Gamers predicted protein Structure better than scientists
Oleg Lischuk, N+1
American scientists have come to the conclusion that computer players are able to predict the crystal structure of proteins more effectively than specialists or computer algorithms. The results of the work are published in the journal Nature Communications (Horowitz et al., Determining crystal structures through crowdsourcing and coursework).
Predicting the structure of a protein (the configuration in space that its amino acid sequence assumes) is of fundamental importance for medicine, biotechnology, bioinformatics and theoretical physics. At the same time, the complex and time–consuming prediction process in modern conditions has limited accuracy - about 85 percent of the models of protein crystal structures contained in databases contain significant errors. Therefore, scientists are constantly working to improve prediction techniques in order to maximize their accuracy. Recently, there has been a significant increase in interest in using distributed computing, machine learning and crowdsourcing platforms for this purpose.
Researchers from the Universities of Michigan, Washington, Massachusetts and Northeastern University in Boston held a competition to predict the structure of a protein from an electron density map obtained by X-ray crystallography. It was attended by two qualified specialists, 61 biochemist students (all of them used special software), two different computer algorithms (Phenix Autosolve and MR-Rosetta) and players in the online game Foldit, which eventually recruited 469 people. The task of this game is to manually select the configuration of the protein that has the least energy, since its three-dimensional structure is most likely to correspond to reality. The more plausible the configuration, the more points the user gets. Gamers can cooperate and form teams working on the same protein.
As a starting material, all participants were provided with an amino acid sequence, secondary structure predictions, and a map of the electronic densities of the yeast protein YPL067C. This protein was chosen because it has no significant similarity to any structure from the worldwide Protein Data Bank (PDB) database. In addition, as has been shown in previous studies, YPL067C can prevent the toxic effect of amyloid, a protein whose accumulation in the brain underlies the development of Alzheimer's disease.
The competitors used different approaches to solving the problem. Scientists and students worked separately, mainly relying on finding suitable positions for large aromatic amino acids. Among the players in Foldit, the best result was shown by a group whose members joined forces: one acted as a pioneer, arranging key elements, and the rest carried out fine-tuning of structures using a variety of approaches and techniques (in the video below).
The scientists tested the results using a special Molprobity program, comparing the key crystallographic statistics of the obtained models. The result of the players in Foldit required correction (removal of disorganized amino acid residues) related to the limitations of the gameplay. As a result, the best of the models created by gamers became the winner of the competition and one of the best models of similar resolution (1.95± 0.25 angstroms) in the PDB.
"This shows that anyone with a good three–dimensional imagination can do something previously available only to scientists, and in this way help scientific progress," researcher James Bardwell commented on the results in a press release from Video gamers outdoor scientists in contest to discover protein's shape. Another author of the paper, Scott Horowitz, expressed his intention to use Foldit for teaching students. Scientists also plan to include some techniques invented by gamers in the laboratory software.
The first version of the Foldit game, developed at the University of Washington as a platform for crowdsourcing research, became available in May 2008. During its existence, gamers have helped make several important discoveries. So, in 2010, more than 57 thousand players who predicted the protein structure more accurately than a machine algorithm were collectively listed as authors of a publication in the journal Nature. A year later, Foldit fans helped to find out the crystal structure of the retroviral protease of the Mason-Pfizer monkey virus – it took them only 10 days to solve a problem that no one could cope with for 15 years. In 2012, gamers were able to improve an artificial enzyme developed by computer modeling to catalyze the Diels-Alder reaction – the addition of 13 amino acids increased the activity of the enzyme by more than 18 times. Subsequently, the enzyme catalyzing this reaction was found in the wild.
Portal "Eternal youth" http://vechnayamolodost.ru
20.09.2016