• We are sorry, but NCBI web applications do not support your browser and may not function properly. More information
Logo of pnasPNASInfo for AuthorsSubscriptionsAboutThis Article
Proc Natl Acad Sci U S A. Nov 22, 2011; 108(47): 18949-18953.
Published online Nov 7, 2011. doi:  10.1073/pnas.1115898108
PMCID: PMC3223433
From the Cover
Biophysics and Computational Biology, Psychological and Cognitive Sciences, Computer Sciences

Algorithm discovery by protein folding game players

Abstract

Foldit is a multiplayer online game in which players collaborate and compete to create accurate protein structure models. For specific hard problems, Foldit player solutions can in some cases outperform state-of-the-art computational methods. However, very little is known about how collaborative gameplay produces these results and whether Foldit player strategies can be formalized and structured so that they can be used by computers. To determine whether high performing player strategies could be collectively codified, we augmented the Foldit gameplay mechanics with tools for players to encode their folding strategies as “recipes” and to share their recipes with other players, who are able to further modify and redistribute them. Here we describe the rapid social evolution of player-developed folding algorithms that took place in the year following the introduction of these tools. Players developed over 5,400 different recipes, both by creating new algorithms and by modifying and recombining successful recipes developed by other players. The most successful recipes rapidly spread through the Foldit player population, and two of the recipes became particularly dominant. Examination of the algorithms encoded in these two recipes revealed a striking similarity to an unpublished algorithm developed by scientists over the same period. Benchmark calculations show that the new algorithm independently discovered by scientists and by Foldit players outperforms previously published methods. Thus, online scientific game frameworks have the potential not only to solve hard scientific problems, but also to discover and formalize effective new strategies and algorithms.

Keywords: citizen science, crowd-sourcing, optimization, structure prediction, strategy

Citizen science is an approach to leveraging natural human abilities for scientific purposes. Most such efforts involve visual tasks such as tagging images or locating image features (13). In contrast, Foldit is a multiplayer online scientific discovery game, in which players become highly skilled at creating accurate protein structure models through extended game play (4, 5). Foldit recruits online gamers to optimize the computed Rosetta energy using human spatial problem-solving skills. Players manipulate protein structures with a palette of interactive tools and manipulations. Through their interactive exploration Foldit players also utilize user-friendly versions of algorithms from the Rosetta structure prediction methodology (6) such as wiggle (gradient-based energy minimization) and shake (combinatorial side chain rotamer packing). The potential of gamers to solve more complex scientific problems was recently highlighted by the solution of a long-standing protein structure determination problem by Foldit players (7).

One of the key strengths of game-based human problem exploration is the human ability to search over the space of possible strategies and adapt those strategies to the type of problem and stage of problem solving (5). The variability of tactics and strategies stems from the individuality of each player as well as multiple methods of sharing and evolution within the game (group play, game chat), and outside of the game [wiki pages (8)]. One way to arrive at algorithmic methods underlying successful human Foldit play would be to apply machine learning techniques to the detailed logs of expert Foldit players (9). We chose instead to rely on a superior learning machine: Foldit players themselves. As the players themselves understand their strategies better than anyone, we decided to allow them to codify their algorithms directly, rather than attempting to automatically learn approximations. We augmented standard Foldit play with the ability to create, edit, share, and rate gameplay macros, referred to as “recipes” within the Foldit game (10). In the game each player has their own “cookbook” of such recipes, from which they can invoke a variety of interactive automated strategies. Players can share recipes they write with the rest of the Foldit community or they can choose to keep their creations to themselves.

In this paper we describe the quite unexpected evolution of recipes in the year after they were released, and the striking convergence of this very short evolution on an algorithm very similar to an unpublished algorithm recently developed independently by scientific experts that improves over previous methods.

Results

In the social development environment provided by Foldit, players evolved a wide variety of recipes to codify their diverse strategies to problem solving. During the three and a half month study period (see Materials and Methods), 721 Foldit players ran 5,488 unique recipes 158,682 times and 568 players wrote 5,202 recipes. We studied these algorithms and found that they fell into four main categories: (i) perturb and minimize, (ii) aggressive rebuilding, (iii) local optimize, and (iv) set constraints. The first category goes beyond the deterministic minimize function provided to Foldit players, which has the disadvantage of readily being trapped in local minima, by adding in perturbations to lead the minimizer in different directions (11). The second category uses the rebuild tool, which performs fragment insertion with loop closure, to search different areas of conformation space; these recipes are often run for long periods of time as they are designed to rebuild entire regions of a protein rather than just refining them (Fig. S1). The third category of recipes performs local minimizations along the protein backbone in order to improve the Rosetta energy for every segment of a protein. The final category of recipes assigns constraints between beta strands or pairs of residues (rubber bands), or changes the secondary structure assignment to guide subsequent optimization.

