Developer Chat

10:52:00 katfish Hey everyone, we're starting the Scientist chat in just a few minutes!
10:52:00 katfish Read up on the subject of our chat here:
10:52:00 katfish http://fold.it/portal/node/996679
10:52:00 Susume2 hi kat
10:52:00 TimovdL Hi Kate
10:52:00 Susume2 is it here or in global?
10:54:00 katfish The plan is to do it here. That work for everyone?
10:54:00 silverbergirc yep
10:55:00 Susume2 yes
10:56:00 katfish Awesome! Just few more minutes and we'll get started. Should be a good one
10:58:00 auntdeen2 hi all - katfish, are you back in the country?
10:58:00 frood66 hey all
10:58:00 auntdeen2 hi frood
10:59:00 Susume2 katfish thank you again for giving up such niceties as sleep to help sort things out with newchapter
10:59:00 auntdeen2 omg yes - much thanks
11:01:00 auntdeen2 looks like beta hasn't played in a while, he was just kicked
11:01:00 betahelix I was?
11:01:00 auntdeen2 nope - he's here, lol
11:01:00 katfish You bet!
11:01:00 auntdeen2 kidding you
11:02:00 betahelix you got me... I was worried for a second
11:02:00 porkythepundit hi all
11:02:00 betahelix Hi porkythepundit! Hi everyone...
11:03:00 betahelix ...we'll just wait for the UW scientists to join us.
11:03:00 betahelix Hopefully the UW server issues have gotten better. Anyone notice any problems yesterday?
11:03:00 auntdeen2 so tell us how you are doing in the meantime… much snow?
11:04:00 betahelix Yep, check it out:
11:04:00 betahelix umassd.edu
11:04:00 betahelix they closed the university yesterday and today!
11:04:00 auntdeen2 lol - imagine that
11:04:00 mimi2 snow holiday
11:04:00 TimovdL Looks like we folders send too much snowflakes on the puzzle
11:05:00 betahelix hahahahahha... exactly!
11:05:00 auntdeen2 you're getting a good example of northeast winter
11:05:00 katfish Ok, it looks like we're ready on our end
11:05:00 katfish So it looks like we have beta, Steven, bkoep, ( and I believe v_mulligan) here to chat with us about newchapter
11:06:00 steven hello
11:06:00 katfish Ahh and here's vmulligan
11:06:00 katfish Welcome!
11:07:00 vmulligan Hello!
11:07:00 katfish Alrighty, shall we get started with the questions that were posted in the blog entry?
11:08:00 BletchleyParkirc hi, yes please do. We are all in listen mode
11:08:00 katfish Alrighty, let's start with brow's question
11:08:00 katfish Did your comparisons of players' solutions and methods for prediction or other puzzle types in Foldit, versus R@H, lead to changes in Rosetta?
11:09:00 betahelix Steven, you want to take this one?
11:09:00 steven Sure
11:09:00 steven This is going to go back a couple of years
11:10:00 vmulligan Sorry to interrupt -- should we switch this to global?
11:10:00 steven One of the first designs that Foldit players created, which scored very well in Foldit, was with a bunch of aromatic residues (TRP, PHE, HIS) on the surface of a protein.
11:11:00 auntdeen2 vmulligan: better here, global was full of classes yesterday
11:11:00 betahelix (@vmulligan I think there are less interruptions here in veteran, and we'll post the chat online afterwards)
11:11:00 steven ok
11:12:00 steven At RosettaCon, a conference held for developers of Rosetta, the design was shown
11:12:00 steven Although this was known as a problem with designs in Rosetta, the exploit that the users of Foldit just exmplefied the problem
11:13:00 steven Because of this, a separate score term, known as the reference energies was implemented into Rosetta
11:13:00 steven This, along with several other score terms, seen in the new chapter
11:13:00 steven are suppossed to help control that problem
11:14:00 steven I also think that the players help inspire scientists, not only in the Rosetta community
11:14:00 steven because of Foldit, we have seen a lot of spin off games to help solve other scientific problems
11:14:00 auntdeen2 Steven - a quick clarification.. "the exploit that the users of Foldit just exmplefied the problem"
11:14:00 auntdeen2 sorry to interupt
11:15:00 steven Exploit might not be the right word. Users of Rosetta avoided designing with phe and trp on the surface
11:15:00 steven because we knew that the score function favored those interactions on the surface
11:16:00 steven Players "exploited" the score function by designing proteins with a lot of those interactions on the surface. When they showed this at RosettaCon
11:16:00 steven everyone in the crowd was very interested in the problem
11:16:00 steven which goes back to my inspiring scientists
11:16:00 steven I know I was and that is why I started working on the score function
11:17:00 steven Does that help clarify things?
