Folding@home

Recent talk of results from Folding@home

Donors are often curious to hear about recent results.  While one can read our papers, listed on our web site, those are fairly technical and intended for a biological or biophysical audience.  The talk I gave at GTC 2010 ("Folding@home:  Petaflops on the cheap today, exaflops soon?") was intended for a computational audience, so it may be more approachable than the papers, especially for those more familiar with the computational side of FAH, rather than the biological.  You can see it here.

I talk a bit about how FAH works and some recent results in protein folding (pushing past the millisecond timescale), protein misfolding disease (Alzheimer's) and viral infection.  If you're curious about how FAH works or what we've done recently, this might be of interest to you.

Two new key results

We have two papers (one that just came out in PNAS and one that's about to come out in Physical Review Letters -- papers #74 and #75 on our papers list) that we're particularly excited about.  They represent some key results that we've learned by examining multiple results from Folding@home.  The resulting picture of how proteins fold is fairly different from the prevailing view, so it will be interesting to see what experiments tell us about the specific predictions made within.  We're excited to see where this goes!

Paper #72: Major new result from Folding@home: simulation of the millisecond timescale

 Simulating protein folding on the millisecond timescale has been a major challenge for many years.  When we started Folding@home, our first goal was to break the microsecond barrier.  This barrier is 1000x fold harder and represents a major step forward in molecular simulation.

Specifically, in a recent paper (http://pubs.acs.org/doi/abs/10.1021/ja9090353), Folding@home researchers Vincent Voelz, Greg Bowman, Kyle Beauchamp, and Vijay Pande have broken this barrier. The movie below is of one of the trajectories that folded (i.e. started unfolded and ended up in the folded state). From simulations like these, we have found some new surprises in how proteins fold. Please see the paper (url above) for more details.

Why is this important?  This is important since protein misfolding occurs on long timescales and this first simulation on the millisecond simulation for protein folding means we have demonstrated our new Markov State Model (MSM) technology can successfully simulate very long timescales.  It make sense to go after protein folding first, since there is a wealth of experimental data for us to test our simulations. 

While this paper on protein folding has just come out, we have already been using this MSM technology to study protein misfolding in Alzheimer's Disease, following up from our 2008 paper. While our previous paper was able to get to long enough timescales to see small molecular weight oligomers, this new methodology gives us hope to push further with our simulations of Alzheimer's, making more direct connections to larger, more complex Abeta oligomers than we were previously able to do.

Release of new Protomol (Core B4) WUs

As announced some time ago we have been working on a new core (Protomol core B4), and it has been looking good in QA so we have started to release Promol core WUs more broadly.  We have a preliminary Protomol FAQ for those who are curious to get more information.

This new core implements the NML (Normal Mode Langevin) method which accelerates the long-time dynamics of the proteins by a factor that can reach speeds up to a hundred times faster than those of molecular dynamics. This method searches for low frequency directives by using normal mode analysis and projects the motion of the molecule along them while resolving the nearly instantaneous motion. If you want to learn more about this method, you should read this pre-publication: Multiscale Dynamics of Macromolecules Using Normal Mode Langevin.

Based on Protomol 3.1, this core and its associated projects have the following goals:

• To validate NML by simulating the folding and dynamics of the Fip35 WW domain.
• To understand the role of mutations on folding.
• To understand the activation of src Kinase, an enzyme that is involved in the onset of some kinds of cancer.
  

For more information about the Protomol core, you should visit the Protomol official site.

The new projects are distributed by a new server (129.74.85.48) which is located at the University of Notre Dame (Indiana) and have the following characteristics:

• p10000 : 544 atoms, 84.48 points, preferred deadline 3.07 days, final deadline 23.04 days. This project uses conventional simulation methods.
• p10001 : 544 atoms, 50.56 points, preferred deadline 1.84 days, final deadline 13.79 days. This project uses the NML simulation method.

We will be posting more information as time goes on.   I'm very excited about the new capabilities here, since NML allows us to algorithmically get an amazing speed up, i.e. without any additional CPU power.    That algorithmic speed up multiplied by the vast power of FAH could mean very significant advances shortly, making 2010 an exciting year for FAH (in many ways)!

IMAG meeting today -- The Impact of Modeling on Biomedical Research

I'm at NIH today for the IMAG meeting on the Impact of Modeling on Biomedical Research.  IMAG stands for the Interagency Modeling and Analysis Group and it constitutes all of the science major funding agencies in the US (NIH, NSF, DOE, and more).  It's a great idea to get these agencies together and well-coordinated in terms of their approach to computer simulation and modeling.

In the meeting, they've broken down all of modeling into 5 scales, and I'm chairing the atomic and molecule scale.  Our charge is to talk about the future of modeling and I'll be highlighting some of the key new results from Folding@home.

