puzzle picture
2027: Design with AlphaFold Predictions
Status: Closed

Summary

Name: 2027: Design with AlphaFold Predictions
Status: Closed
Created: 08/04/2021
Points: 100
Expired: 08/31/2021 - 23:00
Difficulty: Intermediate
Description: Design a protein with the help of the new AlphaFold prediction tool! AlphaFold is an algorithm from the company DeepMind and can predict protein structures with incredible accuracy. We want to know if Foldit players can use feedback from AlphaFold to design new proteins! After you've designed a protein, you can upload your solution to the Foldit server and get back an AlphaFold prediction about how your design will actually fold up. Players should aim for an AlphaFold confidence and similarity of 80% or more. See the blog for more details about the AlphaFold prediction tool.

We are especially interested to see what creative new folds players might come up with. We already know that Foldit players can design simple helix bundles or ferredoxin-like ("surfing hotdog") proteins. We are hoping that AlphaFold will allow players to explore protein designs with all beta sheets, or folds with a recessed binding pocket where a ligand could fit. After this puzzle closes, we'll highlight our favorite designs in the next Lab Report video update on September 1.

In addition to the typical extended chain, this puzzle has four starting structures of designed proteins that you may like to use as a starting point. Reset the puzzle to cycle through the different starting structures. There are no Objectives, but we expect successful designs will still need lots of helices or sheets, with short minimal loops, as well as a closely-packed core of orange hydrophobic residues. AlphaFold predictions will not affect your score. Players may insert and delete residues to a maximum of 120 residues total. This special puzzle will remain online until August 31 at 23:00 GMT.
Categories: Design, Overall

Top Groups

RankGroupScorePoints
1Anthropic Dreams13,411100
2Contenders13,39777
3Go Science13,38458
4Beta Folders13,36043
5L'Alliance Francophone12,79131

Top Evolvers

Top Soloists

RankPlayerGroupScorePoints
1WakeOwl 71 165  13,536100
2dcrwheeler 71 6  13,45098
3Galaxie 2 3 Anthropic Dreams13,41096
4phi16 12 26 Anthropic Dreams13,40893
5grogar7 42 1 Anthropic Dreams13,40491


Need this puzzle? Log in to download.  

Comments

Susume's picture
User offline. Last seen 7 hours 10 min ago. Offline
Joined: 10/02/2011
Which version should we share with scientists?

If I've reached agreement with AF on the topology (confidence >80), and AF and foldit are now duking it out about how big the voids will be or what angle the helix will lie against the sheet (similarity never quite reaching 80), is it really useful to keep pushing for them to agree, or can I call that one done and move on to something else?

And which version do scientists want to see? I guess if I upload mine you can always run AF on it yourself? But if you will discard it due to similarity < 80%, I'll be sad.

bkoep's picture
User offline. Last seen 1 hour 56 min ago. Offline
Joined: 11/15/2012
Groups: Foldit Staff
Good question!

Please share your Foldit-optimized design (not the raw AlphaFold prediction)! We can easily run AlphaFold predictions for your shared designs, and we like to see a players' best attempts in Foldit.

. . . is it really useful to keep pushing for them to agree, or can I call that one done and move on to something else?

We are still uncertain about this. Among previous Foldit designs, we have seen that AlphaFold similarity is slightly better than AlphaFold confidence for predicting successful designs. That suggests that, yes, it's worth it to keep refining your Foldit solution until AlphaFold similarity is greater than 80%.

However, we don't know how well that holds if you are iteratively loading AlphaFold solutions, refining them in Foldit, resubmitting to AlphaFold, re-refining in Foldit, etc. . . . This kind of strategy might "artificially" boost your AlphaFold similarity value without making meaningful improvements to your design.

Either way, I would hope that you could eventually find a high-scoring Foldit solution that also has a high AlphaFold similarity. If the differences cannot be reconciled, that might suggest a deeper problem (so you should share examples with scientists!).

Bautho's picture
User offline. Last seen 1 week 1 day ago. Offline
Joined: 02/16/2012
Groups: Go Science
That's super cool :)

Can you say anything about the time, how long it will take to run the prediction?
Does it depend on the number of residues?

Bautho's picture
User offline. Last seen 1 week 1 day ago. Offline
Joined: 02/16/2012
Groups: Go Science
Foldit crashes

... when loading solution from AlphaFold.
I can share the original design to scientists :)

bkoep's picture
User offline. Last seen 1 hour 56 min ago. Offline
Joined: 11/15/2012
Groups: Foldit Staff
Crash feedback

Yes, please share with scientists! We've received sporadic crash reports, but have had difficulty reproducing them.

Joined: 12/07/2007
Groups: Contenders
I'm seeing this too. Doesn't

I'm seeing this too. Doesn't happen with all AF-generated solutions but when it does it crashes consistently. Scientist-shared solution t4_c, when fed to AF, generates a solution which crashes when you try to load it.

