New protein structures never before seen in nature have emerged from deep learning software deployed by University of Washington researchers and their colleagues. The method is a step towards building distinct proteins with specific properties, a feat that will accelerate drug development.
Deep learning software recently developed by Alphabet’s IPD and DeepMind scientists were shocked with their speed and accuracy in 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 dividing cells.
In the new study, IPD researchers used deep learning to compute entirely new artificial protein structures.
The researchers liken their approach to DeepDream, can output images such as faces based on the photos included. Google calls the resulting images “hallucinations,” a term the team uses for its new proteins. The study was even named “De novo Protein Design Using Deep Network Illusion”.
“We’re not guiding the software towards a particular outcome – these new proteins are just things that computers dream about,” said the first author. Ivan Anishchenko, a postdoctoral scholar at Baker Laboratories, in the statement.
The body makes proteins from amino acid building blocks, joining them together like beads on a string. Similarly, the researchers started the study by computationally generating sequences of 100 amino acids at random. When they ran these sequences into their prediction grid, they didn’t look like anything, not folding into defined structures.
In the next step, the researchers mutated a random amino acid in the sequence, changing it into another amino acid. If the resulting protein looked a little more structured, they added another mutation, in thousands of iterations.
The study produced 129 protein structures that resembled native proteins in their predicted conformation. Some of the structures are even more streamlined, similar to the ideal structure created 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 structure predicted in the initial experiments; three of these were selected for further structural studies in the laboratory and were shown to be in close agreement with the models.
Leonid Sasanov, a professor and structural biologist at the Institute of Science and Technology in Klosterneuburg, Austria, comments on the new study in a tweet: “I admire the courage to call these things ‘hallucinations.’ Great stuff! ”
The new study is a step towards the larger goal of designing proteins that have the potential to act therapeutically – by interfering with enzymes that cause disease or stopping viruses, for example. IPD released a previous study Last month, that went even further towards this goal, showing that deep learning tools can intentionally create proteins with functional modules.
“Discovering how to best use this strategy for specific applications is currently an active area of research, and this is where I look forward to further breakthroughs,” said Baker.
NS nature the study used a deep learning engine that IPD built before the blockbuster RoseTTAfold. The lab is currently implementing RoseTTAfold and DeepMind’s AlphaFold for a variety of protein folding problems, including in their new preprint. Only this November, the lab publish a study predict the structure of protein complexes, which can act like machines to do work in the body.
Several companies discovered that the protein was separated from the IPD by various technologies. Including A-Alpha Bio, which recently raised 20 million dollars, and Cyrus Biotechnology, which recently signed an agreement up to $1.5 billion with Selecta immunobiology. Alphabet also entered the arena this November, launching Isomorphic Laboratory to build DeepMind’s protein folding study.
NS nature The study also included researchers at Harvard University and Rensselaer Polytechnic Institute in Troy, New York.