Different algorithms were used with very different frequencies during the experiment. Some are designated by the authors as public and are available for use by all Foldit players, whereas others are private and available only to their creator or their Foldit team. The distribution of recipe usage among different players is shown in Fig. 1 for the 26 recipes that were run over 1,000 times. Some recipes, such as the one represented by the leftmost bar, were used many times by many different players, while others, such as the one represented by the pink bar in the middle, were used by only one player who chose not to share it with other players. Not surprisingly, the frequency of usage, indicated by the height of the bars, was significantly higher for the publicly shared recipes than the private ones.

Fig. 1.
Foldit recipes are used with very different frequencies. Each bar represents a different Foldit recipe; the height of the bar shows the number of times the recipe was used, and the colors denote different Foldit players. Blue Fuse, at the very left, was ...

Context Dependence.

Observing the breadth of created recipes, it is clear that Foldit players use these algorithms to augment rather than to substitute for human strategizing. Players in essence perform a problem-informed search over the space of strategies, and use recipes to encode specific lower-level strategies. Different algorithms are employed at different stages in gameplay. Fig. 2 shows the relative frequency of recipe use in the opening, midgame, and endgame. Expert players exhibit different patterns of recipe use than the player population as a whole (compare Fig. 2 A and B). The top Foldit players use recipes tlaloc Contract 3.00 and Aotearoas_Romance in the endgame, while most players use them almost equally at every stage in a puzzle (Fig. S2 A and B). Most Foldit players run after-rebuild test exclusively in the opening, but the top players often use it in the midgame as well. Local optimize recipes are heavily used in the endgame by top Foldit players (Fig. 2B and Fig. S2F).

Fig. 2.
Foldit players employ different recipes at different stages of gameplay. The relative frequency of usage for each recipe category is shown for three different stages of gameplay (blue, first third of gameplay; red, second third; green, final third) for ...

Human strategic judgment plays an important role in choosing when and how to employ different recipes. Most recipes heavily rely on the interactive aspects of the game, and are used less as fully automated tools and more as power tools that serve as an extension of the player’s strategy. We have not found any recipes that generate top models without human intervention; instead, Foldit players employ recipes to perform specific tasks at different stages of the folding process. For example, the local optimize recipes that walk along the protein backbone performing local minimizations are useless on the initial state of a Foldit puzzle that starts in an extended chain configuration, because a successful prediction will no longer have that backbone in an extended conformation and will require a new round of local optimization. Many of the aggressive rebuilding recipes require specifying which region of the protein to rebuild, as it is rarely useful to entirely rebuild a protein chain from beginning to end. These recipes are launched by players once they have converged to a low-energy solution and want to search nearby regions of conformational space before local optimization (Fig. S2 C and D); they are most useful in the midgame and often run for long periods without any human intervention. By contrast, all recipes in the set constraints category are designed to be run before launching other recipes or manually manipulating the protein (Fig. S2 G and H).

Overall, the context dependent use of different recipes is key to successful gameplay, but also makes it difficult to quantitatively evaluate the effectiveness of individual recipes because they are optimally used on quite different input protein states. This context dependence also complicates efforts to incorporate the rich new algorithms developed by Foldit players back into Rosetta or other automated methods.

Recipe Evolution.

