Global Grid Computing at NYU:

The Human Proteome Folding Project

By Richard Bonneau

Grid computing can mean different things to different people. For the purposes of this article, we'll define grid computing as the use of the collective processing resources of a network of loosely coupled computers to solve large-scale computational problems. The computers in question are not primarily devoted to the grid: they can be anywhere in the world, can belong to anyone (so long as they choose to participate in the grid), and need only communicate with a central server for brief bursts each time a large part of the computational problem has been completed.

Figure 1. Overall workflow of the Human Proteome Folding Project. Steps carried out on the Grid appear in the block at the bottom left.

This type of grid strategy has resulted in several projects where people throughout the world can devote their computers' processing power to public research efforts during the times when their computers are not in use. One of the best known of these efforts is seti@home (http://seti.org/), a public project dedicated to finding signs of extraterrestrial activity in interstellar radio signals. If the input/output (I/O) needs of the computational problem are small and there is no need for careful timing, these grids can scale to include many hundreds of thousands of computers (as in the case of the project described below) to millions (as in the largest projects such as seti@home). Given that, on average, less than 10% of an off ice computer's daily processing capacity is used, these grid computing projects are tapping a valuable and underused global resource, enabling much larger scale projects than could otherwise be completed.

This article describes the Human Proteome Folding Project, a grid computing effort currently being carried out by NYU in collaboration with IBM (www.grid.org/projects/hpf/). We're using the computing power of millions of computers to predict the shapes of human proteins about which researchers currently know little. From these detailed shapes, we hope to learn about the functions of these proteins, as the shape of a protein is inherently related to how it functions in our bodies.

The resultant database of protein structures and putative functions will let scientists take the next steps towards understanding how diseases that involve these proteins work. We hope that our work on this project will contribute critical public infrastructure for the biological and biomedical community. Only with the amount of computing power available through the World Community Grid (described below) could we hope to complete this project. The scale of the Grid also allows for better sampling, enabling improved accuracy and gene coverage. I'll describe the project in greater detail below, and then go on to explain how you (via your computer) can participate.

PROTEINS

Proteins are the most important molecules in living beings. Just about everything in your body involves or is made out of proteins. They are structural molecules made up of long chains of smaller molecules called amino acids that act as enzymes and important carriers of biological signals. There are 20 amino acids that combine in different ways to make up proteins. As the chain of amino acids is built, the chain folds (like balling up string) into a more compact mass, ending up in a particular shape. This process is called "protein folding." Many of the things that happen in cells are specifically controlled by protein shapes and the functions conferred by those shapes. For example, a protein in a virus or bacterium may have a particular shape that interacts with human proteins or human cell membrane, enabling it to infect cells. This is an oversimplification, but nevertheless, knowing these shapes helps us gain insight into the biology of disease by understanding protein function.

How proteins fold has been a compelling theoretical and experimental research focus since the early 1950s, when Dr. Christian Anfinsen won the Nobel Prize for showing that proteins fold reproducibly. Most proteins fold to a tight ensemble of possible shapes. These ensembles are amazingly reproducible when compared to other polymers, the discovery of which spawned the new field of polymer physics called protein folding. Eventually, methods for folding proteins computationally came on the scene, one of the more successful of them being the Rosetta software program. Less than five years ago, we (the Rosetta developers community) reached a critical milestone, showing that, for small proteins, we can predict structure well enough to predict aspects of function. Ever since, we have been working hard to improve these methods of extracting function from structure prediction, but the computer power needed for these calculations has been a constant limiting factor.

AN EXPANDING PROTEIN UNIVERSE

In recent years, scientists have sequenced many genomes, including the human genome. Between 20,000 and 30,000 genes found in the human genome encode proteins. The collection of all the human proteins in the genome is known as "the human proteome." It is not, however, a trivial task to determine from the sequence of genes their final 3D shape. Protein structure can help us divide proteins into functional classes and thus aid in our organization and understanding of the protein universe. A protein's shape also gives us clues as to the protein's function, but there are tremendous barriers to getting structure for whole proteomes.

PROTEIN STRUCTURE PREDICTION

We've known for a while that we can get good enough structure predictions using Rosetta to predict function for large numbers of proteins. For example, as part of my PhD at the University of Washington (Baker Lab), I focused on improving the Rosetta method and then on predicting the structures of only 500 key proteins. I was terribly limited by available computer resources, however, and predicting the structure of just these 500 proteins made me somewhat unpopular, as my calculations monopolized the Baker Lab computer clusters.

