for how an individual protein might fold are numerous and depend on multiple molecular interactions within the protein and its environment. These interactions are extremely difficult to predict and are constantly shifting during the folding process.
Historically, predicting the folding of even a small protein has taken immense computing power — one group even built a massive supercomputer just for the purpose — with mainly incremental results. Drug companies and researchers have relied on laborious experimental methods to determine the structure of proteins, such as critical drug targets.
Last fall DeepMind stunned the field with its application at a biennial competition of computational and structural biologists. The method relied on a deep learning network to predict structures.
Though DeepMind did not release details at the time, computational chemist Minkyung Baek in the Baker lab and their colleagues began to work on a similar approach. “Our work is really based on their advances,” Baker told Science. The researchers worked with a larger team including researchers at institutions in Victoria, B.C., South Africa and the United Kingdom.
Baekand, Baker and colleagues published their approach last month on the preprint server bioRxiv, and today in peer reviewed form, introducing their new application: Rose TTAFold. In the study, the researchers predicted the structure of hundreds proteins, including many that were previously only poorly understood.
In just the last month, more than 4,500 proteins have been submitted to the Baker Lab’s new server, according to a press release. Rose TTAFold “made it possible to solve the structure of one our enzymes that has caused us a lot of headache,” said Casper Wilkins, an assistant professor in biocatalysis at the Technical University of Denmark, in a tweet.
Rose TTAFold is also fast: it can predict a structure in as little as ten minutes on a gaming computer, according to the lab. This is how the team describes its system:
RoseTTAFold is a “three-track” neural network, meaning it simultaneously considers patterns in protein sequences, how a protein’s amino acids interact with one another, and a protein’s possible three-dimensional structure. In this architecture, one-, two-, and three-dimensional information flows back and forth, allowing the network to collectively reason about the relationship between a protein’s chemical parts and its folded structure.
According to Science, DeepMind’s application is more accurate, but Rose TTAFold performs nearly as well, and also better predicts some aspects of protein structure. In addition, while DeepMind’s application has been run on single proteins, Rose TTAFold can predict how proteins fit together in complexes, molecular machines that do much of the work in the body.
“We hope this new tool will continue to benefit the entire research community,” said Baekand in the press release.