A protein is made on the ribosome by sequential addition of amino acids starting at the free amine end. So it would seem that this end (the N-terminus) will be able to start folding before the rest of the protein, perhaps into a locally stable configuration that might be quite different from that which would be obtained from folding starting with the entire unfolded protein chain. Is this a relevant consideration for Rosetta? I can't find anything on this topic in these forums or on those at F@H.

Wow! I've had the question for some time now and never had a clue how to ask it with the proper terms.

For us lay-people, proteins are produced one amino acid at a time, and so it makes sense that perhaps looking at the entire protein to determine how it will fold is a more complicated problem then the one you really want to solve.

If you could devise a way of predicting the shape as it is being produced, then it would seem to simplify the problem. If the chemical cursors that produce the protein would be physically in the way at a given point in the protein's production, then that's probably not the correct shape.

And how would the hydrophobic and hydrophilic properties come in to play at this point in the protein's life? The team has explained in the past that portions of the protein tend to be attracted to water ("hydrophilic"), and so Rosetta biases towards placing these on the exterior of the shape, as this will be where water is nearby. And tends to be where these sequences are found in the native structure.

And so this raises another question, is the external influence of water molecules similar in magnitude as the protein is being produced, compared to once it's production is completed?

---

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/

And while we're here, I've also been curious how the end which is the "N-terminus" is determined.

---

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/

I cannot comment on how the folding commences, but I'll try an answer to the question about the N-terminus.

The genetic code is read from the DNA by enzymes that generate a messenger RNA, or m-RNA, which carries the information about the coded amino acids out of the cell core. This m-RNA carries special sequences which indicate from which end it must be read.

At another place other RNA, the t-RNA or transfer RNA, is loaded with amino acids. Each t-RNA carries a triplet of bases which fits to a certain triplet in the m-RNA. These triplets are called "codon" and "anticodon". Each amino acid is bound to the t-RNA with its COOH group and the amino group is kind of "dangling" free. The amino group (NH2) of the first amino acid of a protein will later be unbound, so it forms the N-terminus. On the other end the last COOH group will be unbound, forming the C-terminus. This is how it happens:

Now the ribosomes (kind of the protein factories of a cell) come into play: They find the starting sequence on the m-RNA and thread it into the protein synthesis mechanism in the right direction. When reaching the first codon that represents an amino acid, a t-RNA with the proper anticodon is attached to the m-RNA.

The m-RNA is now advanced a little and the second t-RNA is attached. Now the bond between the COOH group of the first amino acid and its t-RNA is broken and the COOH group forms a bond to the amino group of the second one. As you see, the amino group of the first amino acid does not form a bond to another amino acid, it is free and forms the N-terminus.

The first t-RNA is released by this process as it is no longer connected to an amino acid. The whole process commences with the third amino acid (and all those after that): go sideways by one codon, attach third t-RNA, break bond of second COOH end to t-RNA and attach it to third amino end, release second t-RNA, go sideways by one codon, attach fourth...

By breaking the bond between an amino acid and its t-RNA this acid is also no longer connected to the ribosome in any way. This means that the amino acid chain, that will later form a protein, slowly threads out of the robosome. [You can also see this whole process in one sequence of the Rosetta at Home video. Have a look at it, it is quite intuitive if you know what happens.]

Finally the last codon is reached and the last COOH bond to its t-RNA is released without forming a bond to another amino acid. This last COOH group now forms the free C-terminus.

This is how the protein is formed from the amino end (N-terminus) to the COOH end (C-terminus). That is why all proteins are synthesized from the N- to the C-terminus.

Rosetta does not assume that the N-terminus folds first. Is this a valid assumption? In almost all cases, the answer is, yes, it is a valid assumption. Even though the protein is synthesized from N to C, as nicely described by CJ below, this shouldn't affect the final folded structure for a couple of reasons which I will try to explain.
1) the native structure is at a thermodynamic minimum of the energy landscape. That means that no matter what route the protein takes (ie fold N first, or fold C first, or fold some other part first) given enough time it will find the lowest energy native conformation.
2) It is true that for some proteins, if they fold the "wrong" part of the protein first, they might become kinetically trapped in the incorrect conformation. That is, the amount of time it would take for the protein to unfold the incorrect structure and refold into the correct structure is too long on the time scale of the cell to be reasonable. So the cell has evolved mechanisms to overcome this. The main such mechanism is the "chaperone".

