6 replies [Last post]
Joined: 11/14/2009
Groups: None
Time Name Message
12:32 David Hello everybody this is David Baker
12:33 spvincent Hello Dr. Baker
12:33 MellyLam Hello!
12:34 Lazarus -uk hi David
12:34 Christoph hi!
12:34 CharlieFortsConscience :) evening
12:34 David how are things going with the CASP targets?
12:35 Aotearoa omg this is so awesome
12:35 CharlieFortsConscience it's a lot of hard work
12:35 Lazarus -uk you are probably in a better position to tell us how it's going Dr.B
12:35 Aotearoa they have been challanging for sure
12:36 David we are very excited about the mini-CASP results you came up with!
12:37 Aotearoa phew ! i was worried about those as we hadnt had a lot of feedback, so were getting the hang of it?
12:37 spvincent In the R@h forums some time ago you said that people were better at folding than 1 computer but not as good as 40,000 (or something).
12:38 Lazarus -uk hopefully 'all-hands' will be able to refine some of the solutions still further
12:38 David there certainly are cases now where foldit players are doing better than could be achieved even with very large amounts of computing power
12:39 David "all-hands" could be very powerful, but it also could have a pitfall that everybody needs to know about
12:39 spvincent What would that be?
12:39 David in general, we see players sometimes "piling up" too much at the best current solution
12:40 CharlieFortsConscience *walked to death*?
12:40 Aotearoa it's been great to see the new folders performing so well. aap, cfc and Keypad5 ... so many new ones have realy changed the way we fold.
12:40 Lazarus -uk we may be refining the wrong solutions
12:40 mat747 cfc is right
12:40 David it is like gold-mining when everybody starts searching around the best current find
12:40 David but the mother-lode is somewhere else which hasn't been discovered yet
12:40 spvincent Perhaps you could have them as sepearte puzzles then
12:41 Aotearoa more re-starts then david?
12:41 David so in all hands you need to consider not only the best current solution, but the full range of solutions presented, and also keep exploring on your own
12:41 Lazarus -uk is there any way your team could search out the most likely candidates, rather than the highest scoring ones
12:41 beta_helix Aotearoa: yes... we are also excited about the new crop of top folders. As for the mini-CASPs, David already mentioned this great case where we were impressed by how you all used the alignment tool:http://fold.it/portal/node/987564
12:43 David the real problem is that computers currently aren't good at detecting which are the most likely candidates.
12:44 David once you get very close to the solution, the computer can tell because the energy is very low (the score is very high).
12:44 Aotearoa the alignment tool does throw up some crazy template configurations though, with prolines in the middle of sheets and helix's. i dont know if leaving them there is good or bad
12:44 David one of the hopes is that people could be better at deciding what is a likely candidate
12:45 David we could try to decide with are the best candidates, but one of the goals of the casp foldit experiment is to try to avoid "expert" intervention as much as possible
12:46 Lazarus -uk that may be where scoring actually becomes a problem. pulling everyone towards the points and away from the correct solution
12:46 CharlieFortsConscience So we're still ahead in the 'it doesnt look right' factor...
12:46 Lazarus -uk so, we just have to do our best
12:47 David yes--you are absolutely right. the computer has this problem even worse, though--it has no intuition about where things might go so it has to be greedy
12:47 spvincent I sometimes wonder if these scripts we're creating are doing no more than a very bad and ineffecient Rosetta@home
12:48 David I think your scripts and approaches are much more interesting than that. you are coming up with new algorithms that we never imagined
12:48 spvincent Good!
12:48 BletchleyPark good evening David, do you see a role for GPU computing within foldit ?
12:49 David yes-you are very far ahead in the "it doesn't look right" factor! and hopefully as you get more experience this will just keep increasing
12:49 CharlieFortsConscience We react to the constraints placed upon us, and find shortcuts that make sense to our style of play...
12:50 David I think GPUs could play a role--it is really a question of manpower. we are hoping NVIDIA will help us out here...
12:51 Aotearoa we sure could do with a new influx of folders and material we know is "correct" to teach them more
12:51 beta_helix so more quest to the natives?
12:51 Aotearoa *omg please NOooo!* grin.
12:52 spvincent I don't find those puzzles much fun
12:52 mat747 me too
12:52 aap2 i kinda like them