11:17:00 auntdeen2 so our software has been flawed fundamentally from the beginning, yes?
11:18:00 vmulligan @auntdeen2: Yes and no...
11:18:00 steven I wouldnt say your software...all modelling software
11:18:00 vmulligan One of the big challenges for the scientists is coming up with any sort of means of scoring whether a protein structure is "good" (i.e. realistic) or not.
11:18:00 auntdeen2 ahh - thank you - that makes much sense
11:18:00 vmulligan So as Steven says, all protein modelling software has flaws in its scoring function.
11:19:00 BletchleyParkirc Does Rosetta also use the filters found in Foldit ?
11:19:00 auntdeen2 … how have we managed the few CASP sucesses we have had?
11:19:00 mimi2 we've always known that foldit scoring functions were not "ideal"
11:19:00 steven In a perfect world, we would like to score proteins with quantum mechanics, but that requires a lot of computational power, which no computer can do right now except for very small systems
11:19:00 steven Thats the beauty of foldit players
11:20:00 betahelix Foldit players have uncovered some flaws (such as the pi-helix bug in Rosetta that was discovered thanks to all of you!)
11:20:00 steven Players are able to identify things that computers can not
11:20:00 vmulligan (On the quantum mechanics point: Indeed, the quantum equations can't be solved exactly for anything more complicated than a single hydrogen atom. EVERYTHING must be an approximation to some degree. The trick is finding good approximations.)
11:21:00 betahelix (for example, Rosetta would never attempt to change an alpha-helix into a pi-helix)
11:21:00 steven Also, watching things like the black belt folding is inspiring
11:21:00 brow42 regarding pi-helix, since the hydrogen-bond scoring has changed, this should change how pi-helices are scored. Might the 'fix' for pi-helix now be broken?
11:22:00 betahelix that was one issue that was addressed with the new scoring function!
11:22:00 katfish Alrighty, shall we move on down the list of questions?
11:22:00 brow42 I see, ty
11:22:00 katfish I understand Dawn's questions will be addressed by the devs shortly
11:22:00 betahelix did we answer everyone? BP's question?
11:23:00 Susume2 question to tack on the end of the list: is the new scoring function better at picking good casp solutions?
11:23:00 vmulligan ^^The filters question? The short answer is yes, though the choice of filters tends to vary from task to task (the user can set different filters). We have a few that are very task-specific, that are typically not useful for Foldit players.
11:24:00 BletchleyParkirc thank you
11:24:00 spvincent Are there plans to incorporate quantum mechanics into the score function (thinking of last years Nobel prize in chemistry here)
11:24:00 steven I have not seen solutions for the casp problems with the new score function
11:25:00 steven but Rosetta with the new score function is better at identifing folded close to native proteins
11:25:00 TimovdL Have you for instance tried Rosetta on 663 starting with my winning solution?
11:26:00 betahelix I don't believe anyone has specifically tried that out with the new score function, but that is a good suggestion.
11:27:00 vmulligan @spvincent: There are improvements being made all the time to the scoring function, but true QM calculations would be (a) prohibitively computationally expensive, and (b) probably not worth it. The main inaccuracy probably comes from the fact that we're not explicitly modelling the water surrounding the protein, not from the quantum mechanical nature of the protein molecule.
11:27:00 vmulligan I can say more about why water matters (and why it's hard to model the effects of water) if people are interested, but I don't want to derail the chat.
11:28:00 Susume2 would youconsider a blog post on water?
11:28:00 vmulligan I've been thinking about that. I could write one, sure .
11:28:00 spvincent tx for the reply: perhaps not the best place to go off on a tangent
11:28:00 vmulligan I'll put it in a blog post.
11:29:00 Susume2 :)
11:29:00 SethCooperIRC I'll be posting an update to newchapter shortly
11:29:00 BletchleyParkirc You might also think about: Would a model that excludes gravity produce more accurate results for experiments that are done without gravity ? (growing proteins in a space environment)
11:29:00 SethCooperIRC I've also created a forum thread for update notes
11:29:00 SethCooperIRC http://fold.it/portal/node/996719
11:29:00 spvincent What fixes/changes will there be Seth?
