Folding@home

Guest post from Dr. Gregory Bowman, UC Berkeley

Two general objectives of the Folding@home project are (1) to explain the molecular origins of existing experimental data and (2) to provide new insights that will inspire the next generation of cutting edge experiments.  We have made tremendous progress in both areas, but particularly in the first area.  Obtaining new insight is even more of an art and, therefore, less automatable. 

To help facilitate new insights, I recently developed a Bayesian algorithm for coarse-graining our models.  To explain, when we are studying some process—like the folding of a particular protein—we typically start by drawing on the computing resources you share with us to run extensive simulations of the process.  Next, we build a Markov model from this data.  As I’ve explained previously, these models are something like maps of the conformational space a protein explores.  Specifically, they enumerate conformations the protein can adopt, how likely the protein is to form each of these structures, and how long it takes to morph from one structure to another.  Typically, our initial models have tens of thousands of parameters and are capable of capturing fine details of the process at hand.  Such models are superb for making a connection with experiments because we can capture all the little details that contribute to particular experimental observations.  However, they are extremely hard to understand.  Therefore, it is to our advantage to coarse-grain them.  That is, we attempt to build a model with very few parameters that is as close as possible to the original, complicated model.  If done properly, the new model can capture the essence of the phenomenon in a way that is easier for us to wrap our minds around.  Based on the understanding this new model provides, we can start to generate new hypotheses and then test them with our more complicated models and, ultimately, via experiment.

Statistical uncertainty is a major hurdle in performing this sort of coarse-graining.  For example, if we observe 100 transitions between a pair of conformations and each of these transitions is slow, then we can be pretty sure this is really a slow transition.  However, if we only observe another transition once and it happens to occur slowly, who knows?  It could be that it is really a slow transition.  On the other hand, it could be we just got unlucky. 

Existing methods for coarse-graining our Markov models assume we have enough data to accurately describe each transition.  Therefore, they often pick up these poorly characterized transitions as being important (for protein folding, we typically care most about the slow steps, so slow and important are synonymous).  The new method I’ve developed (described here) explicitly takes into account how many times a transition was observed.  Therefore, it can appropriately place emphasis on the transitions we observed enough times to trust while disregarding the transitions we don’t trust.  To accomplish this, I draw on Bayesian statistics.  I can’t do this subject justice here, but if you’re ever trying to make sense of data that you have varying degrees of faith in, I highly recommend you look into Bayesian statistics.

Guest post from Dr. Gregory Bowman, UC Berkeley

We’ve been making a lot of progress with developing Markov state model (MSM) methods for analyzing the data we generate with the help of the FAH community.  For those of you with a theory background, MSMs are just discrete-time master equation models.  For everyone else, MSMs are a way of describing the conformational space a protein (or other biomolecule for that matter) explores as a set of states (i.e. distinct structures) and the transition rates between them.  Much of the theory underlying these methods is quite old but their use has been limited by the challenges inherent to identifying a reasonable set of states. 

During my time in the Pande lab, I worked with Xuhui Huang (now at the Hong Kong University of Science and Technology) to develop new methods for building MSMs from the large data sets we generate with FAH.  Together, we started an open source software package called MSMBuilder (here) to automate the process of building MSMs.  Now a number of more recent additions to the Pande lab are helping Xuhui, Vijay, and me in continuing to develop the software.

As we just released an update to MSMBuilder, I was looking back at some of our user statistics and was pleased to see how quickly our project is gaining traction.   Since its initial release in 2009, there have been over 1,600 unique downloads of MSMBuilder.  One cute feature of our webpage—provided by the SimTk software consortium at Stanford—is that you can go look where all of our users are (here).  Its fun to see that MSMBuilder is being used on 5 continents.  Maybe most importantly, MSMBuilder has been used in at least 40 publications to date.  MSMBuilder is coming up at conferences with increasing frequency too, so I look forward to reporting back on our growth in another year or so.

Guest post from Dr. Gregory Bowman, UC Berkeley

Most rational drug design efforts assume the target protein exists in a single structure and that the structure of one region of the protein--called the active site--allows the protein to perform some function.  Once this assumption is made, the only way to manipulate a protein’s activity is with inhibitors that bind the active site tightly enough to block it from performing its intended function.  Unfortunately, this strategy only works for  ~15% of proteins, greatly limiting the number of proteins we can manipulate for therapeutic purposes.