12:53 David we could set up simple puzzles where you are given some number of models including the native and just have to pick out the native? these would probably be easy and not time consuming.
12:53 CharlieFortsConscience Hmmm... FoldClub has changed so much in the last year, more longer lasting beginner puzzles would help keep newbs around. It's a lot more technical, and scary than it used to be
12:53 Aotearoa That sounds good
12:54 David what do you think of the design puzzles?
12:54 spvincent That would be fine. I think you did that once.
12:54 CharlieFortsConscience :) I personally like them. A lot.
12:54 MellyLam good point CFC. The puzzles are a lot bigger and slower than they used to be, and they take a long time to work out
12:54 Aotearoa design puzzles GREAT David, but we still have to change our limited thinking to research topics and interests to make them stronger. but we need more and more
12:54 spvincent I don't really know I'm doing with them and end up trying things at random
12:54 Aotearoa *pletsch and I would like to use fold.it and take on some projects.*
12:55 BletchleyPark @David: are there collaboratins between the various ' folding' initiatives ? (GPUGrid, Rosetta, FAH come to mind)
12:55 beta_helix we can post beginner puzzles with natives to get new folders to learn the alignment tool (and for you all to practice with them?)
12:56 David to get a feeling for protein interactions and binders, it could be useful to browse the protein structure database which has a free viewer
12:56 Aotearoa Yes Beta !! great idea.
12:56 beta_helix here is the protein structure database for anyone who hasn't been there: http://www.rcsb.org/pdb/home/home.do
12:56 Aotearoa i've noticed more feedback from beginner puzzles than competition puzzles
12:56 aap2 i think i learned alot from the qttn puzzles
12:57 David there are some collaborations on the science side. for example, we are collaborating with some of the FAH researchers now on looking at protein folding kinetics using Rosetta and Rosetta@home
12:57 beta_helix our hope was that by seeing the native, you would be able to see what moves are needed using the alignment tool
12:58 BletchleyPark that sounds like a good initiative to me, they also have GPU experience. GPUGrid sells / licenses its GPU engine
12:58 David we are very excited now about our designed flu virus binder.
12:59 Christoph have you already tested it in the lab?
12:59 Aotearoa GREAT !!!
12:59 Aotearoa Vakobo are you here?
12:59 David we are starting up projects to design binders to other pathogens, and after CASP we hope to get all of your help with this!
12:59 Aotearoa <--- I'm in.
12:59 spvincent Will the CO2 fixation stuff be in FoldIt too?
13:00 David we know that in a test tube the designed protein binds tightly to the flu surface protein. we don't know yet whether the protein can actually block flu infection
13:00 Aotearoa question: do out puzzles on their "paths" to final protein shapes, get included in your final results.
13:01 David we can set up the CO2 fixation problems as foldit puzzles as well
13:01 spvincent Neat
13:01 BletchleyPark @David: is there a limit on how big the fold.it proteins can be ?
13:02 David the only size limit is that manipulations get slower and
slower , as you know. we are working on speedups now
13:02 BletchleyPark very good to hear ! I notice that foldit uses a lot of ' ordinary' floating point math, might be good to make it SIMD aware ?
13:03 beta_helix Aotearoa: those in the Baker lab working on design look at all the foldit solutions, filtering them using different methods
13:03 Aotearoa ty beta: I ask that question because I can go through many different combinations which are low scoring shapes/combo's and various designs but "dont" save them as I go through a range of differing shapes.
13:03 David we are now trying to figure out how to get bacteria to grow and make fuel molecules from solar generated electricity and CO2
13:04 David this would be the ultimate in "clean energy"
13:04 Aotearoa I saw some Ideas from BMW ag in Munich when I was working there David
13:04 David the first challenge is to figure out what enzymes are needed, then we can set up design puzzles to make them
13:04 spvincent Something like that's going to be really needed if the world's going to be fit tolive in.
13:05 CharlieFortsConscience We hear rumours that Nature will publish soon... how is that going? Back from review yet?
13:05 CharlieFortsConscience We hear rumours that Nature will publish soon... how is that going? Back from review yet?
13:05 David yes--it is a super hard problem but something like this has to be done!
13:05 David I expect to hear from Nature any day now--you will be the first to hear!
13:06 Aotearoa *my goat will be very happy to hear that*
13:06 Aotearoa *lol j/k*
13:06 David It is looking like they will do a news article in their front section about game playing and then have the scientific article describing your amazing results in the main part
13:07 David we are getting a huge boost for casp and inhibitor design now by the recent increase in participants in rosetta@home--I hope it lasts!
13:08 BletchleyPark Nature article may spark interest of the Discovery Channel, that would reach quite a few people who may turn out to be players
13:08 David yes-that would be great!
13:09 Aotearoa we realy need a few of us to stop folding for a while and start writing how to use the new interface and how to fold using aap kp5 and the other new folders methods dont we. we are starting to get behind in the wiki updates
13:09 David back to CASP--for problems where the alignment is less good (fewer identical residue pairs), definitely keep in mind that the alignment could be wrong
13:09 David particularly in edge beta strands which are frequently misaligned.
13:09 David for these it could be good to shift left and right one residue at a time
13:10 beta_helix Aotearoa, after CASP (if you all aren't too burnt out) that would be great
13:10 spvincent Do you mean that these are good candidates for register shift?
13:10 Aotearoa I would like to sugggest a new tweak view and tool for helping with that ! will write it in feedback sooner than later.
13:10 BletchleyPark You could send a copy of Nature with some background info to Pioneer productions and see if it sparks interest with them
13:10 David yes! I would focus register shifting on the edge beta strands. doing it at the alignment level shoudl be particularly powerful
13:10 BletchleyPark They do a lot of features for DiscC
13:10 Christoph instead of trying out different alignments, couldn't we simply rebuild those regions?
13:10 Christoph at the end of strands
13:11 David that is worth trying as well-but I would first explore different registers
13:11 CharlieFortsConscience I keep using the phrase 'alignment is just a gift-wrapped head start' - surely the pt gains come with the post threading rebuilding?
13:11 David remembering that regions with strong sequence identity should not be altered
13:12 BletchleyPark @David: is the alignment score holy ?
13:12 David yes--the posthreading is critical-BUT you need to start with the right alignment
13:12 David no--it is not holy. treat it as a rough indicator
13:12 Aotearoa im having problems counting the segments to move the right ones
13:12 BletchleyPark ok, t hanks
13:12 Aotearoa numbers at the top would be nice
13:13 David this is like being in my lab and talking about ten different research projects at the same time!
13:13 beta_helix Aotearoa, we will add numbering to the alignment tool soon!
13:13 CharlieFortsConscience Is it worth then, pursuing a highest thread score?
13:13 Aotearoa *ty beta*
13:14 David it is probably worth trying different alignments, quickly fixing them up, and then making a call about which are worth pursuing
13:14 David maybe the different alignment shifts could be divided up among team members?
13:15 Aotearoa once we have threaded a protein, then made changes. can the Thread (top line) then reflect those changes ?
13:15 Aotearoa *reverse engineer*
13:15 CharlieFortsConscience *...quickly, is a judgement call, with us Luddites on slower technology... but I see what you're suggesting*
13:15 David computers can't solve the problem because again they can't "look ahead" and determine what are the most promising avenues to pursue
13:16 BletchleyPark Has alignment scoring been studied or is it based on assumptions ?
13:16 David yes, it has been investigated, but all scoring schemes are approximations. the reality is the energy of the 3D structure
13:17 BletchleyPark ok
13:17 David I meant "alignment scoring schemes are approximations"
13:17 beta_helix Aotearoa, that would be a bit more complicated... but an iterative thread tool would be a cool idea
13:17 BletchleyPark I understood it that way, thank you
13:18 spvincent Is the energy function finalized or are are small improvements continuously being made?
13:18 David we are constantly working to improve the energy function.
13:18 David we are making slow but steady progress I think
13:18 Aotearoa each puzzle seems to reflect that score differently spvincent.
13:19 Aotearoa its way better than its ever been david
13:19 Christoph are we getting the improvements made to rosetta and rosetta@home in foldit too in regular updates?
13:19 David yes. everytime you have to update foldit you are getting the latest improved version
13:20 BletchleyPark Are the critical routines of the foldit software hand-optimized or is this compiler code ?
13:20 Aotearoa *geek alarm*
13:20 BletchleyPark No, pro alarm
13:20 Aotearoa lol *j/k*
13:20 BletchleyPark This is my profession
13:21 David the core algorithms are all the Rosetta C++ code, similar to what is running on rosetta@home
13:21 BletchleyPark is that open source by chance ?
13:21 David the rosetta code is free to everybody not trying to make a profit
13:22 BletchleyPark great, I'd love to improve it
13:22 phryll yeeeah boooy
13:22 David we charge companies licensing fees, and use this to support the rosetta developer meetings
13:22 phryll :D
13:22 aap2 what if you just happen yo make profit? :P
13:22 Aotearoa BP - pletsch and I are looking for work. <--- no qualifactions but we will fold you anything from cakes to alien lifeforms for you or your clients. FYI.
13:22 BletchleyPark shirts too ? :)
13:22 beta_helix any last CASP related questions for David?
13:23 CharlieFortsConscience I mentioned earlier, FoldClub has experienced a huge evolution in the last year - what plans do you have for us in the future? Any blue sky plans you can tell us, or are we in a consolidation phase...
13:23 phryll ARE YOU N.E.R.D.S.?
13:23 David companies have to pay
13:23 BletchleyPark can we be fed one full CASP puzzle (a big one) ?
13:23 Aotearoa I would just personally like to say, Thank You David Baker ! you have realy had a big impact on my life.
13:23 Christoph there's one with 800 residues I think..
13:24 BletchleyPark I'd like a go at that, just to see where we can take it
13:24 David things are happening so fast with rosetta and foldit it is hard to predict where things will be in a year
13:24 Christoph http://predictioncenter.org/casp9/target.cgi?id=49&view=all
13:24 Aotearoa I'de like to try them all too. even if we do fail, good to make a massive helix and band it up.
13:24 beta_helix the 887 residue CASP9 target that was released yesterday is not a single domain
13:24 phryll IMAGE:
13:25 David big proteins we chop up into domains.
13:25 David the problem with really big domains is that for most people's computers they would be painfully slow
13:25 BletchleyPark maybe as a test puzzle ?
13:26 Aotearoa Asteria is my new wife (computer) She's keen to try some funky stuff.
13:26 CharlieFortsConscience *phryll, restart and follow the tips closely. And welcome to FoldClub*
13:26 David there are so many targets we don't want to waste your time on problems which would be very frustrating
13:26 aap2 perhaps a very big puzzle that doesnt get calculated into the global score?
13:26 beta_helix CFC, I think we lost phryll
13:26 spvincent I doubt if my Mac Miini could handle that.In fact I'm sure it couldn't
13:27 CharlieFortsConscience *shame*
13:27 BletchleyPark I think we're keen on bumping into that performance waal just to see where it is.
13:27 Aotearoa true. we can do that when we've found more cures for things for biofuels and H1N1 important things.
13:27 BletchleyPark wall
13:27 David I always hope that what we do each year would have looked "blue sky" the previous year-one reason I don't like predicting the future!
13:28 David we could post big problems as test puzzles to test equipment on
13:29 BletchleyPark I'd welcome that
13:29 Aotearoa we need a TEST area for crazy stuff like that - perhaps on BETA login not MAIN
13:29 beta_helix we could put a huge CASP target in BETA for you to try out...
13:29 Judecca im getting a new laptop soon so i would be interested in seeing how well it handles a larger protein. if anything just to benchmark it
13:29 David I'm getting hungry--any last questions before I go find something to eat?
13:29 beta_helix exactly, Aot, good call
13:29 mat747 david - i think some video`s on the different subjects in wiki would help, like register shift , things we are doing well on and not some well that. things you would like us to try .
13:29 BletchleyPark @beta_helix, please do, sounds interesting
13:29 beta_helix will do!
13:29 Judecca beta would be a good idea for that, heh maybe add an option to show FPS
13:30 aap2 some visual advice/input would be great
13:30 beta_helix we can easily add FPS
13:30 Aotearoa What are you going to Eat David? your own lab made proteins or something like KFC ?
13:30 spvincent Written stuff works better for me : not videos
13:30 Aotearoa *don't answer that btw*
13:30 David mat: we will be watching and try to give advice and pointers as things move ahead
13:30 CharlieFortsConscience ...we grumble, and we rank up occasionally, but we keep coming back every day. Thanks UW, for keeping us sharp.
13:30 mat747 david - thanks
13:30 David great to talk with all of you!!
13:31 BletchleyPark Many thanks for your time !
13:31 David bye!
13:31 aap2 bye
13:31 mat747 spvincent - both would be good
13:31 spvincent true
13:31 Aotearoa yeah !
13:31 mat747 bye - david
13:31 Aotearoa I waited for two years for that.
13:31 Aotearoa today is the best day of my life.
marie_s's picture
User offline. Last seen 1 year 49 weeks ago. Offline
Joined: 05/18/2008
Groups: None
all ends

