dear mr. baker,

i am one of the hordes of Fad members searching for a new distributed computing project to run once Fad shuts down dec. 16th. so far i have installed rosetta on 2 machines and haven't been able to download any work. i understand you're having problems with your server, but to those of us searching for a new project to support, this doesn't come off too well. to that end i would like to extend an invitation to you to come over to our home bulletin board and talk about your project. would you be willing? if so our address is

http://www.xtremepccentral.com/forums/index.php? if so please us this thread for your reply's.

http://www.xtremepccentral.com/forums/showthread.php?t=21274

we are anxious to know if you can support a mass influx of new people and machines without the stoppages that we see right now.

thank you,

Di
vavega

I took your suggestion, and just posted an introduction on your message board. how many FAD participants are there?

---

The landscapes on which we are searching for the global minimum are very large--we are trying to find a small needle in a very large haystack. My guess is that if we could increase our computational throughput by a factor of ten, many of the search problems would be solved.

David, I am not sure the 'needle in a haystack' example describes very well what we are up against. Haystacks tend to be three-dimensional whereas our problem has several hundred dimensions. Let's assume we are searching a 500 dimensional space and the rmsd value is something like a distance in that space. So if we want to improve our search result from rmsd=2 to rmsd less than 1 we need to find the rmsd<1 volume (the needle) within the rmsd<2 volume (the haystack). Now, since the search space is 500 dimensional the rmsd<1 volume is just 1/2^500 = 1/3e150 the size of the rmsd<2 volume. So assuming the volume of a needle is 1 mm^3 that makes our figurative haystack much larger than the size of the Universe, or to put this differently, we are trying to find a volume smaller than one cubic Angstrom within the size of the Universe.

I guess the point I want to make is that there does not seem to be any hope to find the rmsd<1 volume by randomly searching the rmds<2 volume with any kind of computing power (I hope I don't sound pessimistic ;-) ). The only option I can think of seems to be an algorithm that is able to estimate the direction towards the minimum energy point by evaluating the local properties of the search volume.

To transform this back into your earth explorers example, it seems to me that it is not enough to randomly send more explorers into an area which previous explorers already found to be at a relatively low elevation (this would be similar to randomly searching the rmsd<2 volume in the hope to find the rmsd<1 volume) but the explorers somehow need to be able to at least estimate the direction towards the lowest elevation from observations which they can make at their current location. Simply following the local downward slope may not necessarily be the best strategy (this just gets us into local minima). Perhaps there are other more useful local properties of the search volume like taking local averages of the elevation/energy landscape, looking at higher derivatives, something other than energy altogether ...

Well, sorry David, time to shut up. I guess there is't much we lay people can suggest which the experts haven't already thought of. Anyway, I am really looking forward to see how this project gets along and I definitely appreciate your occasional updates in the forum.

-Hermann

Hi Hermann, you are absolutely right--a needle in the haystack problem in 500 dimensions is intractable because the volume of phase space close to the needle becomes vanishingly small. as you point out, part of the explanation for why so few people are finding points less than 2A rmsd from the native structure in the results plots is that vastly fewer structures exist between 1 and 2A than between 2 and 3A rmsd.
however, things are not as bad as they might sound. while there can be more than 500 degrees of freedom in the problem, because of the large number of constraints (atoms can't be on top of each other, etc) the effective dimensionality of the space being searched is very much smaller.
for the proteins whose landscapes you have been collectively mapping out on this project, it appears that the low rmsd and low energy structures only differ from the native structure in a small number of critical features. for example, in the 1PVA case in the results section, the key feature is the bent helix, which I discussed in the "what we have learned thread" earlier.

---

...how many FAD participants are there?

Thousands. Many of us run many machines. At one point I was running 25 or so, myself. So far, I have 6 on Rosetta, 3 of them duallies. Please understand that there is many more folks coming from FaD. And those who have migrated already are probably not running full on, but ramping up.

You guys sound like you are quickly getting ready for an fast influx . Good. Its' coming.

Cheers,
Scott

---

I am also one of the Fad people who are comming over

on Fad i had anywere between 4 and 8 computers running 24/7 so far i have switched over 1.

Alot of us are staying at Fad till the last days... I will probaly switch before then because it is likely that I will be out to see at that time (US NAVY) but it this project looks good to me so i will probaly have 4 computers running this on boinc as wel as a few other projects.


Nasher

---