11:30:00 TimovdL Tx, seth that was my next question answered before I could ask
11:30:00 spvincent Ah: reading it now
11:30:00 vmulligan @BletchleyParkirc: Gravity is actually the weakest of the four fundamental forces. It has a negligible effect on protein folding behaviour. (We know this because biological organisms work just as well in freefall as on Earth.)
11:31:00 Susume2 is the scarcity of rebuilds being returned in newchapter intentional, and if so what is the rationale?
11:31:00 spvincent I see nothing there addressing the Rebuild issues we've been encountering.
11:31:00 SethCooperIRC I'll be looking at rebuild and the "frozen terminal" issues next
11:32:00 TimovdL I think those 2 issues are the biggest for us players
11:32:00 spvincent tx. Until the Rebuild issue is fixed it's very hard to modify scripts for NewChapter
11:33:00 spvincent Those that involve rebuild anyway
11:33:00 auntdeen2 Seth - I think that the terminal wiggle lock evaporatd when you posted the triple wiggle - but that end is always orientation, still
11:33:00 jeff101 one place where QM matters is when metals and other ligands are included in the proteins
11:33:00 Susume2 it's also hard to generate a lot of different poses to run wiggle tests against
11:33:00 jeff101 wild type SOD has metals in it
11:34:00 brow42 question: if Newchapter ends up running slower, or slow enough that we only play on lowest wiggle power, what are your plans regarding posting large CASP puzzles? You want to play larger puzzles, but have given us a slower client. Any thoughts?
11:34:00 vmulligan @jeff101: Very true. And that is one thing that we're not modelling well right now. Typically, we fix the geometry of metal centres. In the SOD1 puzzle, I stripped the metals out since we weren't interested in binding to that part of the protein -- but yes, you're right.
11:34:00 steven @brow42 there are ways to speed up the new scoring, but they are very technical in nature. If you would like I can expand
11:35:00 brow42 I would appreciate a technical question later, but for now, please comment on what sort of puzzles we can expect for CASP
11:35:00 brow42 *explanation
11:37:00 vmulligan (@jeff101: A better framework for modelling metalloproteins is actually high on the agenda for the developers' meeting in February. That would affect Rosetta before it affects Foldit, though.)
11:37:00 betahelix jflat has been working on a way to let you work on larger puzzles... sadly he is sick today and couldn't make it to the chat
11:37:00 steven Sure! I would love to give an explanation, maybe in a follow up posting is the best way to address that
11:37:00 TimovdL I lost connection, so I resent my question. Can we get an overview of active score parts in newchapter depending on the puzzle type?
11:39:00 BletchleyParkirc Is the score function a serial process or can parts of it be calculated in parallel ?
11:39:00 steven @timovdL are you asking about the new type of scoring terms?
11:40:00 SethCooperIRC the score parts should generally be the same
11:40:00 SethCooperIRC however, there is a new one dealing with ideality
11:40:00 brow42 I believe some players have noticed a new bonding score type, even when the residue has no apparent bonds.
11:40:00 vmulligan @BletchleyParkirc: It is very parallelizable. Interestingly, Foldit is better parallelized than Rosetta. (That's one thing that we're actively working on in Rosetta right now.)
11:40:00 SethCooperIRC right, there was a "cartesian bonding", now called "ideality"
11:41:00 BletchleyParkirc OpenCL / CUDA ?
11:41:00 BletchleyParkirc (@vmulligan)
11:42:00 TimovdL Yes and also the other term now has weight
11:42:00 steven Part of the new scoring includes a thing called the analytical etable, which can be used to parralize the atr, rep, and sol terms in Rosetta
11:42:00 vmulligan @BletchleyParkirc: Foldit currently only uses the GPU for the graphics, not for the score function calculation. We have one person in our lab who has done some development work on GPU processing, and there are some specific Rosetta protocols that take advantage of it, but for the most part, most development work has been for CPUs.
11:42:00 steven I believe the plan is to dump the calculations of atr, rep, and sol onto a gpu
11:43:00 bkoep @brow42, I think our general feeling is that we would rather see improvements in accuracy than in number of residues modeled. We also hope to run some CASP targets this year with predicted contacts.
11:43:00 BletchleyParkirc @vmulligan: I have done work on primegrid in CUDA, that is why I ask.
11:43:00 brow42 ty, bkoep
11:44:00 steven @bletchleyParkirc, I believe we use opencl as CUDA limits you to nvidia
11:44:00 BletchleyParkirc correct
11:45:00 vmulligan There were a few people who asked good questions a while back. Let's go back and field a few of those.