In a recent article published in the Proceedings of the National Academy of Sciences (link), I showed that simulations run on Folding@home can reveal new ways of manipulating a protein's activity.  Specifically, I start off by recognizing that proteins are actually flexible and then use Folding@home to enumerate the different conformations a protein adopts.  I then use statistical analysis to find parts of the protein that can communicate with the active site through a process called allostery.  These regions--called allosteric sites--are attractive drug targets as the binding of small molecules to them can be communicated to the active site, ultimately affecting activity.

As a proof of principle, I showed that my approach can identify a known allosteric site in Beta-lactamase (see figure below).  This protein is an important drug target because it can confer bacteria with antibiotic resistance by breaking down antibiotics like penicillin.  I also use my approach to predict new allosteric sites in Beta-lactamase and two other proteins that play important roles in immune deficiencies and HIV.  Now I'm performing experiments to test my predictions.  It will require a lot more of your WUs, but I hope this type of approach can eventually lead to new pharmaceuticals.

On the left is a structure of Beta-lactamase that most people would think of as the structure of this protein.  However, the right shows a different structure with a drug (cyan) bound in a pocket that isn’t visible in the structure on the left.  Binding of this drug somehow affects the structure near the active site (green).  Using my approach, I’m able to start with the structure on the left and then predict the existence of the structure on the right and the allosteric site the drug is bound to.

A brief update to our previous blog post on the A4 bonus: the bonus is now in effect.

We've noticed a significant number of high priority projects are trailing behind existing projects. Newer projects are aimed at interpreting and guiding experiments where the full power of Folding@home (F@h) is essential to continue pushing the boundaries of scientific and medical discoveries.

The main cause of this issue is the core version needed to run these simulations. Many of our newer SMP projects use the A4 core, which has several scientific advancements, while existing projects use the still important A3 core. The A4 core is not compatible with Clients below version 6.34 and many donors are still running these older Client versions.

This presents an opportunity to encourage people to donate their cycles towards these vital A4 projects. To emphasize the scientific importance of these work units, we are boosting the base points of all A4 work units by 10% when uploaded (Note that this bonus will not be reported by V7 or by the 3rd party applications which project PPD but the points will appear when your statistics are credited). The quick return bonuses will be calculated on top of the increased base points. This will start on Monday July 23, 2012, and we will keep this 10% bonus in effect for at least 3 months as a trial period, but plan to keep it longer, as needed.

To participate, donors should be running a recent version of the F@h Client. We strongly encourage Windows users to update to the much improved V7 Client. Although F@h Client v6.34 or newer is sufficient to participate for any supported operating system. Please note the Linux and OSX V7 Clients are a work in progress and feedback is welcomed.

v7 Clients:
Windows/Linux: Visit our home page, http://folding.stanford.edu/English/HomePage
Mac OSX: v7 for OSX is still in testing. For a beta copy: https://fah-web.stanford.edu/projects/FAHClient/wiki/BetaRelease

Old v6.34+ Clients
Windows/Linux: http://folding.stanford.edu/English/DownloadWinOther  

Guest post from Dr. Xuhui Huang, Hong Kong University of Science and Technology

In this post, I want to introduce a new GPU-powered clustering algorithm we recently developed to analyze the large molecular dynamics simulation datasets generated by Folding@home. Folding@home can generate enormous sets of protein structures. A critical step in analyzing these large datasets involves some form of reduction in the dataset, usually in the form of clustering. We recently developed a GPU powered clustering algorithm using the intrinsic properties of a metric space to rapidly accelerate the clustering. Overall, our algorithm is up to two orders of magnitude faster than the CPU implementation, and holds even more promise with the ever increasing performance in GPU hardware.

This algorithm should facilitate numerous applications. For example, one of the systems we tested our code on is the human islet amyloid polypeptide (hIAPP) peptide, whose aggregation is implicated in Type 2 diabetes.  We hope further analysis of this data will provide insights that will inform the development of treatments for diabetes.