The chaperone: a large protein complex whose purpose is to help other proteins fold properly. Some chaperones work by providing a closed environment in which a protein can quickly and safely sample different conformations, until is finds its native conformation and is released by the chaperone to do its job. It allows the protein of interest to escape from kinetic traps and find its thermodynamic minimum.

The key thing to remember is that Rosetta basically ignores kinetics and focusses on thermodynamics.

V.

---

...if they fold the "wrong" part of the protein first, they might become kinetically trapped in the incorrect conformation.

That sounds a lot like the ranger's presentation of why the Gunnison river is where it is. They said if it had "known" then what it "knows" now that it has ground it's way down through 100s of feet of granite, it would have chosen a different route. But now that it is so deep, the river cannot take the "easier" route.

no matter what route the protein takes... given enough time it will find the lowest energy native conformation.

So, that makes it sound like no matter what random number our machine is given to start with, it should be able to find the lowest energy structure... but we don't want the process to take forever, so we use a "shotgun" approach and fire out 100,000 different starting points. And for some of them, the "route to take" will be comparatively short.

I've been wondering, it would seem if you start at random configuration 1, and test a few possible atomic twists that you would reach the same configuration as someone that starts at random number 2. I have read that there are on average roughly 3 to the n possible conformations, where n is the number of amino acids in the protein. So, a 100 long protein is a modest size it seems, and 3 to the 100 power is 5.14 e+47, so a 5 with 47 zeros after it. (can I call that roughly the square root of a google?? Did you know a "google" is a 1 with 100 zeros after it?) So ANYWAY, my question is, when I complete one of the roughly 100,000 models done for a detailed protein study, how many of these 5.14 e+47 possible conformations have we actually tested? And how common is it for my processing of random model 1 to eventually reach the identical configuration that someone else reaches as they are processing random model 2?

I took a graduate class in artificial intelligence. We were asked to write a program to solve the priests and wolves problem. You have... what was it? 3 priests and 5 wolves on opposite sides of a river and you have to ferry them around to reverse which bank each are on. But if you have too many wolves in the boat they will eat the priests, and more wolves the priests on either shore, the wolves will eat the priests. Basically you have to make a LOT of virtual trips back and forth across the river to find a combination that solves the problem with noone getting killed.

My proud moment was raising my hand in class when the assignment was due, when the instructor asked if anyone completed the assignment. You see, in 1986, the computer we were using didn't have enough memory to store the "game tree" that it was obvious to use to complete the assignment. The rest of the class wrote a program that would work in theory, but maxed out the memory and crashed before it could produce an answer.

The way I solved the problem was to keep a global record of configurations I had reached before. I mean if you are presently studying a situation with 2 priests on the left bank with 2 wolves, and 1 priest and one wolf in the boat, and 2 wolves on the right bank... it doesn't really matter how you arrived at that configuration. But you've been there before and concluded there was no amount of shuffling possible to make it work. So I stashed that configuration away, and from that point forward, before I took my virtual trip across the river, I'd check the list. If the configuration I was about to reach had already been tried, then I'd save the trip!

The other students were actually reaching the same point in the simulation more then once and processing it all over again. They were finding that you could reach the configuration in scores of different ways, some of which took dozens of additional trips. In fact, if you send the boat back with the same number of priests and wolves as it just came with, you can work yourself into an infinite loop! And consume more memory then the machine had, and crash your job.

I realize that there are too many conformations to stash away in a list, and it would probably take more time to maintain such a list and to compare to it prior to each proposed twist and turn... but I was wondering if there is some way to apply the same concept to Rosetta. To end the pursuit when you reach a point that is already known (by others) not to be fruitful.