Bautho's picture
User offline. Last seen 1 week 1 day ago. Offline
Joined: 02/16/2012
Groups: Go Science
Consistency

I agree. Even when submitting the same structure again, it crashes every time you load any of the solutions.

beta_helix's picture
User offline. Last seen 1 hour 30 min ago. Offline
Joined: 05/09/2008
Groups: None
Thanks all, we found the issue :-)

Thank you all for your scientist shares, they helped us uncover the issue.
We'll post a hotfix soon!

Bautho's picture
User offline. Last seen 1 week 1 day ago. Offline
Joined: 02/16/2012
Groups: Go Science
It is working :)

In devprev I can load the solutions, which have not loaded before.
Thanks for the quick fix!

LociOiling's picture
User offline. Last seen 3 hours 54 min ago. Offline
Joined: 12/27/2012
Groups: Beta Folders
multistart! (must learn to read...)

This is a multistart puzzle, which is explained in the last paragraph of the puzzle description, which I hadn't read until just now.

When I started the puzzle, I saw a straight "extended chain", so I assumed that was that....

Resetting the puzzle gives you a different structure. Structure 5 is the extended chain. You can reset by ctrl-r or Undo => Reset Puzzle. The first time you open the puzzle, Foldit picks a structure at random.

The structure number appears briefly in a pop-up when you reset. The segment information window also contains the structure number in the third line. To open the segment information window, hover over any segment and hit the tab key.

I used Jpred to search for matches to the primary structure (amino acid sequence) of structures 1 through 4. Jpred looks for matches to "solved" proteins in the PDB (Protein Data Bank, at rcsb.org), and also in the UniProt database of protein sequences. Unlike the PDB, entries in UniProt may not have a known 3D structure. (Maybe AlphaFold can get that sorted.) Jpred doesn't always find PDB matches.

The results were:

  • Structure 1 - length 43 - JPred finds no matches, but rscb.org finds PDB 5UOI
  • Structure 2 - length 68 - matches 6MRR, player-designed protein Foldit1
  • Structure 3 - length 106 - no PDB match, but lots of UniProt matches
  • Structure 4 - length 109 - matches PDB 6D0T (chains A and B)

So three of the structures have been "solved", and structure 2 was designed by Foldit players in the first place.

Structure 3 shows UniProt matches, where you'll quickly see things like "Cluster: SnoaL-like domain-containing protein" and *Burkholderia pyrrocinia*, part of the Burkholderia genus. Interesting, but not as helpful as having a 3D view available. If you look up "SnoaL", you may find "the proteins in this family adopt a distorted alpha-beta barrel fold", and see PDB 1SJW as an example.

(Edit: search at rcsb.org found match for structure 1, Jpred is not perfect.)

Susume's picture
User offline. Last seen 7 hours 10 min ago. Offline
Joined: 10/02/2011
All Baker lab designs (Foldit being part of Baker lab too)

Structure 3 is the type of curved beta-sheet design that was featured in the Marcos paper from Baker lab - PDB 5L33 is another example.

Joined: 12/07/2007
Groups: Contenders
When running a Remix script

When running a Remix script on some structures, I'm seeing cases where small changes (at least as far as the FoldIt score and structure is concerned) result in major changes to the AF score. Trying to narrow this down a bit, there's a case where changing a lysine to an asparagine results in a drop of 3 points in Foldit which is basically nothing. However it makes a huge difference as far as AF is concerned: dropping the score from (83.0%/80.7%) to (75.1%/61.3%) while doing what looks like a register shift on the sheet structure.

Is this sort of behaviour to be expected? Scientist-shared solution s6_10302 in case its interesting.

bkoep's picture
User offline. Last seen 1 hour 56 min ago. Offline
Joined: 11/15/2012
Groups: Foldit Staff
Expected

Yes, I believe this is to be expected in some cases. The AF neural network is very complex, and two input sequences that are relatively "similar" can in theory produce wildly divergent outputs. Whether or not this accurately reflects any "truth" about the actual folding propensity of these sequences is difficult to say without further analysis or experimental testing. However, there are certainly examples of proteins where minor mutations are known to result in drastically different folding behaviors!

Remember also that the Foldit score only reflects the energy of your solution in its given structure, and tells you nothing about the energy landscape or the energy of alternative structures.

User login
Download links:
  Windows    OSX    Linux  
Windows
(7/8/10)
OSX
(10.12 or later)
Linux
(64-bit)

Are you new to Foldit? Click here.

Are you a student? Click here.

Are you an educator? Click here.
Social Media


Search
Only search fold.it
Other Games: Mozak
Recommend Foldit
Topics
Top New Users
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