Thu Jul 22 10:50:06 PDT 2004 T0248 DUE 26 Aug Thu Jul 22 14:18:09 PDT 2004 Kevin Karplus No strong hits in t2k.best-scores, and only slightly better in t04. Probably a new-fold or difficult fold recognition target. Thu Jul 22 17:29:12 PDT 2004 Kevin Karplus try1-opt2 has a hairpin and a bunch of helices, but they are not well-packed and some of the helices are broken. There are not many RR constraints, and they are pretty weak, but we might as well include them for try2. The t04 alignment seems to have somewhat crisper conservation than the t2k one, so I'll favor it. Thu Jul 29 15:00 PDT 2004 Kevin Karplus try2 looks pretty good up to about P77---after that the helices are not well packed. Fri Jul 30 05:19:08 PDT 2004 Kevin Karplus try1 looks ok at 2ry level, but has poor packing of helices. The rr predictions, though few, may be useful. (I had thought that try1 had been redone, but it seems that the redos started wtih T0149, so these comments may be redundant.) Tue Aug 17 6:00pm Jenny Draper I've also created a superposition of the best t04 alignments: T0248/T0248.t04.undertaker-align.pdb they have some good stuff, though the first two hits are to odd, stretched-out structures (the ones robetta's using). Since we no nothing at all about this protein, and the e-values are, at best, e^-1, i don't see why to believe the odd structures; I'll try for a globular protein... The region from 1-70 in try2 definatly needs some help. I'm going to start a try3 from alignments, using a cost function looser in secondary structure constraints than try2, but with higher weight on pred_alpha_t2k, because the alpha predictions for t2k agree pretty well with the str2 preditions for t04 & t2k (but pred_alpha_t04 is different). I'll also up near_back and way_back, since this is a helix-packing job. I'll also add in the full set of alignments for the two best alignment _models_ that I've seen: 1ku2A/ #nice alignment structure, t2k, model 5 1fkmA/ #nice alignment structure, t04, model 3 Note: t2k best-alignment models 6,7,8 have good alignments to the region 1-70 which is messed up in try2. Wed Aug 18 12:15am Jenny Draper Try3 crashed for some reason (the log file just stops, mid-run) before opt2 could complete. Which really sucks, because opt1 looks pretty good. definatly foamy, and the helix from 80-100 needs to be turned, and the hairpin needs a little tweaking to form the sheet -- but I like this better than try2. I'll re-start try3 as try4, using the same costfunction; hopefully, it will make it... Wed Aug 18 8:15am Jenny Draper Damn. Try4 finished, but it doesn't look nearly as good as try3-opt1; it's got this weird t-shaped structure with a diamond of helices as the base. I suppose it could be fixed, but I think I'll start try5 to polish try3-opt1. Th Aug 19 8:15am Jenny Draper Hmm. Try5 has been blown up a bit, and jumbled. It's definately not packed better hydrophobically, either. It's not improving try3... what am I doing wrong? Sat Aug 21 3:15pm Jenny Draper Try6 is just more of the same jumbled mess as try4. It looks like try6 is getting all of it's score improvements from constraints, so I think I'll drop the overall constraint weight down, and up the break cost. Sat Aug 21 10:45pm Jenny Draper Damn; forgot to save my notes in the README, so when I quit the editor to refresh this morning, they wre gone. Last night, I ran try7/try8/try9, all using the same cc'd try7 cost function, which was generated in an attempt to get try3 to score at the top; the best i was able to manage was try5>try6>try3. In the process, I decided that try5 and try6 weren't really all _that_ bad, they just lack the nice core that I saw in try3. So, I decided to try a run off of each of them: try7:try3, try8:try5, try9:try6. Sun Aug 22 11:15am Jenny Draper Damn. All three were essentially a waste of time. try8 and try9 destroyed the hairpin; try7 went off in some funky direction an did some damage to the hairpin. Try1 looks way better than any of these. I think I need to call in Kevin to work some magic on the .under and .costfcn scripts. Sun Aug 22 1:00pm Jenny Draper Well, with a lack of better ideas, I'm going to start try10 from all conformations, using the try3 basic cost function, which scores try1 > try4 >try5 > try7.... Sun Aug 22 17:43:52 PDT 2004 Kevin Karplus Jenny asked me to look at T0248, because she was feeling frustrated with it getting worse rather than better (certainly a feeling I've had oten enough). Refreshing my memory: this is a mostly helical protein (one hairpin in the middle at K187-Q200) and fairly long (294 amino acids). The t2k and t04 multiple alignments have rather different conservation patterns, with the t04 alignments have much greater diversity. There is only one strong rr constraint: Constraint L236.CB I262.CB -10. 