---

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/

Rosetta does not assume that the N-terminus folds first. Is this a valid assumption? In almost all cases, the answer is, yes, it is a valid assumption. Even though the protein is synthesized from N to C, as nicely described by CJ below, this shouldn't affect the final folded structure for a couple of reasons which I will try to explain.
1) the native structure is at a thermodynamic minimum of the energy landscape. That means that no matter what route the protein takes (ie fold N first, or fold C first, or fold some other part first) given enough time it will find the lowest energy native conformation.
2) It is true that for some proteins, if they fold the "wrong" part of the protein first, they might become kinetically trapped in the incorrect conformation. That is, the amount of time it would take for the protein to unfold the incorrect structure and refold into the correct structure is too long on the time scale of the cell to be reasonable. So the cell has evolved mechanisms to overcome this. The main such mechanism is the "chaperone".

The chaperone: a large protein complex whose purpose is to help other proteins fold properly. Some chaperones work by providing a closed environment in which a protein can quickly and safely sample different conformations, until is finds its native conformation and is released by the chaperone to do its job. It allows the protein of interest to escape from kinetic traps and find its thermodynamic minimum.

The key thing to remember is that Rosetta basically ignores kinetics and focusses on thermodynamics.

V.

Thanks for the info, particularly that on chaperones. Maybe there should be something in the FAQ about them?

In spite of the two reasons you give why this issue is unimportant, it seems that it might be testable ( after a fashion ) relatively easily. Look at the distances between protein atoms in the Rosetta-predicted structure and those in the experimentally-derived structure. If these show a significant bias towards the N-terminus, it might indicate something along these lines is going on.

...if they fold the "wrong" part of the protein first, they might become kinetically trapped in the incorrect conformation.

That sounds a lot like the ranger's presentation of why the Gunnison river is where it is. They said if it had "known" then what it "knows" now that it has ground it's way down through 100s of feet of granite, it would have chosen a different route. But now that it is so deep, the river cannot take the "easier" route.

no matter what route the protein takes... given enough time it will find the lowest energy native conformation.

So, that makes it sound like no matter what random number our machine is given to start with, it should be able to find the lowest energy structure... but we don't want the process to take forever, so we use a "shotgun" approach and fire out 100,000 different starting points. And for some of them, the "route to take" will be comparatively short.

I've been wondering, it would seem if you start at random configuration 1, and test a few possible atomic twists that you would reach the same configuration as someone that starts at random number 2. I have read that there are on average roughly 3 to the n possible conformations, where n is the number of amino acids in the protein. So, a 100 long protein is a modest size it seems, and 3 to the 100 power is 5.14 e+47, so a 5 with 47 zeros after it. (can I call that roughly the square root of a google?? Did you know a "google" is a 1 with 100 zeros after it?) So ANYWAY, my question is, when I complete one of the roughly 100,000 models done for a detailed protein study, how many of these 5.14 e+47 possible conformations have we actually tested? And how common is it for my processing of random model 1 to eventually reach the identical configuration that someone else reaches as they are processing random model 2?

Those are excellent questions. During the protocol we're currently publishing for our structural prediction studies, we actually test for structural convergence of low energy models. Put another we, if see 20 / 50 low energy predictions with very low RMSD to each other, this would be a very high confidence prediction. Enumberating the possibilities for various proteins is something we're currently working on inside of the lab.

The first three degrees of freedom that many people talk about for protein structure prediction are usually phi, psi and omega angles, which describe the torsion angles between C-alpha atoms in the protein backbone. For a protein of N residues, there are 3^N degrees of freedom (as you point out), and 3^N gets big very quickly. Even worse, these degrees of freedom do not always give you the resolution that you need in order to discriminate the properly folded from the improperly folded conformations. So, this low resolution-search isn't sufficient to solve the problem. However, you can use it to very quickly through away highly unlikely conformations, so that you can spend your more detailed (aka higher-resolution) inspections for conformations that are more reasonable.

