About Foldit and RoseTTAfold

Case number:845813-2011936
Topic:Game: Tools
Opened by:ullian
Status:Open
Type:Question
Opened on:Sunday, August 1, 2021 - 13:47
Last modified:Monday, August 2, 2021 - 21:59

Hi! I'm a researcher focusing on serious gaming and also a fan of Foldit!

Recently RoseTTAfold really makes a great hit for its accuracy and efficiency. I noticed that one of Foldit's interesting goals is to help researchers find out if human players' pattern recognition and puzzle-solving abilities can help advance protein-folding software. So I'm really curious about whether Foldit's data makes any difference in the development of RoseTTAfold? Does human players really do better at protein folding?

Any concerned articles or informations would be apreciated! Please contact me on ulliansong@gmail.com.Thank you!

(Sun, 08/01/2021 - 13:47  |  1 comment)


bkoep's picture
User offline. Last seen 6 hours 25 min ago. Offline
Joined: 11/15/2012
Groups: Foldit Staff

Good question! Foldit data was not used in any way to train the RoseTTAFold neural network. RoseTTAFold was trained on solved protein structures deposited in the PDB, in addition to evolutionary sequence information about natural proteins. (Technically, there are a handful of Foldit-designed protein structures in the PDB, but I believe these were excluded from the RoseTTAFold training set, so they should not affect the RoseTTAFold algorithm.)

These deep neural networks are very, very good at protein structure prediction! There will always be cases where the network predictions fall short or do not tell the full story about a protein, and human predictions may still be useful in some of those cases. But, by and large, these neural networks seem to be the better option for raw protein structure prediction. We think that humans have more to contribute to other problems, like protein design or model-building with experimental data.

Sitemap

Developed by: UW Center for Game Science, UW Institute for Protein Design, Northeastern University, Vanderbilt University Meiler Lab, UC Davis
Supported by: DARPA, NSF, NIH, HHMI, Amazon, Microsoft, Adobe, Boehringer Ingelheim, RosettaCommons