Thu Jun 8 08:36:25 PDT 2006 T0325 Make started Thu Jun 8 08:38:31 PDT 2006 Running on shaw.cse.ucsc.edu Tue Jun 27 16:42:26 PDT 2006 Firas Khatib so looking at score-all+servers.unconstrained.pretty, Pmodeller6_TS3 and ROBETTA_TS3 score best (equally), but when we put them in superimpose-best.under they looked identical (which would explain why they both score exactly the same! The try1-opt2 run disagrees with the alignments and Robetta_3 on the region 100-120. There is a break in the try1's helix at E108. One of the reasons for this is because of the region 70-100 that has very little agreement. try1 thinks it's 2 helices, Robetta_3 has it somewhat disordered and the top 3 undertaker-align files are all disordered as well (although they seem to all agree on where it is) There is also disagreement on where 150-190 should go. It also has low secondary structure prediction and Robetta, try1, and the alignments all disagree on where to place it and what is helix and what is not. The turn at 215-230 needs agreement as well. We might want to do a subdomain for 70-190, before that we will blastp that two different sequences (70-100 & 150-190) and look for similar structures in the PDB. Well... that found nothing... "no significant hits found" Looking at all the Robetta models, they seem to all disagree as well for regions: 70-100, 151-190, and 215-230 but they all have agreement with the secondary structure at: 91-97 helix 163-171 helix so we will try to implement that into try2 91-97 is already a helix constraint, but only 0.6 so we will increase this. these are the constraints we changed in try2: HelixConstraint W91 E97 0.9 HelixConstraint T163 Q171 0.7 added this one try2 is running on lopez Tue Jun 27 23:05:36 PDT 2006 Kevin Karplus I'd like to have some suggestions by 8pm on Wednesday (28 Jun) for the preliminary submission. We can replace them later. Wed Jun 28 14:53:32 PDT 2006 Firas Khatib as Kevin noticed during our lab meeting, there are many conserved HIS, looking at try1, they are very close to each other. There are only 3 CYS. Wed Jun 28 17:09:42 PDT 2006 Sylvia Do Starting from try1, we've done this: We're increasing the following constraints to bring the disordered region closer to the core Constraint L64.CB L83.CB -10. 7.0 14.0 0.9 Constraint L64.CB W109.CB -10. 7.0 14.0 0.8 We have this for both try3 and try4, where try4 is based off of try2. For both these runs, we turned off all other rr.constraints. Both try3 and try4 are running on Lopez Wed Jun 28 17:24:28 PDT 2006 Firas Khatib http://predictioncenter.org/casp6/meeting/presentations/DR_assessment_RD.pdf according to the CASP6 disordered prediction assesment, the best predictor was ISTZORAN, so I will try using this on T325 to see if any of these regions: 70-100, 151-190, and 215-230 are predicted to be disordered. (70-100 is not likely to be due to conservation and rr prediction) well, ISTZORAN doesn't seem to be available on the web. I had to go with the 3rd best predictor (according to CASP6) from JonesUCL The results are as expected: http://bioinf.cs.ucl.ac.uk/domout/disopred/1_30_36_179_29-6-2006.jpg other than the 1st MET nothing is highly predicted to be disordered with the regions we were discussing (70-100, 151-190, and 215-230) having a 5-18% chance Wed Jun 28 17:43:42 PDT 2006 Firas Khatib I was able to find the 2nd best predictor online, DISSVM, and am waiting for the results. Wed Jun 28 20:16:52 PDT 2006 Firas Khatib weird... I still haven't gotten an email from DISSVM! well, it looks like try1-opt2 still scores best with the unconstrained cost function! The constraints that we put in try3 and try4 did not seem to work at all! Maybe we need to increase constraints from 10 to 50? try5 will be like try4 but with the CONSTRAINTS weight set to 50 (on shaw) try6 will be like try3 but with the CONSTRAINTS weight set to 50 (on shaw) hopefully when those two are done, we can submit: try5, try6, try1, try2, and try4 Wed Jun 28 21:25:01 PDT 2006 Kevin Karplus The sheets have so much weight in try1-6, that the other constraints are not too important. I'll construct a