I took a graduate class in artificial intelligence. We were asked to write a program to solve the priests and wolves problem. You have... what was it? 3 priests and 5 wolves on opposite sides of a river and you have to ferry them around to reverse which bank each are on. But if you have too many wolves in the boat they will eat the priests, and more wolves the priests on either shore, the wolves will eat the priests. Basically you have to make a LOT of virtual trips back and forth across the river to find a combination that solves the problem with noone getting killed.

My proud moment was raising my hand in class when the assignment was due, when the instructor asked if anyone completed the assignment. You see, in 1986, the computer we were using didn't have enough memory to store the "game tree" that it was obvious to use to complete the assignment. The rest of the class wrote a program that would work in theory, but maxed out the memory and crashed before it could produce an answer.

The way I solved the problem was to keep a global record of configurations I had reached before. I mean if you are presently studying a situation with 2 priests on the left bank with 2 wolves, and 1 priest and one wolf in the boat, and 2 wolves on the right bank... it doesn't really matter how you arrived at that configuration. But you've been there before and concluded there was no amount of shuffling possible to make it work. So I stashed that configuration away, and from that point forward, before I took my virtual trip across the river, I'd check the list. If the configuration I was about to reach had already been tried, then I'd save the trip!

The other students were actually reaching the same point in the simulation more then once and processing it all over again. They were finding that you could reach the configuration in scores of different ways, some of which took dozens of additional trips. In fact, if you send the boat back with the same number of priests and wolves as it just came with, you can work yourself into an infinite loop! And consume more memory then the machine had, and crash your job.

I realize that there are too many conformations to stash away in a list, and it would probably take more time to maintain such a list and to compare to it prior to each proposed twist and turn... but I was wondering if there is some way to apply the same concept to Rosetta. To end the pursuit when you reach a point that is already known (by others) not to be fruitful.

That's a fun story, it's very gratifying to figure out things like that on your own. I don't know if you know it, but you may have rediscovered memoization, which does exactly what you're talking about. This is a way to try and avoid re-computing things many times. The problem (as you point out) is that the function space for this problem is huge, and no single computer can even come close to storing all of the possible answers. You could presumably store everything in a huge database and send things to and from Rosetta clients, but the heterogenous nature of Rosetta@Home makes this logistically hard. We still have a lot of users on dial-up.

There are a variety of tricks to conserving computational power, and we're definitely interested in pursuing them.

Thanks for the info, particularly that on chaperones. Maybe there should be something in the FAQ about them?

In spite of the two reasons you give why this issue is unimportant, it seems that it might be testable ( after a fashion ) relatively easily. Look at the distances between protein atoms in the Rosetta-predicted structure and those in the experimentally-derived structure. If these show a significant bias towards the N-terminus, it might indicate something along these lines is going on.

Hmmm, I admit it, I am guilty of oversimplification. You are right, this is idea is testable, and has been tested, at least after a fashion, and the conclusion is a little more broad than you might expect. It turns out that when you separate proteins by whether they have more long-range contacts or more short-range contacts (a quality we call contact order), the ones with lower contact order (more short-range contacts) fold faster. This applies not only to short-range contacts within the N-terminus, but throughout the length of the protein. Here's the trick: while some proteins may use low contact order as a way of facilitating their folding to the native state, there clearly are proteins extant that have all sorts of contact orders, including high contact orders (where, for example, the N- and C-termini might interact) and they all fold to their native states. So while contact order can be a useful predictor of topology for some proteins, it is important to balance the weight placed on contact order with other physical-chemical qualities that will apply to all proteins regardless of topology.

I do believe contact order is one of the terms in the ab initio scoring function, so it's not ignored.

Hope that helps!
V.

---

So which end is blue in the graphic, and which is red?

---

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/

N-terminal blue and C-terminal red.

So which end is blue in the graphic, and which is red?

---

Something perhaps i understand wrong but it attracts my attention here.

I read here that using random postions in the end the algorythm ends up with the same structure. While other people here pinpointed how many structures there can be.

So combining these thoughts, using pure random numbers isn't effective.
The chance is that numbers who are verry close to eachother are used by different users, who then fine tune end endup with the same result.
Thereby wasting 'global cpu' performance on double calculations.
I think that isnt making a project like this efficient.

