Slashdot Mirror
Folding@Home Reports Success
msheppard writes "This Article describes how the folding@home distributed computing project is reporting that they used the data processed on client machines to "predict the folding rate and trajectory of the average molecule." Too bad Seti@Home hasn't had a hit yet."
Here.
MSNBC Article.
Folding@Home Home
For the real info though check out the Forums
Token link to how my team is doing.
PRIME1
http://www.kubuntu.org/
If you like F@H, check out Distributed Folding.
Look, ma! I'm a karma whore
This(postscript) is the the original paper on the hardness of String Folding problems.
I submitted this as a story, but it was rejected. Google has incorporated distributed computing into its toolbar as an option. The first supported project is Folding@Home, but they will add more projects in the future. Its optional, and currently has only been released to a few toolbar users. It will gradually be released to all users. Check it out at toolbar.google.com/dc/. Google is currently seventh in the team statistics...
Well, I kinda agree and I kinda disagree.
First, you can't expect to go from no success to complete success overnight. People have been trying to fold proteins for some time now and have basically failed because it is freakin' hard. The theory is in principle in place, a least to a first approximation, but the calculations are so intensive that they have basically beaten every comer. As an undergraduate I remember how everyone in the field thought getting bigger and better grants and buying bigger and bigger computers was the answer. Oh to be SGI in those days. They sum up the problem pretty well in the Nature paper, essentially a modern (desktop) computer would require a few decades to crunch through a single useful length simulation. Then you need to do it many times to get a useful answer (say 100-1000). Even supercomputers are going to balk at that kind of calculation. Moore's law what it is, we should then be able to get through an in silico simulation in a week on a single computer (when its this fast crystallography really will be dead) by, oh say 2040 at best. (somebody want to calculate that exactly, 10000yrs -> 0.02yrs is how many doubles). So yes, this hasn't gotten rid of x-ray crystallography just yet.
But this is still really cool. Complaints about interface and maintenance aside, this was a great system. It relied on four pretty bright insights.
First, that distributed computing is essentially the poor man's (cough, the academic's) super computer. Also, it automatically adapts itself to technological improvements. People will buy new computers from time to time and, hopefully, reload your software.
Second, that there was no reason other than no one had sufficiently brute forced the process that the existing methods shouldn't work. They use a bunch of 'cheating' techniques to make this managable during the screen saver timescale, such as a united atom model (I think that means they ignore aliphatic hydrogens) and implicit solvent (don't treat it as individual solvent molecules, just a uniform field). It was an open question as to whether this approach would work at all or if you had to go over to much more explicit methods to get it to work at all. It appears that this has kinda worked with the cheater methods in place.
Third, choice of a test case. Yes they chose something that was small. This isn't surprising. They wanted to be done sometime this decade, remember there is a graduate student as the primary author here. Small was necessary. However they also chose a FAST-FOLDING protein. That was clever. Basically, even with distributed computing, it is still hard to simulate a full microsecond of time. Thus they chose something that had some chance of completing its folding one the time scale that they could look at.
Fourth, they remembered their P-Chem. It is really hard to run these things to completion... so they didn't. You don't have to run the simulation until 99% of the molecules have completely folded, just until an appreciable number have folded and you can extrapolate the behavior from that. They ran a 20ns simulation (at the longest). The thing takes 7us for ~60% to fold. As a result only once in a great ong while did one of the simulations actually produce a folded protein. But by doing it ~10000 times they could figure out how that translated into the rate constant. That's clever.
That said, yes there is a long way to go on this, but its still a really clever paper. No we haven't cured cure cancer yet, but its still progress. And forget an in silico structure of the ATPase, that's largely understood already (check the RSCB/PDB there's a bunch). The real challenge will be getting a structure that size that hasn't been solved by other methods and convincing anyone else that you're right! Disclosure- I don't have PhD in this area yet, but I'm close.
When they can predict the structure of the F1F0 ATPase, then we can throw out crystallography- but it's not going to happen.
(Ignoring for the moment that crystallography has it's own issues. . . at least it can show active sites and quaternary structure)
Well, for the first, we can't throw out crystallography even then. When you're doing a computer calculation, you are in the realm of theory. (even if you have arbitrary accuracy).
You will still need to do experimental verifications now and then.
At the moment, about 2/3 of known protein structures have been mapped through X-ray crystallography. At best the resolutions are about 1.8 Å, which is pretty good. So you can see quite a bit more than quaternary structure!
The other third is done with NMR spectroscopy,
usually with some powerful computing help to figure it out.
And then there are a pitiful few,
done with computers and experimental data.
These structures also have the poorest accuracy.
Note that computers will never, ever be able to figure out a protein structre ab initio. (i.e. without any info except the sequence)
Do the math, say you have 100 amino acids, and you
test say, 4 conformations for each, that's 4^100
combinations to test.. and you test 10 million a second, it'll take you 5E45 years.
Much older than the current universe.
(Disclaimer: I do not -yet- have my PhD in computational biochemistry.. but I'm working on it..)
Actually, we're in the experimental error. Keep in mind that folding time distributions are exponentially distributed (not Gaussian). This means that the std devs will be big just by their nature. 7.5 vs 6 are indistinguishable statistically.