Guest post from Dr. Vincent Voelz, Temple University

Using protein folding simulations alongside experiments remains challenging because the two techniques often "see" very different things. Simulation trajectories "see" every atom in a single protein in microscopic detail, while experiments often "see" only bulk properties averaged over large ensembles of molecules. For example, in the last few years, we have built kinetic network models of ever larger and slower-folding proteins. These models can have huge numbers of states and many possible folding pathways, yet experimental folding kinetics can be fit to models having only two or three states.

In a new paper, we try to bridge these two levels of detail using a combination of simulation and experiment to study the early folding events of ACBP, a 86-residue helix-bundle protein that folds on the ~10 millisecond timescale, one of the largest, slowest-folding proteins we have studied to date. Previous experiments suggested that ACBP folds via a "three-state" mechanism, with an intermediate forming on the ~100 µs timescale. To understand the molecular events underlying the formation of this intermediate, we used Folding@Home to generate tens of thousands of GPU-accelerated trajectories, and stitched these together to build a kinetic network model of the complete folding reaction (see figure below). By comparing our model to the results of state-of-the-art experiments (single-molecule FRET, Trp-Cys quenching, and time-resolved FRET) we found something surprising -- the folding relaxation timescale around ~100 µs corresponds to the heterogeneous formation of unfolded-state structure, rather than some discrete structural state.

This work is exciting because it shows that our models can predict atomically detailed mechanistic information about folding (currently very difficult to obtain experimentally) while simultaneously providing accurate predictions of quantities seen in bulk folding experiments.

Guest post from Dr. John Chodera, UC Berkeley

Kinases [http://en.wikipedia.org/wiki/Kinase] are the molecular logic gates of the cell.  These important proteins integrate critical signaling information in every cell of our bodies, becoming active only when specific upstream signals are received.  However, in many kinds of cancer, mutations can emerge in one or more kinases that cause them to ignore these regulatory signals and become active all the time.  If these kinases are involved in cell division, this can erroneously cause cells to keep dividing even when they shouldn't, potentially resulting in a form of cancer.  

Guest post: Dr. Greg Bowman, UC Berkeley

We just had a protein folding conference at Stony Brook University in New York that was extremely encouraging.  Both the experimental and theoretical communities are very excited about the results we are generating with Folding@home.  In particular, they are excited about (i) our increasing ability to make quantitative connections with experiments and (ii) the long timescale dynamics for large proteins we are now able to capture.  For example, we recently succeeded in folding an 80-residue protein on 10 millisecond timescales (paper is here).  For reference, that’s about twice as many residues and about 1,000 times longer timescales than what most anybody else is able to achieve!    There are now multiple experimental groups who are asking us to make predictions for them to test.  So, we appreciate all your help and have plenty of new calculations for you to contribute to.

To start off FAHcon2012, I gave a talk which included a review of how far Folding@home has come in the last decade.  I showed a slide from the very first talk I gave about Folding@home results.  That talk was given at Columbia University in August of 2000, and I talked about results from our paper in Science entitled "Screen savers of the world, unite!".  That work described the folding of a very small protein (16 amino acids) on a very short timescale (10ns = 10 x 10^-9 seconds!), but still was a major accomplishment for the time.

It's exciting to see how far we've come.  One way to think about it is in terms of how long of a time scale and length scale we can simulate for protein folding and protein misfolding diseases (such as Aß aggregation in Alzheimer's Disease):

Time scales: advancing roughly 1000x every 5 years

2000: 1 to 10ns  (Fs peptide)

2005: 1 to 10µs  (villin, Aß aggregation of 4 chains)

2010: 1 to 10ms  (NTL9, Lambda repressor)

2015: 1 to 10s?

Just breaking past a microsecond was a big deal.  The fact that we can simulate 10's of milliseconds is very exciting, but I'm really excited about where this appears to be leading, allowing us to tackle really challenging and important problems.  It would also mean that through a combination of new methods, algorithms, and hardware advances, we've already increased our capabilities by a million fold in just 10 years (2000 to 2010).  We're looking forward to hopefully making it a billion fold in 2015!

Length scales: advancing roughly 2x every 5 years

2000: 16 amino acids (Fs)

2005:  35 amino acids (villin)

2010:  80 amino acids (lambda, ACBP)

2015:  160 amino acids?

It's also important to note that these are sizes for protein folding.  For other problems, such as protein conformational change, we've already tackled much bigger systems.

I'm really excited to see what the next 5 years will bring!