Discussion of Contact Puzzles

Case number:699969-997478
Opened by:Susume
Opened on:Saturday, April 12, 2014 - 01:17
Last modified:Tuesday, November 8, 2016 - 23:55

Just wanted to collect recent discussion of the contact puzzles in one place where people can find it. Feel free to add.

(Sat, 04/12/2014 - 01:17  |  16 comments)

Susume's picture
User offline. Last seen 2 days 20 hours ago. Offline
Joined: 10/02/2011

Comments about Puzzle 871b, Denovo Freestyle 36 Predicted Contacts:

Comments from chat about how this puzzle works
bkoep: Contacts are actually being scored as constraints in this puzzle, so the scoring mechanism for contacts is completely different from the visualization.
auntdeen2: ? then what good is the visualization? a semi loose guide?
bkoep: We still wanted you to have access to the "contact map"
bkoep: And the constraints match the contacts shown on the map
bkoep: Well, the constraints have different weights, for one, which shows up nicely in the map
bkoep: And the map might also be useful for inferring other information (such as the extent and placement of secondary structure units)
Susume2: the scores are the same when the contact is red as when it is green in the same position
bkoep: Okay, that makes sense.
auntdeen2: so we can't trust the color to know if we have made the contact?
bkoep: In this puzzle, it's less important to "make" the contact, because it isn't scored as an all-or-none bonus
bkoep: As two contacts get closer together, they score more and more points (up to some threshold)
bkoep: The color of the contact band does not reflect the score contribution of the contact
bkoep: this is analagous to how Rosetta approaches contacts, so we wanted to try it out
bkoep: It's worth noting that since the contacts are not scored as "bonuses" but contribute to the normal Stability score, wiggle will actually help you make the contacts if you're close enough.
Susume2: so they are sticky like the ED map is sticky
bkoep: So I would recommend relying on the map and contact visualizations only for large-scale planning and placement of SS elements
bkoep: @Susume, YES, exactly

Susume's picture
User offline. Last seen 2 days 20 hours ago. Offline
Joined: 10/02/2011

Comments about 875b: De-novo Freestyle 37: Predicted Contacts:

jmbrownlee333: one thing I don't get is the distance implied by the contacts. It seems so varied.
TimovdL: I think it is not backbone distance, but also sidechains play a role not sure about it however
jmbrownlee333: contacts from NOE's are pretty much limited to 5 angstroms, or am i wrong about that?
drjr: dunno. but 5 A seems informative
Susume2: what's an NOE?
drjr: Nuclear Overhauser effect
jmbrownlee333: from NMR.
jmbrownlee333: but some contacts seem much longer than 5 angstroms.
drjr: are they green yet?
karstenw: i'm getting all different lengths. i got a green one at 10 band length on the previous protein.
karstenw: had another that wouldn't turn green until 4.83 band legth
jmbrownlee333: right karsten. is there any rhyme or reason to the length that turns green.
karstenw: i assume so, but i don't know what it is.
bkoep: Hey guys, I can explain a little about the different distances required for contacts
jmbrownlee333: please
bkoep: First of all, the contacts are supposed to represent just that--two residues that directly interact with each other
drjr: These experiments, if it is in fact NOE, have to be done with about 20 min per interaction, at least in the 90s, and the sample has to be pretty concentrated
bkoep: Be it hydrogen bonding, electrostatics, or simply butting up against one another, there are many different ways for two resides to "contact" one another
karstenw: ah, makes sense
bkoep: Rather than go through and check each of these possibilities (which could be slow), we make an approximation
bkoep: We look at existing protein structures, and see that when two residues have any of these "contacts", their C-beta atoms are within x distance of one another
bkoep: So we just measure the C-beta/C-beta distances between tow residues to figure out if they are close enough to be "contacting"
karstenw: do we have a max distance that two segs could make contact under the right conditions?
bkoep: Then we notice that different pairs of residues actually have different C-beta/C-beta distance threshold for what constitutes a "contact"
bkoep: For instance, two tryptophan residues (large) often make contacts even when their C-beta atoms are pretty far away (like 12 angstroms)
frood2IRC: so why is this explanation not posted in the puzzle comments? It would avoid much consternation
karstenw: ok, 12 on the long suckers. makes sense
bkoep: On the other hand, two glycine residues (small) that are 12 angstroms apart would never make any meaningful contact
jmbrownlee333: and how does this relate to CASP?
karstenw: what do we think is the shortest distance for contact?
bkoep: @jmbrownlee, CASP now releases some targets with "predicted contacts"
bkoep: There are different ways to predict contacts in a protein structure
bkoep: Sometimes you can get experimental data (for example, from NMR experiments)
frood2IRC: yes bkoep - but can we have this posted straightforwardly?
bkoep: But a lot of times these predictions are generated from a covariation analysis of related proteins.
jmbrownlee333: or chemical cross linking i guess
karstenw: there is a contact between seg 21 and 5 that suggests the sheet order, but of course the contacts could come from another angle where the two "reach out" for each other via there sidechains. unlikelt imo
karstenw: *unlikely
karstenw: more likely they sit right next to each other.
TimovdL: bkoep, is there a hidden score when we get contacts right?
karstenw: alright, i got the contacts i like. i'm running it.
frood2IRC: Timo: interesting answer to that one :(
drjr: Timo -there is. Brow mentioned it. SUm up the parts and subtract it from the segment energy
drjr: The total Contact energy is about 500 for my 871b
TimovdL: So it is inside the segment energy?
drjr: It appears to be
TimovdL: Then it really should be a subscore
bkoep: @Timo, in these puzzles, contacts are being scored as constraints, which actually contribute to the "Backbone" subscore
frood2IRC: so we look at the back bone score yes?
bkoep: We realize this is not ideal, but we're still trying to figure out the best way to use predicted contacts in Foldit.
frood2IRC: sure...but my question remains
TimovdL: Maybe my suggestion I put in feedback would be a way
bkoep: I agree. If we end up using these constraint-type puzzles (as opposed to puzzles with hard bonuses for making contacts), then a separate subscore would be the way to go.
alcor29: Thanks for the explanations bkoep.