In addition to providing tools for recipe creation, we enabled sharing, evaluation, and refinement of recipes. These tools spurred a rapid evolution of recipe use in the Foldit player population from June 2009 [when the graphical user interface (GUI) recipes were first introduced; see Materials and Methods] through the end of August 2010. Fig. 3 shows the number of distinct recipes run every week across this time period, with different shades of gray representing different recipes. Two recipes were particularly heavily used: Quake (in red) and Blue Fuse (in yellow). The essential features of the algorithms contained within these two most popular recipes are similar. Quake is a GUI recipe that repeatedly packs the side chains and minimizes the side chain and backbone torsion angles, increasing and decreasing the strength of an applied set of rubber band constraints to promote annealing of the structure. The Blue Fuse algorithm (a script recipe; see Materials and Methods) is conceptually similar to but simpler than Quake: rather than varying the strength of artificial constraint functions, the strength of atom-atom repulsive interactions is alternately increased and decreased in repeated cycles of side chain packing and complete torsion angle minimization (Fig. S3). In both cases, an initially poorly packed structure is alternately compressed (by minimization with strong constraints or weak repulsive interactions) and then relaxed (by minimization with weak constraints or strong repulsive interactions); this alternation appears to be effective in locating low energy, well packed conformations.

Fig. 3.
Rapid proliferation of Foldit recipes between June 2009 and August 2010. Each different shade of gray represents an individual recipe with the size of the gray bar denoting how many times that recipe was run that particular week. The most heavily used ...

Foldit algorithms evolve in a social fashion, with popular recipes copied and then elaborated upon. An evolutionary tree depicting the evolution of the Blue Fuse algorithm family is shown in Fig. 4; size corresponds to the logarithm of the number of recipe uses and color represents the recipe author. Popular recipes spawn large numbers of descendants, and there are multiple independent lineages each spanning many generations.

Fig. 4.
Social evolution of Foldit recipes. Each circle represents a different Foldit recipe, each color denotes a different author, and the size is the logarithm of the number of times a recipe was used. Arrows represent the process of copying and subsequent ...

While the evolutionary mechanisms by which new algorithms are created (copying and variation) are clear, the mechanisms through which individual recipes spread through the population are less so. The popularity of Quake and Blue Fuse could result from individual player experimentation, player communication through chat, or player communication within teams. Interviews with Foldit players suggest that the primary mechanism by which successful algorithms take over the population is word of mouth. Players test and then recommend new recipes to others and rate recipes on the Foldit website, making decisions in part based on the reputation of the recipe author.

Similarity of Independently Developed Player and Scientist Algorithms.

During this same time period, researchers in the Baker group were attempting to improve the core optimization methods within the Rosetta structure prediction and design program. A breakthrough was made with the Fast Relax algorithm, which achieves more effective energy optimization in less time than previous Rosetta methods. Benchmarks have shown Fast Relax to be considerably more efficient then an older algorithm, Classic Relax (12) used since 2002 (Fig. 5, compare green to red), and it is now used in all Rosetta de novo and homology modeling methods as well as enzyme design protocols.

Fig. 5.
Performance comparison between Rosetta Relax protocols and Foldit recipes. Each optimization protocol was run on a benchmark set of 6,758 models from 62 different proteins, and the average energy was calculated over all models after different numbers ...

The key innovation in the Fast Relax algorithm compared to Classic Relax and earlier methods is alternately increasing and decreasing the strength of the repulsive interactions. As diagrammed in Fig. 6, Fast Relax is comprised of interlaced cycles of full combinatorial repacking of the side chains of the protein (repack) and gradient-based minimization of the backbone and side chain degrees of freedom (minimize); within each round the repulsive weight is increased while from the end of one round to the start of the next (indicated by the arrow) the repulsive weight is decreased 50-fold. Typically 5–15 rounds of this repulsive weight annealing are applied to a given structure and the lowest energy structure encountered (only full repulsive weight structures are eligible) is finally kept as the result.

Fig. 6.
Similarity between the unpublished Rosetta protocol Fast Relax and the Foldit recipe Blue Fuse. Equivalent steps in the two algorithms are represented with the same color. The repulsive interaction strength is reset from high to low in each cycle of Fast ...

There is a striking similarity between the Fast Relax algorithm developed by scientists and the Blue Fuse algorithm developed by Foldit game players. These similarities are evident in the algorithm comparison shown in Fig. 6. Both algorithms ramp the repulsive weight up and down while repacking and minimizing the structure, ultimately selecting as the algorithm output the lowest energy structure encountered. There are minor differences—Blue Fuse begins with a low repulsive weight and only performs a shake before minimizing the structure at the standard weight while Fast Relax does a repack/minimize cycle at each stage—but the similarities far outweigh the differences.