7.0 14.0 0.752 These are on two adjacent helices, but would be difficult to bring together. I made an unconstrained.costfcn, and it likes try2-opt2.repack-nonPC best, despite ther rather ugly "handle" made from N-terminal helices. If those helices were broken up a bit and packed down against the main body, this model would not be so bad. There are some rather awkward helix breaks (like between D59 and F60). My earlier comment "try2 looks pretty good up to about P77---after that the helices are not well packed," is reversed---it is AFTER F60 that things look better. Next best try is try10-opt1 (at the moment---try10 hasn't finished). There is a helix unwinding at V55-D59 in mid-helix, but after that looks pretty good. If we can find a model with a good beginning, we may be able to superimpose on residues F60-A63 and patch this (or try2) with something that works better. I thought try5-opt2.repack-nonPC might be a good choice, but it messes up the helix from K58 on. Wait a sec---the t04 alpha prediction is for there not to be helical conformation there anyway. Ah---the color scripts are still set up for t2k, not the preferred t04. I'll re-do the make, to be sure that the coloring scripts are the preferred ones. It looks like Jenny has been using rather old cost functions that give too much weight to the sidechain term, and not enough for the burial terms. She also hasn't been commenting out the PrintTemplateAtoms line, so has been outputting Template.atoms files on each run. Ahh---and there was no T0248.t04.many.frag.gz file, so she's been running on an incomplete set of fragments (just the t2k ones). Maybe I should do an optimization run with the unconstrained costfcn, either from alignments or polishing the existing models, to see if we can do any better. Still, try10-opt2 has a good shot at being the best even with the newer unconstrained costfcn, since try10-opt1 is doing so well. With inconsistent predictions between t02 and t04, we may have to just present a variety of reasonably-packed models, without really trying to figure out which is best. I'll start try11 from alignments with an unconstrained costfcn. try12 with the same costfcn, but without SCWRLing the alignments and starting from existing models. Since there are more fragments in the fragment library now, I am writing out the Template.atoms file from try11 (but not try12, since it will be the same). Mon Aug 23 12:20:16 PDT 2004 Kevin Karplus try11-opt2 never got the hairpin try12-opt2 polished up try2 try10-opt2 has some holes, looks a bit like try3. I like robetta model2, maybe we should steal some ideas from it---perhaps try optimizing with just it as a starting point, maybe with sheet constraints to bring the 3rd strand into place. robetta-model4 and robetta-model6 are not terrible either. Mon Aug 23 9:30pm Jenny Draper I'm starting try13 from the three robetta models we like (see above; models 2,4,6). I'm using the same constraints as for try3 & try10, but trying to set the weights more like in try12 (Kevin's settings). Hmmm. Robetta model 2 costs 300 points for constriant violations than try7, using the try13 costfunction. Obviously, I'll need to clean these constraints up to fit with the robetta models... Mon Aug 23 11:30pm Jenny Draper OK, I've replaced the helix-defining constraints with those from robetta-model2.helices. The sheet in robetta models 4&6 is the third-strand match that we never got! This brought the constraint cost way down. Try13 is now running on croak... Tue Aug 24 4:00pm Jenny Draper Try13 looks pretty good -- definately better than anything else we have so far. It clearly started from robetta model2, and has almost formed the 3-strand sheet, with strand 3 (~195-200) just starting to form. If we break the domains and strands like in the new script "strands": domain1 1-100 domain2 101-200 domain3 201-end s1 175-182 s2 187-192 s3 195-200 Then s3, which links domains 2 & 3, is pulling apart domain 3 a tiny bit to form the sheet. Domain2 has also blown up a little, with it's last helix (~155-170) pulling out of the bundle to form the sheet. Continuing this theme, the last helix of domain1, which is the linker between domains 1 & 2, has pulled out of domain 1 a bit to pack domain2 differently. The overall packing of the domains has shifted a bit, becoming more compact. It looks like once we have a formed sheet, we'll need to work on the packing again -- maybe by superpositioning & cut-and-paste? The current sheet positionings look good to me (all phobics on the inside in stripes). Tue Aug 24 4:30pm Jenny Draper Try14 is running on baa, attempting to bring the sheet together and improve the packing. Tue Aug 24 16:50:28 PDT 2004 Kevin Karplus I'm not that fond of the two N-terminal helices in try13-opt2. Do we have another model that we can cut and paste from up to about E70---or maybe only up to V51? The top-scorer (try12-opt2) is not a good choice, unless we stop around Y32, since it doesn't have the turn after the first helix. I could try just hacking off the first helix and superimposing that. The first attempt (in superimpose-first-helix.under) did not work---the helix was superimposed over its entire length, despite the "atom" hints given in the command. Maybe I need to remove the first 20 residues from the broken helix before superposing (or I should increase the effect of the atom hints in undertaker). Tue Aug 24 17:27:05 PDT 2004 Kevin Karplus I chose to fix undertaker to give more weight to the hints, and got the superposition I