gitwut's picture
User offline. Last seen 3 years 3 weeks ago. Offline
Joined: 05/18/2012
Groups: Contenders

I think this new style "contact map" puzzle is confusing. Why not just call them constraint puzzles and have a checkbox as to whether constraints are visible or not? Why display the stability or multiplier score when one basically reflects the score and the other never changes?

It also seems to me that they should be different categories since excelling at one type doesn't necessarily mean that you'd excel at the other--not to mention that they behave entirely different.

Please elaborate in layman's terms if I'm missing something obivous--it wouldn't be the first time. If I am, I still seem to be scoring decently in spite of it.

bkoep's picture
User offline. Last seen 18 hours 38 min ago. Offline
Joined: 11/15/2012
Groups: Foldit Staff

Good questions, gitwut.

Most of your questions could be answered with the response that we're just trying out this style of Predicted Contacts puzzles, and many of these ideas are probably not here to stay. We couldn't afford to invest in developing a better-tuned interface for just a couple trial puzzles, so we've been trying to make do with the available contact map infrastructure. I grant that this style is less straightforward, but it has a couple advantages (e.g. the "stickyness" noted by Susume above; and a similarity to Rosetta's approach allows us to do some extra analysis on these results). We are currently restructuring the way Foldit handles contact maps in response to these feedbacks, and a lot of these issues will be addressed in the next release.

We've continued to call them Predicted Contacts puzzles because the theory behind them is the same, despite the change in mechanism: the contact map itself is meant to be a visual tool that provides extra clues about the correct structure, and we hope that Foldit players might identify patterns in the contact map that are not directly apparent from simple constraints. Once we settle on a scoring mechanism, I agree that a separate category for Predicted Contacts would be appropriate.

gitwut's picture
User offline. Last seen 3 years 3 weeks ago. Offline
Joined: 05/18/2012
Groups: Contenders

Thanks for the reply. I can certainly understand the time-saving measures, and hope that a modified interface will prove more useful than what we currently have.

I haven't actually noticed the stickiness factor yet. I'm guessing that it is more subtle than the ED effect? I have experienced that (ED stickiness) and agree with the current sentiment that we should be able to turn them on/off as desired since they stick to the wrong things before reaching better positions.

jeff101's picture
User offline. Last seen 8 hours 15 min ago. Offline
Joined: 04/20/2012
Groups: Go Science

http://fold.it/portal/node/996372 requests one-directional bands, like bumpers, which would help in Contact Map puzzles.

Joined: 09/24/2012
Groups: Go Science

1) Would it be possible to have a table of (approximate) minimum and maximum distance to make a contact between different AAs please?

I take as a mean about maximum 8A but I see that it can go up to 12A, and less for contacting glycines (what is the maximum for GLY-GLY ?).

