New Update in 'Experimental' Update Group!
As many of you are probably aware, Rosetta is the "science engine" behind Foldit. It's been about two years since we last updated the version of Rosetta that Foldit uses under the hood. Since then, there have been many developments and improvements in Rosetta, and we thought it's time to update again.
Here's just a sampling of some of the improvements that come with the updated Rosetta:
* Fixes to some of the random crashes on MacOS
* Fixes to electron density-related crashes on Linux
* Fixes to symmetry-related crashes on all platforms
* Better support for non-protein residues, such as RNA, carbohydrates, lipids and non-canonical amino acids
* Support for modeling membrane proteins
* Improved detection of native-like hydrogen bonding networks
But perhaps the biggest change that comes with the update is improvements to scoring. There's been a *lot* of work put into the Rosetta scoring function recently, and just about every portion of scoring has been re-evaluated and re-optimized. (For those who want a nitty-gritty breakdown of the changes, a comprehensive overview has been published, along with details on how things were optimized.)
Here's a demonstration of the improvement. When looking at the ability to discriminate native-like proteins from non-native-like ones, for many proteins the new scoring function is able to do a much better job than the older scoring function:
Here each red point represents a structure prediction run for a different protein. The discrimination ability of the two scoring functions are plotted, using a metric where 0.0 represents no discrimination between native-like and non-native, and 1.0 represents ideal discrimination. The diagonal line represents no difference between the two scoring functions, and any points above the diagonal line represent proteins where the new scoring function does a better job than the old one.
This also is reflected in the score-versus-rmsd "funnel" plots for the predictions, where the new scoring function does a better job of eliminating false minima (blue) than the older score function does. (In these plots, better scoring structures are lower on the y-axis, and more native-like structures are further to the left. Eliminating false minima means a selection of top-scoring structures is less likely to include non-native-like ones.)
This improvement isn't limited only to protein structure prediction. The new scoring function shows discrimination improvements in a wide range of protein prediction problems, including protein-protein and protein-small-molecule interaction predictions.
There is a slight drawback to these improvements, though. The new scoring function is slower than the current one. (In our tests, it averages about 30% slower.) We don't anticipate this being noticeable in general interactive use, but it may affect things like long-running shakes and wiggles. Most affected will be scripts which use a set number of iterations of shake and wiggle - these will run for longer, and if you've optimized the number of iterations for the current scoring function, the optimal number of iterations may have changed in the new one.
We're excited about these score function improvements, though, and think the better results are worth the slowdown. You might spend a bit more time working on a single structure, but you should hopefully spend much less time working on "bad" or "scientifically uninteresting" structures.
So, if you're feeling adventurous, please help us out by testing the updated version. To do this, switch your update group to `experimental`. In addition to testing how the slowdown affects scripts, we also want to make sure no bugs have slipped in on how the various tools behave. -- Note that, due to the difference in scoring, the puzzles available with `experimental` are not the same as with the `main` and `devprev` clients. If you want to play the regular Foldit puzzles, you'll need to switch your update group back to 'main'. None of the 'experimental' puzzles will count towards your website rank, but they will help us work out any issues prior to releasing it to the general public!( Posted by rmoretti 84 618 | Thu, 02/08/2018 - 21:42 | 3 comments )