A protein structure generated by a machine learning algorithm developed by the University of Washington’s Institute for Protein Design. (UW IPD Photo | Ian C. Haydon)

New protein structures not seen in nature have emerged from deep learning software deployed by University of Washington researchers and their colleagues. The approach is a step toward building bespoke proteins with specific properties, a feat that would accelerate drug development.

“This approach greatly simplifies protein design,” said David Baker, head of the UW’s Institute for Protein Design, in a statement. The research was published Wednesday in the journal Nature.

Deep learning tools recently developed by IPD and Alphabet’s DeepMind stunned scientists with their speed and accuracy at predicting the shape of proteins found in nature. Proteins fold into three-dimensional structures to perform key tasks in the body, from regulating metabolism to cell division.

In the new study, IPD researchers used deep learning to computationally generate completely new, artificial protein structures.

The researchers liken their approach to Google’s DeepDream, which can output images it calls “hallucinations” based on photos fed into it. In the new study, the software outputs new protein structures based on proteins in nature. The researchers embraced the term “hallucination” for the new structures, titling their study, “De novo protein design by deep network hallucination.”

“At no point did we guide the software toward a particular outcome — these new proteins are just what a computer dreams up,” said first author Ivan Anishchenko, a postdoctoral scholar in the Baker lab, in the statement.

The body makes proteins from amino acid building blocks, piecing them together like beads on a string. Similarly, the researchers began their study by computationally generating random chains of 100 amino acids. When they ran these chains into their prediction network, they did not look like much of anything, failing to fold up into defined structures.

In their next step, the researchers computationally mutated a random amino acid in the chain, changing it to another amino acid. If the resulting protein looked a bit more structured, they added another mutation, in thousands of iterative steps.

The study yielded 129 protein structures that resembled natural proteins in their predicted shapes. Some structures were even more streamlined, similar to ideal structures made by protein designers using more conventional approaches.

The team went on to validate some of the structures, synthesizing them in the lab. There were 27 that appeared to match the predicted structures in early experiments; three of these were chosen for more in-depth structural studies in the lab and were shown to closely match the models.

UW IPD postdoctoral researcher Ivan Anishchenko. (UW IPD Photo)

Leonid Sasanov, a professor and structural biologist at the Institute of Science and Technology in Klosterneuburg, Austria, commented on the new study in a tweet: “I admire the bravery of calling these ‘hallucinated.’ Super cool stuff!”

The new study is a step toward a bigger goal of designing proteins with potentially therapeutic actions, such as interfering with a disease-causing enzyme, or blocking a virus. The IPD released a preprint study last month that edges even further to this goal, suggesting that deep learning tools can intentionally generate proteins with functional modules.

“Exploring how to best use this strategy for specific applications is now an active area of research, and this is where I expect the next breakthroughs,” said Baker.

The Nature study used a deep learning tool that IPD built before RoseTTAfold, its recent blockbuster. The lab is now leveraging RoseTTAfold and DeepMind’s AlphaFold for a variety of protein folding problems, including in their new preprint. In November, the lab published a study predicting the structure of protein complexes, which can act like machines to do work in the body.

Several protein discovery companies have spun out of the IPD using a variety of technologies. These include A-Alpha Bio, which recently raised $20 million, and Cyrus Biotechnology, which recently inked a deal worth up to $1.5 billion with immune biotech Selecta. Alphabet also launched Isomorphic Laboratories in November to build off DeepMind’s protein folding research.

The Nature study also included researchers at Harvard University and Rensselaer Polytechnic Institute in Troy, New York.

Like what you're reading? Subscribe to GeekWire's free newsletters to catch every headline

Job Listings on GeekWork

Find more jobs on GeekWork. Employers, post a job here.