2) Is there any maximum to consider it as a contact? For example, if I have a distance of 2 A, is it still considered a contact in the contact map?

bendbob's picture
User offline. Last seen 39 weeks 4 days ago. Offline
Joined: 05/06/2011
Groups: None

Is there any central location that tells us how and what each of the user interface pieces are? The bands seem to be the most discussed. I have been trying to understand what I'm looking at as a whole. I see black dots. I see green dots (OK squares really, but at the low zoom they are dots on my screen.) I see the idea of a diagonal with the colors for sheets and helices. [BTW 2 different sources for contact maps mention the colors for sheets and helices - the two sources are opposites!]

Having a clear definition of the parts and what they do, if they move around the grid, if, if, if...if.

My issues is what the black squares are for. Am I supposed to use them objectively or are they only informational? If as an object what do I do to interface with them. If informational, what is a black square telling me. Does it matter if the black squares are off the diagonal? Are to make the black squares move inward toward the diagonal.

Having the UI of the contact map selecting and making bands is like whack-a-mole. Is there a method to correctly apply the contact map to the current state of the protein? Perhaps displaying where the black squares are while in the Original UI. So, as we move through the 3-D space the black squares move, or...?

There are many lingering questions that pop-up as I work with this tool and they have not been included in my questions.

jeff101's picture
User offline. Last seen 8 hours 15 min ago. Offline
Joined: 04/20/2012
Groups: Go Science

http://fold.it/portal/node/997598 has more info about Contact Map puzzles.

Also, http://foldit.wikia.com/wiki/Distance_Maps and http://memorize.com/distance-and-contact-maps show Distance and Contact Maps along-side their respective 3D images for many different proteins. Images are classified as in the SCOPe database (http://scop.berkeley.edu), and http://memorize.com/distance-and-contact-maps can shuffle the images in multiple-choice and matching modes to help you learn the patterns in them.

Joined: 09/24/2012
Groups: Go Science


Now the made contacts get a reward of 50 pts. See

The contacts are identified by Foldit, from low distance to high distance (up to 12 A), with a mean of about 8 A. In addition, there is a probability from 0 to 1 that the contacts are existing. Moreover, we heve "long distance contacts" that are possibly "false" contacts: contact from the protein to it's neighbourg protein. The player should delete thes contacts because they don't show the direction to the rightstructure. These "contact to extern protein" are on the surface of the protein. If we have a good starting structure, we an easily delete them (they cross the entire protein), or using tools like BACA to delete the "unmade contacts greater than 13 A" for example.

The problem rised by Loci in his recipe is:
ALL predicted contacts receive a bonus of 50 pts.
EVEN the less probable contacts and of course even the "long distance" contacts to other proteins.
This uniform rewarding system could encourage us to invent unrealistic proteins with forced unlikely but rewarded contacts.

Susume's picture
User offline. Last seen 2 days 20 hours ago. Offline
Joined: 10/02/2011

On puzzle 1304 I am not seeing the contact score going up and down by jumps of 50 pts - I think the contact scores are weighted by likelihood as usual (for example 50 base * 0.5 likelihood = 25 points).

Joined: 09/24/2012
Groups: Go Science

Ah strange. The statistics I see on a solution for 1304:

Number of made contact : 39
Bonus : 1451.5 pts
Max points per contact: 49.9
Mean points per contact: 49.9
Max possible bonus: 5864.6
Min distance of made contacts: 4.5
Max distance of made contacts: 12.2
Mean distance of made contacts: 7.1
Shorter unmade contact: 5.6

It might depend on the way I achieved this result. Not sure it's representative.

It's true that, in the blog over contacts, they say they can play with the bonus (no bonus, bonus, penalty for unmade contacts etc). I suppose that they decide on each case depending on the protein and what they want us to achieve.

Susume's picture
User offline. Last seen 2 days 20 hours ago. Offline
Joined: 10/02/2011

When I sum the *weights* of all my made contacts (not the count of made contacts), and multiply by 50, I get my total contact score.

Joined: 09/24/2012
Groups: Go Science

Ah you are right, sorry ! I didn't use the recipe well to calculate the bonus. Indeed they are weighted. Here are the statistics made with contractulator (on another solution):
Contact made: 40
Bonus: 1523.9
Max points: 50
Max weight: 1

Joined: 09/24/2012
Groups: Go Science

oops sorry, this text isn't correct (see replay by susume)

jflat06's picture
User offline. Last seen 3 days 16 hours ago. Offline
Joined: 09/29/2010
Groups: Window Group

Just as a side note - it might be good to make a new thread on this if we want to continue discussion. Some of the old posts in here are likely to be very confusing to new players, since they deal with the old version of contact puzzles, which were significantly different. Discussion is definitely welcome, though.


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