Members of Foldit's online gaming community have solved the structure of a protein-cutting enzyme from an AIDS-like virus whose configuration has eluded researchers for more than a decade.
Within three weeks, the Foldit Contenders group, which comprises around 15 active players, used the protein-folding game in which players compete to create the lowest-energy model to successfully create a model of a retroviral protease that closely matched researchers' experimental data.
With the gamers' model in hand, a team comprising investigators from the University of Washington, Poland's A. Mickiewicz University, and the Academy of Sciences of the Czech Republic, was able to solve the enzyme's structure using molecular replacement. The technique is often used in x-ray crystallography to address phase problems that can occur while determining structures from diffraction data.
Retroviral proteases play an important role in how the AIDS virus matures and proliferates and are currently being studied as potential targets for drugs that can block their activity. The Foldit players solved the structure of the Mason-Pfizer monkey virus, which causes AIDS in monkeys.
Once the structure was solved, the researchers identified novel structural features that could provide "opportunities for the design of antiretroviral drugs, including anti-HIV drugs," they wrote in an article published in a recent issue of Nature Structural & Molecular Biology that discussed the gamers' findings.
The authors also note in the paper that these results mark the "first instance that we are aware of in which gamers were able to solve a longstanding scientific problem."
They also pointed out that the results "indicate the potential for integrating video games into the real-world scientific process," adding that "the ingenuity of game players is a formidable force that, if properly directed, can be used to solve a wide range of scientific problems."
The Problem with Proteins
"Trying to determine the 3D structure of a protein experimentally is very expensive and very time consuming and in some cases such as this one, experimental methods don’t even work," paper co-author Firas Khatib, a postdoctoral researcher in David Baker's lab at UW, told BioInform.
Methods that scientists use for this task include x-ray crystallography, nuclear magnetic resonance, and dual polarization interferometry.
In most cases where protein structures are solved using x-ray crystallography, researchers use molecular replacement — which relies on the existence of a similar previously solved protein as a model to find the unknown structure — to solve the final structure, Khatib explained.
Usually, researchers "have all this experimental data, and what you need is a model that will fit this data to then be able to get the correct structure because otherwise the experimental data doesn’t make any sense," he explained.
However, the M-PMV protease proved to be an "extremely nasty" protein that resisted all attempts to understand its structure, Khatib said. Although researchers had solved the protein using NMR, the models that they generated were inaccurate with a lot of variation between the structures.
One of the authors on the current paper, Frank DiMaio, a postdoctoral researcher in UW's biochemistry department, and other colleagues published a paper in Nature earlier this year describing the use of the Rosetta structure-prediction software, developed by Baker's group at UW, to step in when experimental methods failed.
In that paper, the authors noted that methods like molecular replacement fail when the starting models have less than "30 percent sequence identity" and suggested that "combining algorithms for protein structure modeling with those developed for crystallographic structure determination" could help address the problem.
Using Rosetta in combination with Autobuild chain tracing, the researchers obtained high-resolution structures for eight out of 13 x-ray diffraction datasets that crystallographers had been unable to solve. The M-PMV protease was among the five that remained a mystery.
With this challenge, "we wanted to see if human intuition could succeed where automated methods had failed," Khatib said in a statement.
Using Foldit, which is also based on the Rosetta algorithm, the gamers were able to generate models that were good enough for the researchers to refine and, within a few days, determine the enzyme's structure. Additionally, surfaces on the molecule stood out as likely targets for antiretroviral drugs.
One of the members of the Foldit Contenders group, who wished to be referred to by her screen name, Mimi, told BioInform that finding the solution took "team effort," in which members improved on each other's structures.
The Nature Structural & Molecular Biology paper explained that one player used partial threading with Foldit's alignment tool to improve the starting NMR model while two other teammates made improvements to the core of the protein model and tucked in a loop to generate the final structure.
Mimi, who has been playing Foldit for about three years, said that solving these structures is "very much a question of using your judgment as to what shape you felt the protein ought to be in."
Basically, "you are looking for a shape that is the lowest energy possible and looking to make it as compact as possible," she explained.
CharlieFortsConscience, another member of the group who also wished to remain anonymous, expressed similar sentiments.
