Gaming the System: Video Gamers Help Researchers Untangle Protein Folding Problem

[cegras]

http://www.scientificamerican.com/ar...olding-problem

When the researchers analyzed the strategies employed by a group of 57,000 Foldit players, they found that humans were better at some aspects of pattern recognition and protein structure prediction than current computational software. In fact, the gamers outperformed the computer on five out of 10 puzzles and delivered similar results on three other puzzles by using more varied approaches to solutions not used by the computer. The findings appear online August 4 in Nature (Scientific American is a part of Nature Publishing Group).

http://fold.it/portal/info/faq

1. Pack the protein

The smaller the protein, the better. More precisely, you want to avoid empty spaces (voids) in the structure of the protein where water molecules can get inside. So you want the atoms in the protein to be as close together as possible. Certain structures, such as sheets, will even connect together with hydrogen bonds if you line them up right and get them close together. This is also good. Key word: Compact.

2. Hide the hydrophobics

Hydrophobics are the sidechains that don't want to be touching water, just like oil or wax. Since most proteins float around in water, you want to keep the hydrophobics (orange sidechains) surrounded by as many atoms as possible so the water won't get to them. The other side of this rule is that hydrophilics (blue sidechains) do want to be touching water, so they should be exposed as much as possible. Key word: Buried.

3. Clear the clashes

Two atoms can't occupy the same space at the same time. If you've folded a protein so two sidechains are too close together, your score will go down a lot. This is represented by a red spiky ball (clash) where the two sidechains are intersecting. If there are clashes, you know something is wrong with your protein. So make sure everything is far enough apart. Key word: Apart.

[h0bbes]

I'm not bothered to remember my protein folding lectures but i feel that this is way over-simplified. It's not always about avoiding having voids (what about enzymes or channels), or having the hydrophobic molecules tucked inside. There are already software out there which are pretty good at dealing with steric clashes

[saaya]

yeah, this sounds too simple... proteins arent fixed 3d objects, they are constantly wobbling around and bending and folding when they interact...
sounds cool, but i doubt this will be very productive... if it would be, then theres something seriously wrong with their software thats supposed to do this lol...

[cegras]

I find it an interesting study in the holistic abilities of pattern recognition compared to brute force algorithms.

[h0bbes]

So many protein folding software use pattern recognition from what i remember but it just very difficult to predict the secondary structure of a protein.

[Serra]

Well the goal right now is not to actually do useful work, it's to show that people CAN do better than computers as a proof of concept. If that pans out then pharmaceutical companies can hire rooms full of people to play a more advanced version of this 'game' rather than buying large computers (which sounds more expensive until you realize that could be outsourced and done for about $1-$2/day/worker in many countries). That much is, I think, pretty much pointless given that computers will continue to get faster and people will stay at the same speed (though I suppose such people who do this 40 hours a week would get better at it too - to a point - and may actually widen the gap for a period of time).

But what they also get is a lot of data to analyze what makes people more successful in different areas of this specific application, which can then be used to create better programs for computers to do this work faster.


Edit: One interesting comparison here is that of chess. The results people are turning in now are compared to those being turned in by custom-built hardware and software. If we were playing chess, it would be like playing chess against a computer with a 2700+ rating. But the computer is being beaten by regular people, not 'masters'. How much wider would the gap be if people developed an actual training method for this type of work and dedicated themselves to it? If a folding master is to a regular folder what a chess master is to an average player, the results would really be astonishing... computers wouldn't catch those people in 10 years (if ever).

[h0bbes]

There has been a tremendous improvement in folding software in the past decades. Artificial neural networks seem to be working quite well but we are still a long way from predicting general protein structure.

Also for hydrophobic molecules, it is energetically favourable to have hydrophobic molecules tucked away from water but depending on their role, they might be directly in contact with water. Its about the overall energy state of the protein ( sum of all the interactions between all the amino acids and the environment) and this doesn't necessarily mean that every single amino acid has to be in it's most energetically favourable position.

[cegras]

The thing is though, we don't really know how the brain works. Well, I'm not majoring in neuroscience, but the actual nature of the 'transistors' and how they fire is lost to us. The way I see it, it is like a 'cascade' reaction where one stimulus input is exponentially amplified really quickly to give lots of meaningful results.

Chess programs still can technically be beaten. If you 'game' the system by making a lot of stupid, repetitive moves, eventually most algorithms will make a mistake or be backed into a corner. So basically, it comes down to how fast can you brute force it (by optimizing based on an algorithm) versus how effective can looking at the overall picture be.

[XSAlliN]

Not another subject where gaming is suppose to show you're smarter in some way.

Probably next, will see a news about guys that masturbate 10 times a day, which would imply based on some results that in terms of imagination they outperform architects with 10 years of experience.

