here (I think it will show my results but don't let that put you off).

There are quite a few RMSD's of 0.01 yet a range of energy of about -40ish to -160ish.

So, as I understand it, even when the backbone is very close to the right structure the packing of the sidechains is causing the energy fluctuations.

It raises the questions, how close to the RMSD do you have to be to be medically significant? Or is RMSD not really a good indicator of this at all? Is energy (medically) any better?

Like I said interesting results.

The WU's you speak of are the ones that ran slow and gave low scores under 5.34. I think that these WU's will be resent this week as the results looked so promising. Yes those are my lowest RMSD's Ever.

As far as the importance of decoy to RMSD deviation, the energy fluctuations do indicate more of a side chain deviation with the backbone so close.

The project goal is to perfect the program in ten years, However Dr. Baker said that mass clustering around both RMSD and lowest possible energies is not absolutely required for useful predictions.

Hope this helps.

---

The question has been asked before, but it was long enough ago that I'm going to have trouble digging up the link. I believe the results was that RSMD in the 1.0 range is "close enough" in many cases for pharmacuticals.

In you think about it, a drug is targeted at one specific portion of the protein (the portion that's exposed on the outside). And so if your prediction on the other side of the protein is where you're off, you may still be successful with creating a drug. And, in fact, may still have a perfect prediction for the "southern hemisphere" if you will, of the protein.

---

Add this signature to your EMail:
Running Microsoft's "System Idle Process" will never help cure cancer, AIDS nor Alzheimer's. But running Rosetta@home just might!
https://boinc.bakerlab.org/rosetta/

Even when we are getting RMSDs of 0.1, these predictions don't always include the lowest energy prediction - graphically, the lowest energy is sometimes way off to the right somewhere. Without an experimentally determined native structure, there is no way to know the RMSD of the lowest energy prediction.

It seems then, that if "mass clustering around both RMSD and lowest possible energies is not absolutely required for useful predictions," then clustering near 0 RMSD doesn't really mean much. I'm guessing (and please correct me if I'm wrong) that what is being looked for is a "clustering" of similar structures at low energy levels. A few structures with 0.1 RMSD doesn't count for much because we won't know that their RMSD is so good; instead, a group of the lowest energy predictions with very similar structures (and slightly higher RMSD) are a better indication that the predictive methods have been successful.

---

Clicking though the new results plots and from what can be learned from the WU names I get the impression that a lot of interesting things are going on with the current round of WUs, though this is of course all just guess work. I wonder whether any of this would be suitable for some slightly more detailed explanations, perhaps in David Baker's Rosetta@home journal, than we have been given so far.

Another, lets call it a dream, I was having lately would be some kind of 'experts´ corner', somewhere on the website, with more in depth material for those with some background or interest in science to dig their teeth in. I was thinking in terms of something like a repository of powerpoint presentations from internal seminars, students papers/thesis, conference presentations or posters (those that are not already linked in the publications section on the Baker Lab website), etc. And perhaps it would even be useful for the Baker Lab team members to have these things handy on the Web. Also, I understand that some kind of licence agreement is required to get access to the Rosetta source code, but perhaps part of the software documentation could be placed freely on the Web ? Or, another idea, are there perhaps some internal log files to keep track of all those different WUs that could be made available ?

I guess it needs to be understood that the Baker Lab team members would not have the time to answer all sorts of specific questions related to this material but I am guessing there must be lots of Rosetta@home participants out there who could learn a lot from this kind of material. Well, I thought I just mention this - of course I am not sure this is really feasible and could be done without investing too much time... -H. ;-)

---

Team betterhumans.com - discuss and celebrate the future - hoelder1in.org

I've been worrying about RMSD and energy for a while, because I think it is
possible to add a type of RMSD into a real prediction (which of course you
don't have a known structure to test RMSD against), it's not perfect and
may be impractical but let me explain.

If you save, the backbone of each of the predicted structures as well as
the energy, you could relate them to each of the other predicted backbones.

If you now took this, and statisticly fit these into the patterns known
from the RMSD information already obtained from known structure runs
(including heuristics and homolgues differences), you should be able to get
a RMSD type view.

I'm not sure if what I am saying is clear, so let me give an example:

this has a unique shape to its data distrobution, as does this

and so forth...

So if you got back say 10,000 results with 3000 different backbones, saved
each of these 3000 backbones with a frequency returned (and energy), then
went to your known runs and fit the new data with the old (known) data you
should in theory be able to tell (depending on which approach etc...) even
how far you are from the correct RMSD.

I hope this makes some kind of sense, as I'm not good at expaining my thoughts.

So if you got back say 10,000 results with 3000 different backbones, saved
each of these 3000 backbones with a frequency returned (and energy), then
went to your known runs and fit the new data with the old (known) data you
should in theory be able to tell (depending on which approach etc...) even
how far you are from the correct RMSD.

I hope this makes some kind of sense, as I'm not good at expaining my thoughts.

It makes sense but only if you really trust your methods. It seems like there will always be the chance that it actually looks some thing like this.

Of course, you should be able to tell with that one because there would be too much variation among the structures of the many predictions near the bottom of the energy range. On the other hand, you might have one outlier in a set like the ones you point to that would mess things up.

I think the best indication would bbe some sort of RMSD calculation between the lowest energy predictions. If all or most of the lowest energy predictions are starting to look pretty much the same, then you have something.

---

I'm a little sorry to say this, but the very low RMSD workunits of 0.1 are due to a bug in how the RMSD values were read out from the reported structures into the database that makes the cool graphs. That same issue leads to the sharp vertical lines at 1.0, 2.0, and 3.0. I've fixed this in the new apps being tested on ralph. My apologies for the confusion!

An interesting scientific point, though, is that we never expect to get better than about ~0.5 A from the "right" answer, because the true structures of proteins have thermal fluctuations at about this level. So if you experimentally solve the structure of a protein under slightly different conditions, in different crystal forms, or with different techniques (e.g., nuclear magnetic resonance spectroscopy vs. crystallography) you'll get slightly different answers. Our goal, of course, is to predict structures within this kind of precision (0.5 A); as pointed out below, structures with accuracies of better than ~2.0 A would still be extremely valuable for pharmaceutical applications.

Even when we are getting RMSDs of 0.1, these predictions don't always include the lowest energy prediction - graphically, the lowest energy is sometimes way off to the right somewhere. Without an experimentally determined native structure, there is no way to know the RMSD of the lowest energy prediction.

It seems then, that if "mass clustering around both RMSD and lowest possible energies is not absolutely required for useful predictions," then clustering near 0 RMSD doesn't really mean much. I'm guessing (and please correct me if I'm wrong) that what is being looked for is a "clustering" of similar structures at low energy levels. A few structures with 0.1 RMSD doesn't count for much because we won't know that their RMSD is so good; instead, a group of the lowest energy predictions with very similar structures (and slightly higher RMSD) are a better indication that the predictive methods have been successful.

---