Rosetta@home

Like all BOINC projects, Rosetta@home uses idle computer processing resources from volunteer computers to perform calculations on individual workunits, which are sent to a central project server where they are validated and assimilated into project databases upon completion. The project is cross-platform, and runs on a wide variety of hardware configurations. Users can view the progress of their individual protein structure prediction on the Rosetta@home screensaver.

In addition to disease related research, the Rosetta@home network serves as a testing framework for new methods in the aforementioned areas of structural bioinformatics. These new methods are then used in other Rosetta-based applications, like RosettaDock and the Human Proteome Folding Project, after being sufficiently developed and proven stable on Rosetta@home's large and diverse collection of volunteer computers. Two particularly important tests for the new methods developed in Rosetta@home are the Critical Assessment of Techniques for Protein Structure Prediction (CASP) and Critical Assessment of Prediction of Interactions (CAPRI) experiments, biannual experiments which evaluate the state of the art in protein structure prediction and protein-protein docking prediction, respectively. Rosetta@home consistently ranks among the foremost docking predictors, and is one of the best tertiary structure predictors available.^[3]

Computing platform

Both the Rosetta@home application and the BOINC distributed computing platform are available for the Microsoft Windows, Linux and Macintosh platforms (BOINC also runs on several other platforms, e.g. FreeBSD^[4]). Participation in Rosetta@home requires a central processing unit (CPU) with a clock speed of at least 500 MHz, 200 megabytes of free disk space, 256 megabytes of physical memory, and Internet connectivity.^[5] As of July 1, 2008, the current version of the Rosetta application is 5.98^[6] and the current BOINC program version is 5.10.^[4] Standard HTTP (port 80) is used for communication between the user's BOINC client and the Rosetta@home servers at the University of Washington; HTTPS (port 443) is used during password exchange. Remote and local control of the BOINC client use port 31416 and port 1043, which may need to be specifically unblocked if they are behind a firewall.^[7]

Rosetta@home screensaver, showing the progress of a structure prediction for a synthetic ubiquitin protein (PDB ID: 1ogw)

Like all BOINC projects, Rosetta@home runs in the background of user's computer using idle computer power, either at or before logging in to an account on the host operating system. The application takes a lowest-priority status compared to all other tasks. In other words, Rosetta@home frees resources from the CPU as they are required by other applications so that normal computer usage is unaffected. To minimize power consumption or heat production from a computer running at sustained capacity, the maximum percentage of CPU resources that Rosetta@home is allowed to use can be specified through a user's account preferences. The hours of the day during which Rosetta@home is allowed to do work can also be adjusted, along with many other preferences, through a user's account settings.

Rosetta, the underlying application run on the Rosetta@home network, was rewritten in C++ to facilitate easier development than that offered by its original version, which was written in Fortran. This new version is object-oriented, and debuted on February 8, 2008.^[9]^[10] Development of the Rosetta code is done by Rosetta Commons. The software is freely licensed to the academic community; it is available to pharmaceutical companies through a fee.^[11]

Project significance

The first close to atomic-level resolution, ab initio structure prediction ? CASP6 target T0281. Rosetta produced a model for T0281 (superpositioned in magenta) 1.6 � RMSD from the crystal structure (blue).

Protein 3D structures are currently determined experimentally through X-ray crystallography or nuclear magnetic resonance (NMR) spectroscopy. The process is slow (it can take weeks or even months to figure out how to crystallize a protein for the first time) and comes at high cost (around $100,000 USD per protein^[12]). Unfortunately, the rate at which new sequences are discovered far exceeds the rate of structure determination -- out of more than 6,600,000 protein sequences available in the NCBI non-redundant (nr) protein database, less than 48,000 proteins' 3D structures have been solved and deposited in the Protein Data Bank, the main repository for structural information on proteins.^[13] One of the main goals of Rosetta@home is to predict protein structures with the same accuracy as existing methods, but in a way that requires significantly less time and money.