11:47:00 vmulligan Oh -- looks like SethCooperIRC fielded those.
11:47:00 vmulligan OK, any more questions?
11:47:00 betahelix katfish, there was Dawn's
11:48:00 katfish I believe Seth can address those
11:48:00 betahelix which was the next one down the list, right?
11:48:00 betahelix http://fold.it/portal/node/996679#comment-25898
11:48:00 katfish SethCooperIRC?
11:48:00 vmulligan He's just reading it over .
11:50:00 SethCooperIRC so for that comment
11:50:00 SethCooperIRC with respect to #1
11:51:00 SethCooperIRC current wiggle would do different things on different iteratiosn
11:51:00 SethCooperIRC so, if you ran it for 2 iterations, it might try different things that running it for 1 iteration twice
11:52:00 SethCooperIRC the version of wiggle in newchapter now does the same thing every iteration
11:52:00 TimovdL How manjy different parts are there?
11:53:00 brow42 that is important to know, ty
11:53:00 Susume2 and yet in newchapter 10 loops of 2 iterations is outscored by 20 iterations
11:53:00 SethCooperIRC that is also correct
11:53:00 SethCooperIRC or, that would make sense
11:53:00 MikeCassidytoo why?
11:54:00 MikeCassidytoo if its doing all the same
11:54:00 BletchleyParkirc 10A+10B is not equal to 20A
11:55:00 TimovdL 10A+10A should be
11:56:00 MikeCassidytoo in newchapter now does the same thing every iteration does not this imply A = B?
11:56:00 BletchleyParkirc old wiggle = 10A then 10B, new wiggle is 20A
11:56:00 BletchleyParkirc if A scores, but B doesnt then new wiggle outscores old wiggle
11:56:00 Susume2 not comparing ols to new, comparing 10*2 in new to 20 in new
11:56:00 Susume2 *old
11:57:00 SethCooperIRC basically, as you run wiggle it can construct sort of an estimate of the scoring landscape
11:57:00 vmulligan As wiggle runs, there is a particular quantity that is estimated, and the estimate improves the longer it runs. Whenever you restart wiggle, that value is reset. (For those who know multivariate calculus, the Hessian matrix is being estimated from the gradient, and improved with each iteration.)
11:57:00 SethCooperIRC if you stop wiggle and restart it, it resets that estimate
11:58:00 BletchleyParkirc thanks for clarifying, I was wrong.
11:58:00 Susume2 that means checking periodically if wiggle has stabilized yet reduces the work it will do; so we need to know for our scripts abouthowlong wiggle takes to stabilize, and change the hard-coded iterations every time wiggle changes
11:59:00 TimovdL Can Wiggle be made such that 2 subsequent calls without any intervening potentially protein changing calls it will still use the matrix?
11:59:00 porkythepundit perhaps a function of fetching the old estimate could be included?
11:59:00 TimovdL Or an extra boolean parameter
11:59:00 Susume2 or we could assume, or demonstrate, that wiggle is still choosing good poses even when it is doing less work
12:00:00 jeff101 or output the Hessian to a variable in LUA and let it be input into the next wiggle command
12:00:00 brow42 it may be simpler to implement a call-back into Lua instead of cacheing the wiggle state
12:00:00 SethCooperIRC we had also discussed adding an additional score change threshold argument to the lua wiggle function
12:01:00 vmulligan @TomovdL: I like that idea (caching the Hessian). The one embarassing thing is that our code isn't set up to do that easily.
12:01:00 vmulligan We'd have to reset it if the protein were changed manually, though.
12:01:00 TimovdL That is easy, check the score and the banding
12:02:00 porkythepundit yes
12:02:00 SethCooperIRC also for the blog comment
12:03:00 SethCooperIRC it seems #2 is addressed by the wiggle power option?
12:03:00 porkythepundit would such extra functions impact the speed significantly?
12:03:00 TimovdL But for now I hope that the devs are concentrating on the rebuild and other problems we found
12:03:00 Susume2 @Seth the score change parameter might do waht we need
12:04:00 TimovdL Problem with that parameter is working not on CI 1
12:04:00 vmulligan (@porkythepundit: If you mean caching the Hessian, it would slightly increase memory usage, but shouldn't have a huge impact in speed, I'd think. It's doable, but as TimovdL said, the current priority is getting NextChapter ready to roll out).