Instead of random numbers isnt it better to use a model like distributed guesing
If you have a number lower then 20 and have ten people then let the first try one and two, the next will try 3 and 4, and so on..
So in the first round all possible combinations are done, and no double calculations are done. Effectivly all people have been used.

In folding this would mean the task would be pre devided into sub categories of shapes to test, and become a kind of efficient b-tree problem.





for the current program:

### it would be nice to have some effects in the video output, if one is finding a lower energy level, beating the old one, a bit of light, an email or a sound. perhaps a color indication of how low this energy is compared to the best known so far ?

### and just for the comparison be able to rotate the "low energy" and "native" model by mouse moving just for visual comparison, it would be fun to see how close it is to something.

### it would be nice to have some effects in the video output, if one is finding a lower energy level, beating the old one, a bit of light, an email or a sound. perhaps a color indication of how low this energy is compared to the best known so far ?

### and just for the comparison be able to rotate the "low energy" and "native" model by mouse moving just for visual comparison, it would be fun to see how close it is to something.

Well, it doesn't occur dynamically on your machine, because you don't have all the data about all the other predictions... but once you are done crunching a task, you can view your results and see how they compare with others. Click "[Home]" at the top of this message board page, then down on the left under "Returning Participants" click the "Results" link.

You are also shown how your current model compares to the prior models you have crunched on a given task with the little red dots in the lower righthand box with the scattergram. Although sometimes your current model has to change energy levels, which rescales the graphs, in order to see the red markers.

You can also manipulate the graphic and rotate the structures. It is a little picky and works best if you first expand the graphic to full screen. You can rotate (with left mouse button and drag), and also enlarge and reduce the scale (with right mouse button and drag up or down). You can also change the coloration of the sidechains by pressing the letter "C" key while displaying the graphic.

I'm not clear on your logic in presuming there is any wasting of global CPU going on. When there are
500,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000
conformations to study, it would seem your approach may require many more then the 6,000,000,000 people on earth to exhaust the search space.

The random numbers are only used as a starting point for a model, which is then searched and studied by the Rosetta science to see if it is likely what the native structure looks like.

---

Rosetta Moderator: Mod.Sense

I'm not clear on your logic in presuming there is any wasting of global CPU going on. When there are
500,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000
conformations to study, it would seem your approach may require many more then the 6,000,000,000 people on earth to exhaust the search space.

The random numbers are only used as a starting point for a model, which is then searched and studied by the Rosetta science to see if it is likely what the native structure looks like.

Then I just wonder what if not so random numbers where given for a start.
its just that if this caclulation ends up with the same structures. then i think there is some distributed cpu-time waiste in here.

Not completly sure of this but it also means that starting from some shapes you endup with a prefered shape, so a number of scenario's end up in the same final shape, after some time their different calculations will match.
If that is true then there is a part of math that this model could be improved.

Altough i don't know the math below it, and probaply personaly cannot improve it. I think for distributed calculations, one computer could tell some other computers, "okay guys, when your close to this b-tree, then stop calculating as it ends up with this shape" The nodes should give once in while such feedback of their current progress, and have it compared amongst other half finished calculation. The comparison could be done by a seperate server who also distributes the work loads.

A simple shape comparison could be done by for example by providing distance information of only a few atoms in the model. This simple aprouch would also shorten verry quick for the required comparissons.

If the native protein structure really is at a thermodynamic minimum as stated earlier in this thread, then why aren't the Rosetta-predicted minimum energy structure and the structure that results in the lowest RMSD value one and the same? In an ideal world wouldn't you expect the RMSD value of the lowest energy structure to approach zero if the native structure really was at its thermodynamic minimum?

An insufficient number of decoys possibly? RMSD maybe not the best metric for determining closeness to the experimental result? ( I seem to recall reading something about this but can't find the thread ). Crystallization distorting the native structure in some fashion?

---

Perhaps this is the post and thread you were thinking of?

---

Rosetta Moderator: Mod.Sense

Perhaps this is the post and thread you were thinking of?

Yes it was : thanks.

---