What we'd ideally like to do in the Human Proteome Folding Project is attempt to predict the shape of all proteins of unknown function. Due to the massive scale of such a project, this is not possible using existing supercomputing centers or clusters here at NYU, or anywhere else that I know of. One of the features of the problem, however, is that computations on different proteins can be run completely separately, and, in fact, single proteins can be separated into hundreds of smaller independent jobs (a so-called "embarrassingly parallel problem")¹ Although the supercomputing resources here at NYU are impressive², they would not be the most cost-effective solution, since parallel problems like ours do not require shared memory or tightly coupled I/O. An attractive aspect of our problem is that data sizes can be kept very small, with large data volume steps separated from processing- intensive steps.

WORLD COMMUNITY GRID PILOT PROJECT

The World Community Grid was started by IBM as a public supercomputing resource; all results are made public, and research efforts are carried out on the non-IBM side by non-profit entities. In 2004, IBM was looking for a biological application they could use to help test their grid (also with the aim that work on the Grid would benefit humanity and help generate positive publicity). Given the inherent grid-ability of our project, we were a natural fit for the World Community Grid. We proposed the Human Proteome Folding Project, and were accepted as the World Community Grid's pilot project.

Initial work consisted of defining the project's scope by selecting which proteins were to be folded. In the initial phase of the Human Proteome Folding Project (HPF1), we folded 150,000 proteins of unknown function. In the current, second phase (HPF2), we're focusing on approximately 10,000 proteins (150 genomes) of high interest, including human secreted proteins, cancer biomarkers, and small secreted plasmodium proteins. Another task was to construct the Rosetta grid-client. This consisted of changing our code to match special grid-client library calls and testing the server-side pipeline (scripts for breaking work into smaller pieces and recollecting the finished work units). Rick Alther and Viktors Burstis were instrumental in the completion of this work. Robin Wilner and Bill Bovermann led the overall coordination of all the other aspects of the Grid beyond the science, such as the public relations efforts needed to persuade people to download the client.

The Workflow

The overall workflow of the Human Proteome Folding Project is shown in figure 1. The preprocessing steps consist of the sequence analysis needed for making the work units. We first exhaust all means of finding sequence matches (the most common way of annotating proteins), and only when we have confirmed that a protein, or a part of a protein, has no matches to known structures do we send it out to the Grid to be folded. This process helps us conserve resources; the Grid is large, but we still need to maximize the utility of the 500,000 CPU years³ it will give us.

These preprocessing steps are shown in the box at the top left of figure 1. The box at the lower left shows the most CPU-intensive part of the calculation, and the heart of the calculation performed on the Grid. The rightmost boxes depict our meta-database (structure, sequence, localization, and other annotations integrated for proteins in the 150 genomes analyzed) as well as our planned interface to this metadatabase.

The large data volumes of HPF1 and the voracious appetite of the Grid for work units posed a challenge, but we were able to complete HPF1 without any down time, any fake work units, and any data loss on the part of NYU or the Institute of Systems Biology (ISB) in Seattle, due in large part to heroic efforts by NYU, the ISB, and IBM team members.

Figure 2. The HPF, Phase 1 client. Upper panels show the structure and simulation scores, as well as progress. The lower bar provides tools for management of the client and its priorities on the volunteer computer.

THE RESULTS OF PHASE ONE

Work on HPF1 is drawing to a close, and results from HPF2 (the current phase, which I'll describe below) are beginning to arrive at NYU servers. We are now working around the clock—partly due to the fact that the project involves researchers around the globe—to process the results and get them out to the community in our protein domain annotation database.

Alas, predicting the structures is only the first challenge, and current research in my lab centers on integrating the structure-derived data with other sources of protein function (sequence-based, expression and localization measurements, etc.). The beta-release of the database and our release of the results on a model organism (yeast) are the center of our first paper. Soon after we have debugged the process using the yeast proteome, we'll release the rest of the 150 genomes folded during HPF1 via our meta-database.