[FischOderAal]

The thing is though, we don't really know how the brain works. Well, I'm not majoring in neuroscience, but the actual nature of the 'transistors' and how they fire is lost to us. The way I see it, it is like a 'cascade' reaction where one stimulus input is exponentially amplified really quickly to give lots of meaningful results.

While I'm not majoring in neuroscience either, I do know about artificial neural networks and how they work. Or do you mean the stuff that's going on "physically"?

[h0bbes]

Artificial neural networks in bioinformatics, is something like an artificial intelligence software used for protein folding. It's purpose is not to mimic our brain (well it is modelled after our neurons but anyway:P), but from what i remember we provide the software with information/rules so that it can make a more educated decision.

[cegras]

I mean more biologically; I was envisioning the brain as taking one signal and amplifying it / doing a lot of parallel analysis on it at once. All in all it's extremely holistic.

[tajoh111]

This is more of a biochemistry problem which I have a degree in. There is more to it than just hydrophobic interaction, you still have to worry about polarity, charges and disulfide bridges and a whole bunch of other stuff. In addition, there are other things that affect a protein shape such as if it just a domain that is part of a larger quanternary structure. Proteins will sometimes not need to have the same fold the same because they are joined together and they interact with other proteins to obtain a state of lowest energy. You will also not just deal with beta-pleated sheets and alpha helixes. Proteins get damn complex. Most of the time, when people see models of proteins, they just get colored helices, with alpha and beta turns. They forget that these things are made of amino acids which each have their own way of behaving.

What ultimately rules how a protein will shape is the combination of the above but most importantly thermodynamics. Basically a protein will take a shape where it has the most amount of entropy and it is not based all on hydrophobic interaction. You have to worry about the polarity, charge of side chains(which changes with pH) and also steric effects and mobility of certain groups in bending. Histidine doesn't like to bend so much because it has such a large sidechain with a ring.
Its the amount of variables that make protien folding a task best left up to computers.

[h0bbes]

Isn't it the least amount of entropy ? Of course hydrophobic interactions are not the only thing when it comes to protein structure but lots of protein folding methods start with trying to pack the hydrophobic core first. Cysteine residues are quite important (disulphide bridges) and can be quite useful in structure determination but they are not very common in proteins from what i remember.

Anyway there are a lot of things we dont understand about protein folding, such as how everything starts. Some scientists believe the hydrophobic core is formed first, others believe that different parts of the protein fold independently and then they all come together etc. It is in no way random because it would then take too long for a protein to fold, instead of fractions of a second which is what happens in reality.

Determining the structure of a protein can be done with computational methods to a reasonable degree of accuracy if it is about a protein which has a closely related protein of which we solved the structure.

X-ray crystallography remains the best way to solve a structure!

[tajoh111]

Isn't it the least amount of entropy ? Of course hydrophobic interactions are not the only thing when it comes to protein structure but lots of protein folding methods start with trying to pack the hydrophobic core first. Cysteine residues are quite important (disulphide bridges) and can be quite useful in structure determination but they are not very common in proteins from what i remember.

Anyway there are a lot of things we dont understand about protein folding, such as how everything starts. Some scientists believe the hydrophobic core is formed first, others believe that different parts of the protein fold independently and then they all come together etc. It is in no way random because it would then take too long for a protein to fold, instead of fractions of a second which is what happens in reality.

Determining the structure of a protein can be done with computational methods to a reasonable degree of accuracy if it is about a protein which has a closely related protein of which we solved the structure.

X-ray crystallography remains the best way to solve a structure!

Entropy is the amount of disorder, the greater the amount, more of it, means lowest energy. If we minimize entropy we maximize the energy state and it is not energy favorable. Less entropy mean a state that requires more energy to maintain because it requires more order. Remember gibbs free energy and you will remember why. E.g exposing non polar hydrophobic groups with water requires energy and to go to a lower energy state it will fold inward. Intermolecular H bonding is super important too.

Your right it not random, it's all being controlled by being as low energy as possible. Of course their are some times chaperone proteins to help the folding process.

I agree with the X-ray crystallography being a good technique. Disulphide bridges can be a lot more common in the right environment and are part of some pretty important structures.

[h0bbes]

I haven't done any thermodynamics for two years and I don’t currently have the time to remember Gibb’s laws :P I forgot about chaperone proteins and they certainly overcomplicate the situation when it comes to predicting protein folding.

Disulphide bridges are indeed very important, especially for quarternary structure (eg insulin).

X-ray crystallography is the best method we have so far in order to determine protein structure but its major limitation of course is trying to get the protein to crystallise in a good crystal and that's probably half of the work the crystallographer has to do. It has by far the greatest resolution of all the other techniques but in general it is better to combine different techniques in order to get the best possible structure. Even if bioinformatics cannot yet be used efficiently or reliably enough to predict protein structure, it is obviously a very useful tool already.