Hello everyone,
I just wanted to show an example of the Rosetta energy function that is used in Foldit.
We have evidence that the Rosetta energy function is able to pick out the native conformation when compared to many other models.
In the attached plot, the y-axis is the Rosetta score (negative is better here) and on the x-axis is how far off the model is from the native (so closer to the left is better). The wiggled native proteins are shown in blue on the very bottom left with all the black dots representing Rosetta predictions.
If you were to look at all models with a Rosetta score of exactly -170 (draw a line horizontally across the graph at -170) you can see that these models are very different from one another. If you look at the most successful predictions (shown by the red arrow) you can see that they have the best Rosetta score and are closest to the correct native structure. Based on this, we believe that our energy function is good at picking out folds that are similar to the native.
You'll notice that most of the Rosetta predictions on this graph are far from the native structure.
We are hoping that Foldit players will be able to more efficiently make accurate models of protein structures, because the automated methods are essentially random searches.
This figure is taken from the Science article:
"Toward High-Resolution de Novo Structure Prediction for Small Proteins"
Philip Bradley, Kira M. S. Misura, and David Baker
Science 16 September 2005: 1868-1871.
Where exactly is the "attached plot"?
I don't see one.
It is a link, and now I've embedded it in the main post as well.
Thanks for adding the image, I wasn't able (and still aren't?) to see any link to that graph. It makes the story a lot more clear, lol.
I can see how these results validate the Rosetta scoring method, great!
It brings up two questions (right now at least, as I'm sure there are more questions hanging out in my head, but they are slackers..)
-
(with the risk of sounding ultimately stupid…) How do you know those correct native structures?
-
did foldit players work on computer generated structures of proteins with a known correct native structure, and if so, what does the graph look like if you plot their solutions as well?
I subscribe to Science so I was able to read the original article online. The article and post answer two nagging questions I've had about how foldit can work while ignoring sidechain charge and how accurately it is in identifying the native state.
Any more posts about the science or links to good references would be appreciated.
the native structure is often determined through x-ray crystalography, a complex time-consuming process performed in the laboratory. It is likely that proteins determined in this fashion also have some small amounts of discrepancy, but are generally considered to be accurate.
A number of folidt puzzles were done on the proteins for which we know the native structure. Sometimes the players get it exactly right, other times not. The grand challenges, are designed to show the capabilities of this new computational framework.