HPF2:HIGH-RESOLUTION PROTEIN FOLDING

The second phase of the project (HPF2) is a refinement, using Rosetta in a mode that accounts for greater atomic detail, of the structures resulting from the first phase, HPF1. The project focuses on human secreted proteins, including proteins in the blood, mucus, tears, and the spaces between cells. These proteins can be important for signaling between cells and are often key markers for diagnosis; they have even been found to be useful as drugs, when synthesized and prescribed to people lacking the proteins. HPF2 also focuses on key secreted pathogenic proteins. This phase of the project dovetails with efforts at the Institute for Systems Biology to support predictive, preventive, and personalized medicine, under the assumption that these secreted proteins will be key elements of this medicine of the future.

Figure 3. The HPF, Phase 2 client. The panel to the right shows the protein as it folds. The left panels show three components of the score that Rosetta uses to judge protein structures and ultimately return the best-scoring structures to the central server.

This second phase of the Human Proteome Folding Project continues where the first left off. The two main objectives are to: 1) obtain higher resolution structures for specific human proteins and pathogen proteins, and 2) further explore the limits of protein structure prediction by further developing Rosetta structure prediction. Thus, the project addresses two very important parallel imperatives, one biological and one biophysical. With the second phase, we are aiming to increase the resolution of a select subset of human proteins. Better resolution is important for a number of applications, including, but not limited to, virtual screening of drug targets with docking procedures and protein design. The second phase of the project will also serve to improve our understanding of the physics of protein structure and help us to further develop our state-of-the-art program, Rosetta.⁴

Figure 4. The structure on the left is an example of a Rosetta prediction. The right structure is a structurally similar protein that provides, via structure-structure similarity between the predicted structure (left) and the known structure (right), some clues as to the function of PF0677. (Taken from Bonneau et al. 2001)

POTENTIAL FOR FUTURE PROTEIN GRID COMPUTING AT NYU

There are a number of ways we can increase the degree to which spare computer processing cycles are used. Making small private grids—NYU-wide or department-wide, for example—can provide more modest but still quite significant resources for many different analysis pipelines (preprocessing protein sequence for the larger grid, analyzing mass spectrometry data, turning cross links into structures, learning tissue-specific regulatory networks, etc). Another option is for an NYU-wide installation of the World Community Grid client, followed by an NYU-wide attachment to the HPF2 project, a much simpler option, in my opinion. To date, only a small fraction of current global CPU capacity is being used, and grid computing projects have lots of room to grow.

HOW TO PARTICIPATE IN THE GRID

Don't be left out of the satisfying fun that is grid computing! Anyone wishing to participate should visit www.worldcommunitygrid.org/ to download the free software. Thanks to IBM and United Devices, installing and managing the software client is secure and easy. We also have mechanisms for helping whole institutions push out installations; interested parties should contact IBM via the "About Us" form within the "Become a Partner" section of the World Community Grid website.

Once the United Devices or open source BOINC client is installed, whether on Windows, Macintosh or Linux, the next step is to attach your client to HPF2 by typing in the following project URL when prompted: www.worldcommunitygrid.org/projects_ showcase/viewHpf2Research.do. If you wish, you can also join a team by creating your own or selecting an existing one, the New York University team, for example. Another Grid project of note is working to dock small molecules into HIV protein structures in a search for possible new HIV drugs. I encourage anyone reading this article to convince multiple friends to download the client, as well!

For more information about the Human Proteome Folding Project, select the project name from the list in the "Research" section of the World Community Grid website: www.worldcommunitygrid.org.

Footnotes

1. Wikipedia, http://en.wikipedia.org/wiki/Embarrassingly_parallel
2. See the following Connect Magazine articles for more about the Max supercomputer: www.nyu.edu/its/pubs/connect/fall05/ackerman_supercomputer.html and www.nyu.edu/its/pubs/connect/spring06/allison_max.html
3. The CPU, or Central Processing Unit, is a computer component that interprets instructions and processes data in computer programs. CPU years are a unit of measurement for processing time.
4. We are members of Rosetta Commons, the Rosetta developers' community (www.rosettacommons.org/).

Author Biography

Dr. Richard Bonneau recently joined NYU's Departments of Biology and Computer Science as part of a joint initiative of Computation in Science and Society and NYU's Center for Comparative Functional Genomics. He has played a critical role in the development and deployment of Rosetta, the state-of-the-art protein-folding program described in this article. See http://cs.nyu.edu/~bonneau/ for more about Dr. Bonneau's research efforts.