A new study examines the structure of hundreds of protein complexes, tiny machines in cells that control everything from energy use to DNA replication. Researchers at the Institute for Protein Design at the University of Washington helped lead the study, which harnesses new deep learning tools and could lead to new ways of treating the disease.
This spring, the IPD unveiled its deep learning software for predicting protein folding, following on from a similar tool built by DeepMind, a subsidiary of Alphabet. The tools stunned researchers with their speed and accuracy in predicting how proteins transform into three-dimensional shapes.
Proteins are made up of building block chains of amino acids, but they have to bend properly in order to function. IPD’s RoseTTAFold and DeepMind’s AlphaFold have been used to predict the shape of thousands of proteins since their release.
Inside cells, proteins often interact with each other in machine-like protein complexes that perform various tasks. Many approved drugs also interfere with protein complexes, such as chemotherapy drugs that hijack the machines involved in DNA replication and cell division.
“To truly understand the cellular conditions that give rise to health and disease, it is essential to know how different proteins in a cell work together,” said the co-lead author. Ian humphrey, a graduate student from the laboratory of the head of the IPD David Boulanger, in a press release.
In the new study, Humphrey, Baker and their colleagues model most of the protein interactions that occur in yeast Saccharomyces cerevisiae. This single-celled organism resembles human cells in the way it performs basic functions such as growth, division, waste removal, and environmental sensing, all controlled by protein complexes.
Yeast contains around 6000 proteins. To predict which of these proteins might interact, the researchers turned to evolutionary biology. As proteins evolve, they often accumulate mutations in tandem – if one building block is changed in a protein, a corresponding building block is changed in a partner protein. Such changes in tandem ensure that the complex remains intact.
The researchers identified pairs of proteins that acquired mutations so closely linked, suggesting that they might interact physically. They then deployed RoseTTAFold and AlphaFold to model the three-dimensional shape of these interacting proteins.
After sifting through millions of potential pairings, Deep Learning Tools extracted 1,506 potentially interacting proteins. From these proteins, the tools succeeded in modeling 712 protein complexes.
Over 100 of the protein interactions had not been identified previously. One of the new complexes contains a protein involved in DNA repair and cancer progression, and another contains an enzyme involved in neurodevelopmental disorders and cancer.
The new findings open the door to future investigations examining the complexes and how they function. And they could ultimately lead to drugs that interfere with the cellular machinery involved in the disease.
“These models give hypotheses to be tested by the experimenters”, Qian Cong Recount Science, who published study Thusday. Cong is co-author with Baker and was a researcher at UW before becoming an assistant professor at Southwestern Medical Center at the University of Texas last year.
The new findings also paved the way for further studies using RoseTTAFold and DeepMind to map the universe of human protein complexes.
The study involved computer experts, evolutionary researchers, and structural biologists who helped interpret three-dimensional protein models.
“As computational methods become more powerful, it’s easier than ever to generate large amounts of scientific data, but to make sense of it all, you still need scientific experts,” Baker said in the statement. “This is community science at its best. “