12:05:00 SethCooperIRC and for #3, we have changed an iteration to not mean convergence, but rather a unit of work
12:05:00 Susume2 @Seth # is not addressed with reference to iterations, althoguh wiggle power seems to solve the speed problem. We still have a large number of scripts with hard coded numbers of iterations that will not operate as intended
12:05:00 SethCooperIRC so all the points that would be found on the first iteration before are now spread out across several iterations
12:07:00 TimovdL On Low it looks like 2 iterations are enough for near convergence
12:07:00 katfish Ok everyone, we need to start wrapping things up. Last call for comments and questions!
12:08:00 robgee how much better will it be, when newchapter works.
12:08:00 jeff101 centroid vs full atom mode ... any fixes/updates coming soon?
12:08:00 jmbrownlee333 can u repeat what the current timtabele plan is
12:08:00 spvincent I feel that overall the changes in wiggle will be positive: allows for much greater control oin scripting rather than having to guess at the number of iterations.
12:08:00 spvincent But I would like to see the Rebuild issues prioritized.
12:09:00 Susume2 I think we still have to guess at the number of iterations
12:09:00 SethCooperIRC the current plan is still to release newchapter this coming Monday
12:09:00 auntdeen2 a quick explanation of how the different 3 wiggles work?
12:09:00 SethCooperIRC currently rebuild, frozen terminal are my priorities
12:09:00 bkoep @jeff101, after newchapter is up and running, we will release another centroid puzzle with multiple improvements since the last centroid puzzle. We just wanted to make sure all is good in newchapter first.
12:10:00 betahelix @jeff101 that is one of the puzzles that is waiting on the newchapter update
12:10:00 auntdeen2 the 3 wiggle differences are vital to understand, so that players with older computers can figure out how to play...
12:11:00 vmulligan @auntdeen2 and @SethCooperIRC: Would a blog post about the 3 different wiggles be doable? Then it might be more accessible than if it were buried in the transcript of this chat.
12:11:00 auntdeen2 that would be fine
12:11:00 betahelix with figures!
12:11:00 bkoep @auntdeen2, in short, Med and High power wiggles allow small changes in bond lengths and bond angles (which are kept constant in Low power).
12:12:00 katfish Cool, so we will have a few new blog posts for you to look forward to!
12:12:00 TimovdL More general, we would like a list of the biggest differences in newchapter. Especially if they can influence scripts
12:12:00 brow42 bkoep, that is very important to know, people would never be able to fix their cutpoints on low power, it sounds like.
12:12:00 auntdeen2 thanks vmulligan & bkoep
12:13:00 MikeCassidytoo As far as I can tell High Wiggle crashes when used in rebuild
12:13:00 bkoep @brow42, correct, Med power wiggle should be sufficient for fixing cutpoints
12:13:00 SethCooperIRC correct, at low power, cuts will not be fixed
12:13:00 SethCooperIRC and in general idealization will not be dealt with
12:14:00 SethCooperIRC (at low power)
12:14:00 auntdeen2 will there be a way to script the powers?
12:14:00 SethCooperIRC we will likely add scripting support for wiggle power, yes
12:15:00 SethCooperIRC it may not happen before Monday though
12:15:00 mimi2 ny eta?
12:15:00 mimi2 any
12:15:00 auntdeen2 wait - then how can any scripts be done? what is the default?
12:16:00 MikeCassidytoo You click which you want run the script
12:16:00 SethCooperIRC the default it whatever it is set to in the behavior options
12:16:00 SethCooperIRC so you can run the same script with different powers
12:16:00 MikeCassidytoo I have been doing that since the powers were added
12:17:00 katfish Ok everyone, it's time for us to sign off for the day. Thanks for coming out!
12:18:00 katfish I'll have the transcript up on the website later today -- promise!
12:18:00 vmulligan And thanks for the interesting questions!
12:18:00 SethCooperIRC thanks!
12:18:00 bkoep Thanks for the questions, everyone!
12:18:00 jeff101 thanks
12:18:00 mimi2 thanks for the respones
12:18:00 auntdeen2 thanks
12:18:00 katfish Take care everyone!
12:18:00 MikeCassidytoo Thanks for the discussion
Get Started: Download
  Windows    OSX    Linux  
Windows
(7/8/10)
OSX
(10.7 or later)
Linux
(64-bit)

Are you new to Foldit? Click here.

Are you a student? Click here.

Are you an educator? Click here.
Search
Only search fold.it
Recommend Foldit
User login
Soloists
Evolvers
Groups
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, RosettaCommons