Progress in protein structure prediction is evaluated in the biannual Critical Assessment of Techniques for Protein Structure Prediction (CASP) experiment, in which researchers from around the world attempt to provide a protein's structure from the protein's amino acid sequence. High scoring groups in this sometimes competitive experiment are considered the de facto standard-bearers for what is the state of the art in protein structure prediction. Rosetta, the program on which Rosetta@home is based, has been used since CASP 5 in 2002. In the 2004 CASP 6 experiment, Rosetta made history by being the first to produce a close to atomic level resolution, ab initio prediction in its submitted model for CASP target T0281. Rosetta@home has been used in CASP since 2006, where it was among the top predictors in every category of structure prediction in CASP 7.^[14]^[15]^[16] These high quality predictions were enabled by the computing power made available by Rosetta@home volunteers.^[17]

Rosetta@home is also used in protein docking prediction, which determines the structure of multiple complexed proteins, or quaternary structure. This type of protein interaction affects many cellular functions, including antigen-antibody and enzyme-inhibitor binding and cellular import and export. Determining these interactions is critical for drug design. Rosetta is used in the Critical Assessment of Prediction of Interactions (CAPRI) experiment, which evaluates the state of the protein docking field similar to how CASP gauges progress in protein structure prediction. The computing power made available by Rosetta@home's project volunteers has been cited as a major factor in Rosetta's performance in CAPRI, where its docking predictions have been among the most accurate and complete.^[18]

In early 2008, Rosetta was used to computationally design a protein with a function never before observed in nature.^[19] This was inspired in part by the retraction of a high-profile paper from 2004 which originally described the computational design of a protein with improved enzymatic activity compared to its natural form.^[20] The 2008 research paper from David Baker's group describing how the protein was made, which cited Rosetta@home for the computational resources it made available, represented an important proof of concept for this protein design method^[19] This type of protein design could have future applications in drug discovery, green chemistry, and bioremediation.^[19]

Disease related research

In addition to basic research in predicting protein structure, docking and design, Rosetta@home is also used in immediate disease related research.^[21] Numerous minor research projects are described in David Baker's Rosetta@home journal.^[22] Other disease related work includes:

Amyloid illnesses

A component of the Rosetta software suite, RosettaDesign,^[23] was used to accurately predict which regions of amyloidogenic proteins were most likely to make amyloid-like fibrils. By taking hexapeptides (six amino acid-long fragments) of a protein of interest and selecting the lowest energy match to a structure similar to that of a known fibril forming hexapeptide, RosettaDesign was able to identify peptides twice as likely to form fibrils as are random proteins.^[24] Rosetta@home was used to in the same study to predict structures for amyloid beta, a fibril-forming protein that has been linked to Alzheimer's disease.^[25] Preliminary but as-yet unpublished results have been produced on Rosetta-designed proteins that may prevent fibrils from forming.^[26]

Anthrax

Another component of Rosetta, RosettaDock,^[27]^[28]^[29] was used in conjunction with experimental methods to model interactions between three proteins -- lethal factor (LF), edema factor (EF) and protective antigen (PA) -- that make up anthrax toxin. The computational model accurately predicted docking between LF and PA, helping to establish which domains of the respective proteins are involved the LF-PA complex. This insight was eventually used in research resulting in improved anthrax vaccines.^[30]^[31]

Herpes simplex virus 1

RosettaDock was also used to model docking between an antibody (immunoglobulin G) and a surface protein expressed by herpes simplex virus 1 (HSV-1) which serves to degrade the antiviral antibody. The protein complex predicted by RosettaDock closely agreed with hard-gotten experimental models, leading researchers to conclude that the docking method has potential in addressing some of the problems that X-ray crystallography has with modeling protein-protein interfaces.^[32]

HIV

Malaria

In research involved with the Grand Challenges in Global Health initiative,^[36] Rosetta has also been used to computationally design novel homing endonuclease proteins, which could eradicate Anopheles gambiae or otherwise render the mosquito unable to transmit malaria.^[37] Being able to model and alter protein-DNA interactions specifically, like those of homing endonucleases, gives computational protein design methods like Rosetta an important role in gene therapy (which includes possible cancer treatments).^[21]^[38]

Development history and branches

Originally introduced by the Baker laboratory in 1998 as an ab initio approach to structure prediction,^[39] Rosetta has since branched into several development streams and distinct services. More than 7 years after Rosetta's first appearance, the Rosetta@home project was released (i.e. announced as no longer beta) on October 6, 2005.^[9] Many of the graduate students and other researchers involved in Rosetta's initial development have since moved to other universities and research institutions, and subsequently enhanced different parts of the Rosetta project.