Having played the game for two and a half years, "you can't help but develop a strategy that's particular to you and the way you think, and that underpins the other reason why Foldit is so powerful," he told BioInform in an e-mail.
Although he has a background in biochemistry and immunology, CharlieFortsConscience said he is one of the few in Foldit's gaming community that have a scientific background — a fact that demonstrates that a knowledge of the field isn't a requirement to play.
"We all can add our own little bit of human intuition to the process, and it can have the most amazing consequences," he said. "There is a wonderful principle here that keeps being confirmed over and over — whereas computers are unbeatable at crunching numbers, human beings can look at a virtual protein in front of them, and know when something doesn't look right."
Citizen Science
Foldit was developed in 2008 as a collaborative effort between researchers in UW's departments of computer science and engineering and biochemistry as a way to engage non-scientists' help in solving protein prediction problems.
The game was conceived as an offshoot of the Rosetta@Home distributed computing project that runs the Rosetta software on donated PC cycles. The goal was to create a more interactive version of the software that would have a competitive angle to attract the interest of users.
Like many online games, Foldit has several communities, leaderboards, competitions, and teams. The developers post weekly protein structure puzzles that gamers can sink their teeth into.
Usually, "we take a partially folded protein ... post it on the game and the players are able to manipulate the structure," Seth Cooper, a professor at UW's department of computing science and engineering and one of Foldit's creators, told BioInform. "While they are playing, they are being recorded on our servers so we can see the kind of structures that the players come up with."
These servers usually record "tens if not hundreds of thousands of structures," he said.
The developers use a software program that analyzes the structures and generates a score that describes how well a protein is folded.
"It doesn’t need to know the answer," he said. "It's based on the principles of how proteins fold. You can give it a structure and it will tell you how well folded it is but it can't tell you what the best fold of the structure is.
In the case of the M-PMV protease, "we were able to take the [structures] that scored the best and were therefore the most likely to be the right answer ... and gave it the experimentalists who were able to confirm that what the players had found was the correct structure using the real experimental data."
Prior to taking on real-life situations the team put Foldit and its players through their paces using test datasets, Cooper said.
The results were described in a Nature paper published in 2010 in which the team evaluated players' abilities to solve structural prediction problems by posting 10 "blind structures" for which "neither the target protein nor homologous proteins had structures contained within publicly available databases for the duration of the puzzles."
Among other findings reported in that paper, the developers said that players were able to perform "major protein restructuring operations to change beta sheet hydrogen-bond patterns" — a feat that proved difficult for automated methods.
In another outing discussed in the Nature Structural & Molecular Biology paper, Foldit players participated in the ninth Critical Assessment of Techniques for Protein Structure Prediction experiment held last year.
According to the paper, Foldit did not perform to expectations in the Template-Based Modeling category, in which participants make predictions about protein structures that have known homologs, but it did well in the Free Modeling category. In this category, where participants predict structures de novo, it produced "one of the best overall predictions" for one of the protein targets, the researchers said.
In spite of shortcomings that were highlighted by the CASP experience, the researchers point to this recent success with the M-PMV protein as proof of Foldit's value.
The developers already have other challenges up their sleeves for Foldit participants.
For example, they are asking players to design non-existent proteins that could inhibit flu proteins. Other projects involve creating proteins for biofuels, to degrade environmental waste, and design nanodevices.
One of the goals of this paper is that "scientists will see the power of the online gaming community and [realize that] it's not just for proteins," Khatib said. "There are many applications we believe where citizen science can help solve scientific problems."
Some other games that engage non-scientists in scientific problems include EteRNA, an online RNA folding game that one of Foldit's developers now at Carnegie Mellon University co-developed with Stanford University scientists; and Phylo, which was developed by researchers at McGill University and challenges players to find the best possible alignment for genetic sequences (BI 12/3/2010).
"One of the things that this [challenge] shows is that with a video game, you are able to combine the things that people are good at, which in this case was the spatial reasoning and problem solving, ... with what the computers are good at, which is optimization and crunching numbers and that sort of thing," Khatib said. This way, "both the humans and computers can both do what they are best at and use this force for solving new and difficult problems."
Have topics you'd like to see covered in BioInform? Contact the editor at uthomas [at] genomeweb [.] com