Google cloud runs thousands of protein simulations simultaneously

[PoppaGeek]

Read more at - Ars Technica

We know how to get snapshots of what proteins look like. These static pictures tell us where all the atoms of a protein reside within a crystal, which gives us a sense of their structure and lets us design drugs that fit neatly within that structure, altering its activity.

But, in actual cells, proteins are nothing like the static, rigid structures found in crystals. Instead they writhe, buffeted by Brownian motion and constantly shifting among similar energy states. Until we develop a microscope that can resolve all this motion, the best we can do is to run molecular simulations on our computers. Unfortunately, most proteins have a lot of atoms to keep track of, which makes those simulations extremely computationally expensive.

Now, some researchers have figured out how to run the simulations on Google's cloud computing architecture. Although each of the individual simulations is short, they can be aggregated to provide a picture of long-term behavior. And, with this method of aggregating them in place, the system should be able to work with just about any cloud service available.

From comments:

David KonerdingSmack-Fu Master, in training jump to post
Hi folks-
I'm one of the paper authors and would like to a few minor clarifying comments:

1) The computer time was provided free to the scientists.

2) Kai Kohlhoff, who was a postdoc in Vijay Pande and Russ Altman's labs, joined Google as a Visiting Faculty. He used roughly half a billion CPU hours for this calculation (and some others).

3) Much of the work was based on software developed by the Folding@Home team. In many ways, Exacycle resembles F@H in design and ran a binary resembling F@H's core code (gromacs). Further, we used MSM (https://simtk.org/home/msm-database) although rewritten in Google Flume, to do the data analysis.

4) As pointed out, we did not carry out a single 2ms simulation. The results were derived from many shorter simulations. Some of us believe this is actually a better sampling method than single, long trajectories, although that's an open question. I'll note that my PhD work, executed 13 years ago, was comprised of 4 10ns simulations.

5) Regarding the comments that this is mainly a computational advance; we didn't pick GPCRs, or simulate drug binding by chance, and our predictions are testable in a lab. In general, I would agree that pure folding simulations are unlikely to produce pharma-accurate predictions in the near future. I *hope* that our technology and models will become increasingly accurate over time, and potentially help in the development of new drugs that reduce the cost of health care and reduce side effects. But we have a lot of work to do.

6) We just published *another* paper related to this, but focusing on improving Rosetta's force field:
http://onlinelibrary.wiley.com/doi/10.1 ... ated=false

[0ne.shot]

" In general, I would agree that pure folding simulations are unlikely to produce pharma-accurate predictions in the near future."

That's interesting.