The Foldit players’ algorithmic discovery is in a rich tradition of softening the repulsive forces to enhance sampling in protein folding calculations. The earliest methods replaced entire subsets of atoms with simplified centroid side chains (11), and subsequent methods softened repulsive interactions in full atom representations (13). Classic Relax expanded on this approach by gradually ramping up the repulsive term during the course of a simulation. Through social evolution of recipes, Foldit players independently rediscovered the utility of initially softening the repulsive interactions. The players went beyond previous approaches by introducing the sawtooth-like repeated annealing of the strength of the repulsive interactions as in Fast Relax. This sawtooth profile likely induces repeated compaction and expansion of the protein chain, which evidently helps access new energy minima.

Performance Comparison.

To determine how the Blue Fuse algorithm compared to both the Classic Relax and Fast Relax protocols, we ran all three algorithms on an in-house standard benchmark set consisting of 62 proteins with a range of structural diversity, including monomeric as well as multimeric structures with and without ligands. For each protein we included both native and close-to-native structures as well as nonnative Rosetta models for a total of 6,758 structures. The CPU time required for each of the methods can be varied by changing the number of iterations in the outer loop, and we recorded the average energy over all of the 6,758 models as a function of CPU time. Fig. 5 shows that the Blue Fuse algorithm (in blue) performs more efficiently than Classic Relax (in red), but not as well as the newer Fast Relax protocol (in green). Thus, the algorithm encoded in the most popular player recipe is not only structurally similar to the Fast Relax but also more efficient than previously published Rosetta algorithms.

Rosetta optimizations are simplified to run at interactive speeds suitable for human exploration in Foldit (SI Text), and hence Fast Relax in Rosetta utilizes more powerful elementary optimization modules than Blue Fuse in Foldit. To evaluate the performance of Fast Relax relative to Blue Fuse–when given access only to the simplified optimizations in Foldit that are available to game players–we encoded the Fast Relax algorithm using the Foldit script interface. Fig. 7 shows the performance of this Foldit script version of Fast Relax (in dark green) on the same benchmark set; although it is still able to sample lower energies than Blue Fuse, it takes longer to do so. Notably, Blue Fuse outperforms this Foldit version of Fast Relax for CPU times less than 200 s (Fig. S4), and in practice Foldit players run Blue Fuse for an average runtime of 122 s (dotted line in Fig. 7, Fig. S4). Thus, given the tools available in the game, and for the typical times these tools are used within the game, the player discovered Blue Fuse algorithm is actually superior to Fast Relax.

Fig. 7.
Performance comparison between Blue Fuse and Rosetta Fast Relax protocol using the streamlined Foldit minimization and side chain optimization routines. The dark green line shows the performance of Fast Relax encoded using the Foldit scripting interface. ...

We tested other popular recipes (shown in Table S1), but found none that minimized the energy as quickly as Blue Fuse. Many recipes yielded lower energies than Blue Fuse but took much longer to do so. While repeated iterations of Blue Fuse fail to decrease the energy further, Foldit players discovered that by combining different recipes together with Blue Fuse, lower energies can be achieved than with any of them alone; for example, A Deep Breath in Fig. 5. Again, context dependence is important: the authors of A Deep Breath, for example, recommend adding rubber bands to a protein before running.

Discussion

The introduction of tools for Foldit players to codify their strategies spawned a flurry of creative activity and the creation of a remarkable diversity of folding algorithms. This diversity emerged through the evolution of player recipes over the course of a year since their introduction. Particularly remarkable is the convergence of this evolution onto an algorithm similar to one recently developed by scientific experts over the same period of time. Both algorithms achieve faster and more effective energy optimization than previous methods, and within the context of the game, the player algorithm is the most efficient. Foldit players have been at a considerable disadvantage compared to scientists in developing new algorithms as the scripting language only exposes a small fraction of the variables and base algorithms in the Rosetta codebase. We are now generalizing the scripting language to allow control over more of these features, and look forward to learning what Foldit player ingenuity can do with these additional capabilities. More generally, the explosion of Foldit algorithms and the convergence on the best algorithm discovered thus far by scientists highlights the potential of scientific discovery games for significant contributions to science and technology, particularly in the creation and formalization of complex problem-solving strategies.

