Results - CASP 9
CASP assessment summary
In the official CASP ranking we were ranked group 119 TBM and group 54 Link in the free modeling assessment.
This year's casp we predicted all models, which were sent out. At first the template based modeling results for Human groups were quite disappointing, because a place of 119 is simply not favourable, however once you inspect the results it can be easily explained. This year we focused on few high quality templates. Our approach in comparison to last year was not to take many different templates, relax them quickly and then select the best one, but to really do changes to the few template proteins, which we thought to be the best.
The problem now was that since we had a much smaller pool to choose from the template structure which we thought to be the best usually was not. The strategy we employed therefore had a fault directly at the beginning: Picking the best template structure.
From that we have learnt one very important thing: We must not chose the template structures by ourselves. We are good at actually picking a good structure once we have many models, but we are not good to make a worse structure better than the best which was built: The refinement process failed, because it would simply take too long.
For the Free Modeling we also focused more on template based models than last year, too. We thought to extract similar tertiary structure segments by that; because of this there was less time for the bottom-up protein construction approach than last year. However in the free modeling evaluation we did not too bad by itself, the results are convenient; of course we hope to do better next time.
But this is what CASP is all about. You propose a new algorithm to predict protein structures, you test it and you see if you fail or not.
Thanks again for all your support,
Martin and Timo
Previous status reports
The first hard target was T0531. For this target no good template structures could be found. After initial fragment model generations, we submitted the structures for structure building and rating to POEM@HOME, where we found 3 dominant folds.
The first fold is energetically most stable (and therefore usually the one we pick). It is a long extended beta sheet with many many hydrogen bonds stabilizing between the sheets.
Best raw energy structure for target T531
Usually when we discover a structure as stable as this one, we submit it immediately. This time however, we noticed that the disulfide bridges are unpaired (the red sticks, sticking out). Disulfide bridges usually stabilize the protein by a large amount of energy - therefore we also compared this structure with the remaining ones in the population.
Second best energy structure for Target T531
Obviously in this structure the disulfides are also unpaired, but at least 2 bridges are in the vicinity of each other. The last one is found in a loop allowing for the refolding of this loop.
Structure with all disulfide bridges developable
Finally the population contained a structure where all disulfide bonds can theoretically be linked. We started the simulation from here with an emphasis of disulfide bond generation and submitted the resulting structure. Obviously we cannot say yet if this structure is correct or not, because the final results are not out, but it was one of the most interesting structures to date.
Preliminary reference structures for Target T523 were released. We got an excellent RMSD of 1.162 for the models with your help - the comparison picture can be seen below:
Comparison between our prediction and the experimental structure of Target T523
Additions Thursday, July 22nd.
CASP is nearly over and some results have already been released by the experimentel groups. One of those results is T566, which occures as a dimer (complex of two times the same protein) in nature. Usually these proteins are a bit harder to predict, because they stabilize themselves and are not as stable left alone. Our models fits the experimental structure perfectly however with a resolution of 2 Angstroem. The picture (our structure is red, the experimental one green) can be seen below.
Comparison between our prediction and the experimental structure of Target T566
One target, which had an interesting prediction process was T537. Before submitting this protein, we actually thought we did a mistake, because the model does not look very protein-like. It's a long tube of beta sheets attached to each other. While it does exhibit a lot of secondary structure, the fold is seldom to be seen. The experimental structure agrees mostly with our predicts. Pictures from the front and through the 'tube' below.
Front view of experimental (green) and prediction (red) structures for target T537
Side view of experimental (green) and prediction (red) structures for target T537