you could change the scoring: take a complex formula that takes into account major impacts on low-scoring proteins but also small progress on proteins rated high: a combination of % growth, progression and absolute final scores. The time we guess the algorithm, we will work on all solutions.

Joined: 09/18/2009
Groups: SETI.Germany
And...

Maybe it is really better to launch multi-start-puzzles not in a single puzzle, but like the CASPs as a,b,c,d.
Many scripts use the reset/restore (very)best functions, and it is hard to handle with another starting position if you already have brought one of them to high score.

beta_helix's picture
User offline. Last seen 5 days 8 hours ago. Offline
Joined: 05/09/2008
Groups: None
too many puzzles?

Crashguard, the reason we stopped doing a,b,c,d puzzles (and instead created the new alignment tool where you can load the different templates yourselves) is because in the other 2 chats players mentioned that there we too many puzzles already.
Right now we have 3 puzzles up, 1 template and 2 denovo CASP targets. If we gave you a puzzle for each denovo Rosetta model, that would be 11 puzzles up right now!

Joined: 09/18/2009
Groups: SETI.Germany
Okay.

This makes sense.
Then it would be helpful to clear the "very best" result.

As Dr Baker said:
"in general, we see players sometimes "piling up" too much at the best current solution"
[...]
"it is like gold-mining when everybody starts searching around the best current find but the mother-lode is somewhere else which hasn't been discovered yet"

So, if you play a multi-start or all-hands, and you've reached high score, you can't go back to the middle, only by loading another solution.
Or ... you can, but if you did some mistake at tweak/nudging, you can't do a reset.

Let me give an example:
You have played a start position through, resulting a score of 11,000.
You take another start position and improve it, resulting 10,750 points.
Then, you do a wiggle, get 10,800p and try to nudge at a sidechain, but it is the wrong one, making the score go back to 10,700.
Now you can't reset to the 10,800 state, because "restore very best" would take the solution with 11,000 points. It is unusable in the second turn.
If you had the abilitiy to clear the "very best" result at taking another start position, this won't have happened.
Sure, there is the possibility to use the undo bar, but it only has a limited amount of states, vanishing fast if you use some scripts.

beta_helix's picture
User offline. Last seen 5 days 8 hours ago. Offline
Joined: 05/09/2008
Groups: None
will do

That makes sense Crashguard.
You would then save your 11,000 point solution, "clear very best" and basically start over, right?

We will try to implement this soon.

Until then, hopefully you can use the quicksave slots for that 10,800 solution before trying a nudge or something.

Joined: 09/18/2009
Groups: SETI.Germany
Yes...

That's right!
At the moment, I use regular save and quicksave slots indeed, to try another starting position or other alignments.
Maybe it's best to be asked to clear the "very best" result if you reset the puzzle itself, this would give not so much problems.

Thanks in advance!

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