Superposition of Rosetta-designed model (red) for TOP7 onto its x-ray crystal structure (blue, PDB ID: 1QYS)

RosettaDesign

RosettaDesign, a computational approach to protein design based on Rosetta, began in 2000 with a study in redesigning the folding pathway of protein G.^[40] In 2002 RosettaDesign was used to design TOP7, a 93-amino acid long α/β protein that had an overall fold never before recorded in nature. This new conformation was predicted by Rosetta to within 1.2 � RMSD of the structure determined by x-ray crystallography, representing an unusually accurate structure prediction.^[41] Rosetta and RosettaDesign earned widespread recognition by being the first to design and accurately predict the structure of a novel protein of such length, as reflected by the 2002 paper describing the dual approach prompting two positive letters in the journal Science,^[42]^[43] and being cited by more than 240 other scientific articles.^[44] The visible product of that research, TOP7, was featured as the Protein Data Bank's 'Molecule of the Month' in October 2006;^[45] a superposition of the respective cores (residues 60-79) of its predicted and x-ray crystal structures are also featured in the Rosetta@home logo.^[46]

RosettaDock

RosettaDock was added to the Rosetta software suite during the first CAPRI experiment in 2002 as the Baker laboratory's algorithm for protein-protein docking prediction.^[49] In that experiment, RosettaDock made a high-accuracy prediction for the docking between streptococcal pyogenic exotoxin A and a T cell-receptor β-chain, as well as a medium accuracy prediction for a complex between porcine ?-amylase and a camelid antibody. While the RosettaDock method only made two acceptably accurate predictions out of seven possible, this was enough to rank it seventh out of nineteen prediction methods in the first CAPRI assessment.^[49]

Development of RosettaDock diverged into two branches for subsequent CAPRI rounds as Jeffrey Gray, who laid the groundwork for RosettaDock while at the University of Washington, continued working on the method in his new position at John Hopkins University, and members of the Baker laboratory further developed RosettaDock in Gray's absence. The two versions differed slightly in side-chain modeling, decoy selection and other areas,^[50]^[51] but both the Baker and Gray methods performed well in the second CAPRI assessment, placing 5th and 7th respectively out of 30 predictor groups.^[52] Jeffrey Gray's RosettaDock server is available as a free docking prediction service for non-commercial use.^[53]

In October 2006, RosettaDock was integrated into Rosetta@home. The method used a fast, crude docking model phase using only the protein backbone, followed by a slow full-atom refinement phase in which the orientation of the two interacting proteins relative to each other, as well as side-chain interactions at the protein-protein interface, were simultaneously optimized to find the lowest energy conformation.^[54] The vastly increased computational power afforded by the Rosetta@home network, in combination with revised "fold-tree" representations for backbone flexibility and loop modeling, made RosettaDock 6th out of 63 prediction groups in the third CAPRI assessment.^[3]^[18]

Robetta

The Robetta server is an automated protein structure prediction service offered by the Baker laboratory for on-commercial ab initio and comparative modeling.^[55] It has participated as an automated prediction server in the biannual CASP experiments since CASP 5 in 2002, performing among the best in the automated server prediction category.^[56] Robetta has since competed in CASP 6 and 7, where it did better than average among both automated server and human predictor groups.^[16]^[57]^[58]

In modeling protein structure as of CASP 6, Robetta first searches for structural homologs using BLAST, PSI-BLAST, and 3D-Jury, then parses the target sequence into its individual domains, or independently folding units of proteins, by matching the sequence to structural families in the Pfam database. Domains with structural homologs then follow a "template-based model" (i.e., homology modeling) protocol. Here, the Baker laboratory's in-house alignment program, K*sync, produces a group of sequence homologs, and each of these is modeled by the Rosetta de novo method to produce a decoy (possible structure). The final structure prediction is selected by taking the lowest energy model as determined by a low-resolution Rosetta energy function. For domains that have no detected structural homologs, a de novo protocol is followed in which the lowest energy model from a set of generated decoys is selected as the final prediction. These domain predictions are then connected together to investigate inter-domain, tertiary-level interactions within the protein. Finally, side-chain contributions are modeled using a protocol for Monte Carlo conformational search.^[59]

In CASP 8, Robetta was augmented to use Rosetta's high resolution all-atom refinement method,^[60] the absence of which was cited as the main cause for Robetta being less accurate than the Rosetta@home network in CASP 7.^[17]

Comparison to similar distributed computing projects

There are several distributed computed projects which have study areas similar to those of Rosetta@home, but differ in their research approach:

Folding@home

Of all the major distributed computing projects involved in protein research, Folding@home is the only one to not use the BOINC platform. Both Rosetta@home and Folding@home research protein misfolding diseases (e.g. Alzheimer's disease), but Folding@home does so much more exclusively. Instead of using structure- or design-based methods to predict amyloid behavior, for example, Folding@home uses molecular dynamics to model how proteins fold (or potentially misfold, and subsequently aggregate). In other words, Folding@home's strength is protein folding, while Rosetta@home's strength is protein design and prediction of structure and docking. The two projects also differ significantly in their computing power and host diversity. Averaging about 2.5 petaFLOPS (2500 teraFLOPS) with a host base that includes the Playstation 3 and graphics processing units, Folding@home has more than a 35-fold advantage in computing power over Rosetta@home, which averages 66 teraFLOPS with a host base consisting only of PC-based CPUs.

World Community Grid

Both Phase I and Phase II of the Human Proteome Folding Project (HPF), a subproject of World Community Grid, have used the Rosetta program to make structural and functional annotations of various genomes.^[61]^[62] Although he now uses it to create databases for biologists, Richard Bonneau, head scientist of the Human Proteome Folding Project, was active in the original development of Rosetta at David Baker's laboratory while obtaining his PhD.^[63] More information on the relationship between the HPF1, HPF2 and Rosetta@home can be found on Richard Bonneau's website.^[64]

Predictor@home

Like Rosetta@home, Predictor@home specializes in protein structure prediction. Predictor@home also has plans to develop new areas for its distributed computing platform in protein design and docking (using the CHARMM package for molecular dynamics),^[65] further likening it to Rosetta@home. While Rosetta@home uses the Rosetta program for its structure prediction, Predictor@home uses the dTASSER methodology.^[66]

Other protein related distributed computing projects on BOINC include QMC@home, Docking@home, POEM@home, SIMAP, and TANPAKU (site in Japanese). RALPH@home, the Rosetta@home alpha project which tests new application versions, work units, and updates before they move on to Rosetta@home, runs on BOINC as well.^[67]

Volunteer contributions

Rosetta@home depends on computing power donated by individual project members for its research. As of July 29, 2008, over 47,000 users from 165 countries were members of Rosetta@home, together contributing idle processor time from over 85,000 computers for a combined performance of over 64 gigaFLOPS.^[68]

Bar chart showing cumulative credit per day for Rosetta@home over a 60-day period, indicating its computational power during the CASP 8 experiment

Rosetta@home users who predict protein structures submitted for the CASP experiment are acknowledged in scientific publications regarding their results.^[72] Users who predict the lowest energy structure for a given workunit are featured on the Rosetta@home homepage as 'Predictor of the Day', along with any team of which they are a member.^[73] A 'User of the Day' is chosen at random each day to be on the homepage as well from users who have made a Rosetta@home profile.^[74]