For more information, please check out:

http://www.imagwiki.org/mediawiki/index.php?title=IFM_Announcement

FAH work highlighted in Biomedical Computing Review

Two key parts of FAH technology -- OpenMM (the software that powers FAH on GPUs) and MSMbuilder (the algorithms that FAH uses to stitch together hundreds of thousands of donor simulations to get coherent results) are highlighted in this months Biomedical Computing Review.  You can download a copy here

http://biomedicalcomputationreview.org/5/4/index.html

Importance of software and data integrity

The scientific code underlying our research has a lot of dials and switches.  It can be very tempting to play with those switches to optimize the calculations.  However, many of the code settings can impact scientific results in non-obvious ways.  It can also hurt scientific reproducibility.  It is therefore extremely important not to modify the scientific cores or data for Folding@Home calculations. Returning any work results with modified cores or data can taint both your results and the results of any WU's calculated on the basis of those.  Thus even well-meaning optimization can potentially hurt the scientific value of the project substantially.  This is why the EULA specifically prohibits such modifications.

When you donate to Folding@Home, we give you a number of "dials" at the client level that you can use to adjust your contributions.  Many donors find that tuning these "dials" and their machine configuration can optimize their contribution.  Please use only the controls provided by the client and the operating system; do not modify the cores or Folding@Home data in any way.  Doing so is the equivalent of providing a tainted donation.  When we detect such tainted donations, we may remove points awarded, decline further donations by blocking work assignment, or take further action.  We greatly appreciate your contributions to Folding@Home; please don't devalue your collective work by undermining the scientific results.

A new type of work unit

We are preparing for the public release of a new work unit category: extra-large advanced methods work units. Some background information is provided below; we'll update this thread with more details as they emerge.

Why a new work unit category?

We have some specific projects where we 1) have large simulation systems and 2) want to get results fast. As multi-core processors get more powerful, we can perform calculations on Folding@Home that

previously required supercomputing clusters.

What's different about these work units from the donor perspective?

These work units are special SMP work units that have larger upload and download sizes, shorter deadlines, and require more memory and CPU resources. That's why we've created a new category.

Is there any points incentive for running these work units?

The base value of these work units corresponds roughly to what an SMP work unit using the A2 core would yield on an equivalent calculation. However, because fast completion is a scientific priority for these work units, we are doing a **trial** of a new bonus scheme where faster WU completion yields a points bonus.

What systems can run these work units?

Right now, only Linux and OS/X systems can run these work units, and they require 8 or more cores. We prefer 8+ *physical* cores, although fast Core i7 machines that are dedicated folders have proven

sufficient during the testing process. The points incentives are designed to match appropriate resources to points value; if your machine is marginal for the extra-large work units, you're probably

better off running standard SMP.

Does this have any relation to the large-points value work units and recent high-scoring users?

Yes. The initial projects are 2681 and 2682, valued at ~25K points base. Although these point values seem high, the work units are correspondingly larger, so the base PPD (points per day) value is

roughly comparable to standard SMP.

A collaborator has donated a large amount of compute time to this project; those clients were initially running under username Anonymous/team 1. To give proper credit for the donation, we have

changed the username to PDC, team 1. During the period of this donation, there are at any time between 100 and 400 8-core clients running under this username (800-3200 cores total).

Please stay tuned for further details regarding the upcoming release.

MD Workshop

Next week, we will be having a Molecular Dynamics (MD) workshop (see http://simbios.stanford.edu/calendar.htm for details) for three main pieces of Folding@home-related software, namely OpenMM (the GPU acceleration code behind the FAH GPU client), MSMbuilder (the software used to parallelize calculations over all of Folding@home, i.e. how to make all the individual donors' calculations work together) and OpenMM/Zephyr (a tool intended to make molecular dynamics simulation easier for non-experts, built from the accelerated OpenMM codebase).

We're very excited about not just making the key software pieces that drive FAH available to other researchers, and moreover, we're working hard to teach others how they work and how to use them efficiently. 

You can learn more about these projects at their simtk.org web pages as well:

OpenMM:  http://simtk.org/home/openmm

MSMBuilder:  https://simtk.org/home/msmbuilder

Zephyr:  https://simtk.org/home/zephyr

Folding@home talk at PARC today, 4pm

Today (January 8, 2008) at 4pm, I'll be giving a talk at PARC, giving an update about Folding@home.  If anyone is local to the Palo Alto area, you can be there in person.  Note that they also do live streaming and should have a video on line (check out http://www.parc.com/forum for details).

UPDATE:  The talk can be found here:

http://www.parc.com/cms/get_article.php?id=799