Hi Everyone!

Check out our recently published paper. We used Roseta@home and charity engine (which also uses BOINC) resources extensively for this project. Thanks everybody!



A) predicted structures with co-evolution contacts used to guide structure prediction.
B) predicted structure
C) correct answer!

We were inspired by these results and went on to predict structures for ~60 other proteins with similar data.

Below is a summary:

The three dimensional structures of proteins are specified by their amino acid sequences and are critical for proteins to carry out their biological functions. Therefore, to understand how a protein works, it is important to determine its structure, but this is very challenging. It is possible to predict the structure of a protein with high accuracy if previous experiments have revealed the structure of a similar protein. However, for almost half of all known families of proteins, there are currently no members whose structures have been solved.

The three-dimensional shape of a protein is determined by interactions between various amino acids. During evolution, the structure and activity of proteins often remain the same across species, even if the amino acid sequences have changed. This is because pairs of amino acids that interact with each other tend to ‘co-evolve’; that is, if one amino acid changes, then the second amino acid also changes in order to accommodate it. By identifying these pairs of co-evolving amino acids, it is possible to work out which amino acids are close to each other in the three-dimensional structure of the protein. This information can be used to generate a structural model of a protein using computational methods.

Now, Ovchinnikov et al. developed a new method to predict the structures of proteins that combines data on the co-evolution of amino acids, as identified by GREMLIN with the structural prediction software called Rosetta. A community-wide experiment called CASP—which tests different methods of protein prediction—showed that, in two cases, this new method works much better than anything previously used to predict the structures of proteins. Ovchinnikov et al. then used this method to make models for proteins belonging to 58 different protein families with currently unknown structures.

These predictions were found to be highly accurate and the protein families each have thousands of members, so Ovchinnikov et al.'s findings are expected to be useful to researchers in a wide variety of research areas. A future challenge is to extend these methods to the many protein families that have hundreds rather than thousands of members.

Check it out: http://elifesciences.org/content/4/e09248

Any questions are welcome =D

Nice job on the new publication! The results are really fascinating and this co-evolution assisted prediction method seems promising. Does this prediction method have any major drawbacks or limitations? Lastly, what kind of lift was this in terms of CPU hours or credits (whatever is easier to size?)

Also interesting to see the baker lab actively leveraging Charity Engine's resources. Would those CE contributions showed up in the Rosetta@Home credit clock or were the extra FLOPS unreported on R@H's homepage? Just curious...

Again, thanks for sharing!

Tim

PLEASE post this on the homepage.

---

Thanks TIM!

The limitations of the method is the requirement of sequences. For any protein family we want to model we need 1000s of non-redundant sequences. So if it's a protein that is only found in a few sequenced organism, we won't be able to use co-evolution data.

Not exactly sure how many core hours we spent in total, but not including the CASP calculation and other benchmarking, I used 870K workunits from which 6.2 million models were produced. Each workunit is 3-6 hours I believe.

Charity Engine works in chunks, when they have some idle compute time they add them to the BOINC network (Rosetta@home is one of the preselected projects). We got a couple million core hours of compute from them! These compute nodes will show up in the normal boinc stats.

Nice job on the new publication! The results are really fascinating and this co-evolution assisted prediction method seems promising. Does this prediction method have any major drawbacks or limitations? Lastly, what kind of lift was this in terms of CPU hours or credits (whatever is easier to size?)

Also interesting to see the baker lab actively leveraging Charity Engine's resources. Would those CE contributions showed up in the Rosetta@Home credit clock or were the extra FLOPS unreported on R@H's homepage? Just curious...

Again, thanks for sharing!

Tim