Materials and Methods

In order to empower the widest possible pool of active players to create recipes, including those without basic programming skills, we provided two recipe creation pathways. The first recipe creation tool, provided to Foldit players in June 2009, was a simple block-based visual programming interface where different actions are added from a pull-down menu. Available actions include adding or adjusting the strength of rubber bands (soft constraints which connect amino acids and pull on them independently of the player), restoring previous structures (allowing backtracking), and invoking optimization methods such as shake and wiggle. Recipes created using this initial GUI (referred to as GUI recipes) could not utilize conditionals or loops, or modify certain properties such as the clash importance, which lets players adjust the strength of the Rosetta atom-atom steric overlap energy term. To support more advanced recipes, we added a text-based interface, using the Lua scripting language (14), to the game in October 2009. In addition to having many more Foldit actions available, this interface gives players the ability to control the flow of recipes (using conditionals and loops), allowing for the creation of much more complex algorithms (these are referred to as script recipes).

We analyzed the evolution and use of Foldit recipes during the CASP9 structure prediction experiment from May–August 2010 (15). Because Foldit is an online game, we were able to track recipe usage and analyze the recipes developed and employed by Foldit players during this period.

Availability

Foldit player recipes are available on the Foldit website: http://fold.it/portal/recipes, with instructions on how to download them described on the Foldit wiki: http://foldit.wikia.com/wiki/101_-_Cookbook.

Supplementary Material

Supporting Information:

Acknowledgments.

We thank the members of the Foldit team for their help designing and developing the game and all the Foldit players who have made this work possible. This work was supported by the Center for Game Science, Defense Advanced Research Projects Agency (DARPA) Grant N00173-08-1-G025, the DARPA Protein Design Processes (PDP) program, National Science Foundation (NSF) Grants IIS0811902 and IIS0812590, the Howard Hughes Medical Institute (D.B.), a Henry Wellcome Postdoctoral Fellowship (M.D.T), Adobe and Microsoft. This material is based upon work supported by the National Science Foundation under Grant 0906026.

Footnotes

The authors declare no conflict of interest.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1115898108/-/DCSupplemental.

References

1. Lintott C, et al. Galaxy Zoo: ‘Hanny’s Voorwerp’, a quasar light echo? Mon Not R Astron Soc. 2009;399:129–140.
2. Westphal AJ, et al. Non-destructive search for interstellar dust using synchrotron microprobes; X-ray Optics and Microanalysis: Proceedings of the 20th International Congress; Melville, NY: American Institute of Physics (AIP); 2010. pp. 131–138.
4. Cooper S, et al. The challenge of designing scientific discovery games; Proceedings of the Fifth international Conference on the Foundations f Digital Games; New York, NY: FDG ACM; 2010. pp. 40–47.
5. Cooper S, et al. Predicting protein structures with a multiplayer online game. Nature. 2010;466:756–760. [PMC free article] [PubMed]
6. Rohl CA, Strauss CE, Misura KM, Baker D. Protein structure prediction using Rosetta. Methods Enzymol. 2004;383:66–93. [PubMed]
7. Khatib F, et al. Crystal structure of monomeric retroviral protease solved by protein folding game players. Nat Struct Mol Biol. 2011;18:1175–1177. [PMC free article] [PubMed]
8. Foldit Wiki. http://Foldit.wikia.com/
9. Banerji M, et al. Galaxy Zoo: Reproducing galaxy morphologies via machine learning. Mon Not R Astron Soc. 2010;406:342–353.
10. Cooper S, et al. Analysis of social gameplay macros in the Foldit cookbook; Proceedings of the Sixth international Conference on the Foundations f Digital Games; New York, NY: FDG ACM; 2011.
11. Levitt M, Warshel A. Computer simulation of protein folding. Nature. 1975;253:694–698. [PubMed]
12. Kuhlman B, et al. Design of a novel globular protein fold with atomic-level accuracy. Science. 2003;302:1364–1368. [PubMed]
13. Levitt M. Protein folding by restrained energy minimization and molecular dynamics. J Mol Biol. 1983;170:723–764. [PubMed]
14. The Programming Language Lua. http://www.lua.org/

Articles from Proceedings of the National Academy of Sciences of the United States of America are provided here courtesy of National Academy of Sciences
PubReader format: click here to try

Formats:

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...

Links

  • MedGen
    MedGen
    Related information in MedGen
  • PubMed
    PubMed
    PubMed citations for these articles

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...