Thursday, March 4, 2010

Today I found an interesting site called Foldit, where you can download a program that lets you try folding a protein graphically. It differs from sites like Folding@home, which just draw on your computer's spare processor cycles, in that it actually hopes to use human intuition to help solve the protein folding problem, and compare it to the performance of computer algorithms on the same task. I downloaded it, and wanted to give my opinion on it.

First for the pros:
I am very happy (and pleasantly surprised) by the overall idea, that of finding humans who have unique mental capabilities that may be useful for solving real-world problems in biology. It often seems that the only kind of unusually gifted minds (whether you call them prodigies, savants, etc.) that most people know about, or at least care about, are "Rainman"-like human calculators, and then only as some kind of "freak show". I do believe there well could be people who have an intuitive sense of the forces stabilizing proteins, and could fold at least small proteins intuitively, just science never hears of them. And it's not just with this problem by any means. I think there are lots of optimization sort of problems in all areas of science and technology where certain humans could outperform computers--yet the dogma in the field is that there is no use in humans trying to do these tasks themselves.

In the modern world of the internet, I think the time has very much come for people who know or suspect they have such unique talents to network with each other, and share both ideas and life experiences. Whether it is to actually figure out how to use those skills in the real world, or just to relate to the feeling of alienation that arises when one feels his/her thinking style is not accepted by others, this is a definite niche for a social networking site. In fact, it was exactly this type of group I was looking for when I stumbled across Foldit. In case you're wondering, I did not find any such groups, so I will probably be starting one in the not too distant future

But back to Foldit, I quickly realized there are many things that make it less than ideal, both as a way to discover hidden talent and as a means of solving the folding problem:
1) Lack of transparency:
When one looks at the ranking system for users, it becomes clear that the judgment criterion is not similarity to some known native structure, but score by a Rosetta-type scoring function. The specific function used, however, is obscure. If this were just some video game, I could see the makers wanting to keep the exact scoring system secret, in order to not make the game too easy. However, if this is even pretending to be science, then a very open disclosure of the nature of the problem is in order, even if the scoring function is very realistic. One might worry that using a real native structure as the reference would allow cheating by those who know the structure--but CASP has found a logical way around this, by using unpublished structures.
2) Manipulation is unwieldy:
The only way to fold a protein is to bend the backbone by dragging it from one point, and then the rest follows in a "springy" fashion. There's no good way to "tweak" individual residues. Also, there should be some sort of way to specify helix formation and beta-sheet pairing on some global level, and then build these elements as "ideal" units that can then be bent to allow packing.
3) Side chain handling is weird:
Side chains do not have full mobility, they "spring" only to common rotamers that are reasonably clash-free. While this can be useful, especially for beginners, as it prevents really bad things like sticking a methionine through the "hole" in a benzene ring, it also means that it is not possible to stretch a side chain to form, e.g., a hydrogen bond, and then allow the structure to relax with this already made. That goes for everything, in fact--while interactive molecular mechanics is way cool, it's very good to be able to turn OFF the force field, to allow getting out of local minima. Especially when the mechanics is not "real" molecular mechanics. Also, Asp/Asn and Glu/Gln are impossible to tell apart, and the "flip" of Asn/Gln amides is invisible. This is not great for folding, and downright awful for ligand modeling (see below).
4) Ligands are included, but almost definitely not well:
A few of the tutorial levels involve moving a non-protein ligand into position. This was probably introduced to allow enzyme/receptor design. However, one must wonder what kind of force field is used to describe protein-ligand interactions. The whole idea of Rosetta is very "globalist", and for something like ligand recognition, which could be dominated by precise hydrogen bonding and stacking geometries, it is unlikely that this program does it justice. Add to that the insistence that every side chain be in a common rotamer at all times, and it is not even possible to move them into ligand-centric positions. So, the utility of this feature is questionable, in my opinion.

So the conclusion is, two thumbs up for the effort to give human intuition a chance to hack away at real scientific problems. But, this program may well do a rather poor job at this, as well as in predicting structures. It was rather easy to get within the top five scores for some proteins, and yet I did not feel as if I was actually creating good-looking structures. Most likely, the top scorers will be those who can best get around the awkwardness of being stuck in local minima by the limited controls. And if you want to test your ability to predict structures for your own sake (as opposed to impressing others), maybe downloading old CASP targets is a better bet.

No comments: