New set of questions.

1) is is possible we are attempting to "smooth" things too early?

In the first stage we spend a few moments looking through a space of conformations. One of those is then selected and we then spend the bulk of our time trying smaller adjustments to the structures. Yet, my experience so far does not lead me to believe that this mechanism is tending toward a result.

2) Is this really the most effective search?

In the earlier Top Predictions, we saw that the searches did not seem to cover the space. We had a couple threads discussing this with to my mind no conclusion why the coverage was not over more of the space.

Again, I am not seeing as effective of adjustments as I would expect.

3) What I am asking now, is, what is our coverage like now?

Might we not be better served by doing more of the high activity searching looking for the "corner"; returning the best "n" candidates and repeating. From these, select the "best" and try from there.

4) The processing if I am following it correctly is a two stage process. Yet it seems to me that the first stage does select some good starting points, but the relax does not do enough. In particular I am not seeing struture "rotations". Perhaps we need 3 stages?

In the example images 2a and 2b the basic starting point is good. Yet, topologically the structures could have been more closely related to the target if portions were rotated 180 degrees.

In 2b, for example, the red to brown segments and the green to blue were close, but, needed to be rotated up turned 180 degrees and laid back down. What I am suggesting is a multi-pass. If we pre-process as suggested above, then in the second stages good candidates like this would be subjected to a process intermediate in, ahem, violence. With the first stage we want to cover as much random territory as we can.

In the second stage we would attempt rotations and adjustments that would attempt to preserve the fine structure detail, yet improve the posture.

From there refinements using the current relax.

I doubt I am being clear again, but, the first stage I have the feeling we are not doing enough of. The second stage we are doing too much of and it is trying to give a 5 digit accuarcy of position for a raft shooting rapids. It is not seeming to bridge the gap.

Ok, let me have it ... :)

---

Might we not be better served by doing more of the high activity searching looking for the "corner";

So perhaps this post by DB:

Another possibility we have been considering is to generate a uniformly spread out set of ab intio structures using our in house machines, and then send them out to BOINC for relaxing. The advantage is that communication is easy on our local clusters. The disadvantage is that it will take a large amount of time to do the ab initio runs locally, even though they are individually very quick, and that the movies would not be as much fun to watch since most things happen in the ab initio part.

A compromise would be to carry out a first set of runs using BOINC as you are all doing, and then develop a shorthand description of the oversampled and undersampled regions of the landscape based on these results. Then, in a second set of runs, this information would be sent out to all processors and used in your steps 4-5. We already tried an early version of this approach in the 1n0u bab runs, and it looked pretty promising.

from the "CPU's talking" thread addresses your concern that we don't search the "corners" enough ? Generally, I have the impression that better covering the parameter space is high on the agenda of the Rosetta team.

I have been doing some thinking about this, and I'm afraid I'm not at all familiar with proteins or their folding. However, I have some experience desigining and optimizing optical lens systems, and the optimization seems very similar.

I think the problem may be that you can only go so far in the relax stage until you hit a local minimum. The computer decides that it can't go any further and stops. For algorithms as complicated as this, there are usually a TON of local minimums near the true minimum, and the real challenge is to find your way out of the higher valleys and into the best regions.

I'd suggest a series of pseudo-random changes AFTER the relax stage, but instead of randomly changing any joint in any direction, change every joint in, say, three or four opposing directions.

If it's anything like lens design (and I'm not a bit certain it is) then you might find that the lowest of THOSE energy states actually gives a lower global minimum after a further relax cycle.

Of course in lens design, we use algorithms that find the slope of the optimizing function rather than just trying different similar orientations at random. But that shouldn't matter TOO much. My intuition is that if you can jump out of the local minimum (that the current relax finds quite well) then you have a decent chance of finding an even BETTER minimum.

How to do that in a reasonably sized WU... I'll leave that up to you.

---

Ok, let me have it ... :)

still processing...

[still processing...

Jack,

Cool :)

The bad news is that with ability to work limited it gives me more time to think about this stuff ...

I kinda like Deamiter's suggestion. That type of action might allow the "rotation" effect I have not seen happen. Again, though, I suspect that we are hamstrung a little by the 2D of 3D ...

---

Thinking about it some more, I should mention that I have very poor intuition about HOW to jump out of local minimums after the first relax stage. I don't understand the rules for protein folding, much less which rules you're using in your program. I do believe, though, that you would see a very significant improvement in the end results if you ran a randomize script that was designed to push the protein further than the normal relax.

How far to push each connection, and which connections to push (in sequences etc...) I have no idea. Since you're not using minimization functions (but random testing) I'm not sure how close you get to the true minimum anyway...

Just my musings really, but in lens design, we're taught (along with picking the right starting point -- something not feasable here) to spend some time "jinking" the system after we hit a minimum to see if the computer can come up with a better solution. Since you're already effectively doing that, perhaps a stage with smaller random adjustment followed by a stage with larger random adjustment, and then a third smaller random adjustment.

Ideally, the WU would run until the sequence found no further candidates for the smaller relax stage... But then again, IDEALLY, each WU would be run on an uber-computer at 10WUs/sec!

---