was looking for (in first-helix+try13.pdb). I'll make a chimeric model in decoys/T0248.try13-patched.pdb This model has really bad clashes, but may be easily fixed by an optimization run. I'll try that in try15. Wed Aug 25 12:10:09 PDT 2004 Kevin Karplus The best scorers with the unconstrained costfcn are try12-opt2, try15-opt2, and try14-opt2. I created a new cost function (light-constraints.costfcn) that has helix and strand constraints (plus some RR constraints), but gives them only a weight of 5. With this costfcn the order is try15-opt2, try14-opt1 (beating -opt2), try13-opt2, try12-opt2. Rosetta still likes try1-opt2.repack-nonPC best, followed by try12, try15, try13. Maybe I should do a polishing run with the light-constraints costfcn, and see if we can get the better parts of try14 and try15 combined. What I'll submit: best of the try13+ runs (borrowed heavily from robetta) try12-opt2 (polished-up version of try2) try10-opt2 try1-opt2 T0248-1oo0A-t2k-local-str2+CB_burial_14_7-0.4+0.4-adpstyle5 ?? I'm trying to get a reasonable diversity of different predictions. I could drop the template if there is another class of predictions that I've missed. Wed Aug 25 12:40:47 PDT 2004 Kevin Karplus I see that I set the weight for ReduceClash rather too high in try16---it's only been running 8 minutes, so I'll kill it and restart it with better weights (more crossover and tweaking). Wed Aug 25 1:15pm Jenny Draper I've made a hand-edited version of try15, T0248/T0248.try15-opt2.repack-nonPC__handEdit.pdb which rotates the N-terminal helix around to follow it's neighbor. Kevin, what do you think about it? Your ordering of models to submit looks good to me; I'll defer to your judgement on which of the try13 derivatives is best. Wed Aug 25 13:58:38 PDT 2004 Kevin Karplus I see that try16-opt1 has given up the carefully fixed initial helix of try15. I can't find the file hand edited file that Jenny refers to---it isn't in the main directory or the decoys. directory. From learithe@soe.ucsc.edu Wed Aug 25 14:13:25 2004 MIME-Version: 1.0 Date: Wed, 25 Aug 2004 14:13:24 -0700 (PDT) From: Jenny Draper To: Kevin Karplus cc: learithe@soe.ucsc.edu Subject: Re: T0248 In-Reply-To: <200408252103.i7PL3YVU011100@cheep.cse.ucsc.edu> Doh. I thought I'd written it the T0248 dir, but instead I wrote it to my home dir. Way to many terminals open... It should be there now. Let me know what you think. I'm afraid of doing "polishing runs" on pre-try12 models, since all I ever managed on thelm was to blow them up and mangle them... -Jenny On Wed, 25 Aug 2004, Kevin Karplus wrote: > I can't find the hand-edited file you refer to. > > I think I may want to create a chimera between try16 and try15, since > try16 threw away the work with the initial helix. > > You could try doing polishing runs one starting from just try12 models > one starting from just try10 models, using copies of try16.costfcn. > ============================================================ Wed Aug 25 14:29:42 PDT 2004 Kevin Karplus The hand-edited helix has really bad clashes, so scores very poorly with the try16 costfcn. They may be fixable with optimization, since it IS close to knob-and-hole fitting. try16 is really just try15 with the first helix messed up, so it may be worth doing a polishing run on the hand-edited model. I'll also try polishing runs on try10 and try12. Wed Aug 25 15:48:22 PDT 2004 Kevin Karplus try17 (from hand-edited try15) try18 (from try10) try19 (from try12) It looks like try17-opt1 just threw the helix back out away from the model, avoiding the clashes and closing the gap, but basically moving back to try15. Our models for this whole target are so open that it probably doesn't matter much whether we pack the first helix better or not. Wed Aug 25 16:17:14 PDT 2004 Kevin Karplus try17-opt2 scores a little better than try15-opt2, but is just as ugly. Still, it will probably be our first choice. Meanwhile, try18 and try19 have not even gotten to the point of finishing opt1 (cluck and squeal are about half the speed of cheep). Wed Aug 25 17:18:06 PDT 2004 Kevin Karplus Top scorers are now try16-opt2 \ try17-opt2 \ try15-opt2 > basically same (from robetta) except first helix try14-opt1 / try13-opt2 / try19-opt1 .... many others try18-opt1 I'll submit try17-opt2 from robetta try19-opt1 from try12 try18-opt1 from try10 try1-opt2 full auto try1-opt2.repack-nonPC best Rosetta When try18 and try19 finish, I'll probably replace their models with the opt2 versions. Thu Nov 18 23:43:11 PST 2004 Martina Koeva Based on the smooth gdt scores: best sam-t04 17.6354 (try13-opt1-scrwl/try13-opt1) best submit 17.5369 (also model1) model1 17.5369 auto 10.4405 align 7.0422 robetta best 19.0205 (robetta model2) robetta1 8.2283 Fri Nov 26 06:24:43 PST 2004 Kevin Karplus Domain : T0248_1 : FR/A : NT=79 : 21-99 Domain : T0248_2 : NF : NT=87 : 107-193 Domain : T0248_3 : FR/A : NT=87 : 199-285 T0248 models borrowed heavily from robetta-model2, which does better than our models, except on T0248_3, where try6-opt2 is our best model. This is probably not a good model to use in the CASP6 talk, since all the good stuff came from robetta.