try7 that is a bit more balanced. Wed Jun 28 21:29:39 PDT 2006 Kevin Karplus try7 started on cheep Incidentally, the PrintTemplateAtoms line should be commented out after the first run, unless new PDB chains are aligned to. Otherwise you keep getting new identical copies of Template.atoms.gz, which slows down rsync commands. Wed Jun 28 21:36:34 PDT 2006 Kevin Karplus Looking at the RR constraints, I think that the try1-opt2 sheets may be off by 1. In particular, I think that the strands V32-A37 and V207-C212 may be flipped over. The rr constraints would cluster much better if the two strands were aligned one off from where they are. I'm not sure which way to push them, though. We may have to hand-edit alignments to effect this change. Wed Jun 28 21:47:13 PDT 2006 Kevin Karplus It looks like one of the simplest ways would be to use just the constraints from the alignment to 1v6tA. I'll try that for try8. Wed Jun 28 21:53:38 PDT 2006 Kevin Karplus try8 started on orcas. Wed Jun 28 21:56:35 PDT 2006 Kevin Karplus I killed try7 and try8, because I just realize that the all-align.a2m.gz file was missing, and so only a tiny subset of alignments was being tried. No wonder things weren't looking too promising! The runs will be slower now that all-align.a2m.gz is there. try7 restarted on orcas. try8 restarted on cheep. Wed Jun 28 22:35:38 PDT 2006 Firas Khatib that is very odd about the all-align.a2m.gz file! Also, I commented out PrintTemplateAtoms in try2.under (since we did have all those conversations about it earlier this week). try4 and try5.under have it commented out as well, but I think try3 and try6 copied try1 and therefore were not commented. This needs to be added to the checklist for sure! Wed Jun 28 22:59:42 PDT 2006 Kevin Karplus I looked at the make.log file, and the all-align.a2m.gz bug was due to an error in Make.main that I was debugging at the time. It was fixed that day, but I did not notice that it had affected T0325. Wed Jun 28 23:37:37 PDT 2006 Kevin Karplus The best scores with try8.costfcn are currently try1 try2 try6 try5 try8-opt1 The best scores with try7.costfcn are currently try1 try2 try5 try6 try8-opt1 though try7-opt1 is likely to beat try8-opt1 when it is done. I'll do a preliminary submission of try1 try2 try6 try5 align1 (from 1v6tA) Thu Jun 29 08:10:33 PDT 2006 Kevin Karplus The best scoring models with different costfcns: try1: try2, try1, try4, try6, try3, try5, try7 try2: try2, try1, try4, try6, try3, try5, try7 try3: try2, try1, try4, try6, try3, try5, try7 try4: try2, try1, try4, try6, try3, try5, try7 try5: try6, try5, try2, try4, try3, try1, try8 try6: try6, try5, try2, try4, try3, try1, try8 try7: try1, try7, try2, try8, try6, try5, try3 try8: try8, try1, try2, try7, try6, try5, try4 unconstrained: try1, try7, try2, try4, try3, try8, try6 try7-opt2 has some bad breaks, but is otherwise a good TIM-barrel wannabe. try8-opt2 has a poor sheet structure---it looks like it is trying to form a barrel with only 6 parallel strands. I don't like it. Neither try7-opt2 nor try8-opt2 are good enough to displace the models from the preliminary submission, but try7 might bear further work (perhaps crossing it with try1). Thu Jun 29 09:33:07 PDT 2006 Firas Khatib so is the submission still going to be: try1 try2 try6 try5 align1 (from 1v6tA) or should we replace align1 maybe with try4 which scores in the top 4 for every costfcn? Thu Jun 29 14:02:57 PDT 2006 Kevin Karplus try4-opt2 is pretty non-compact, so I think I'd rather submit the alignment. If there are any *new* models to look at, I'll consider them. Thu Jun 29 15:55:15 PDT 2006 Firas Khatib we are going to use try7's constfcn (which seemed to work well) but using try1-opt2 as input (with ReadConfromPDB) this is try9 running on camano Thu Jun 29 16:10:45 PDT 2006 Sylvia Do We are going to increase the break costs of try7 from 50 to 75. try10 is running on Orcas. Thu Jun 29 16:52:53 PDT 2006 Firas Khatib I took try8 and added these constraints to it: Constraint M179.CB N211.CB -10. 7.0 14.0 0.9 Constraint L64.CB I74.CB -10. 7.0 14.0 0.8 Constraint V80.CB A143.CB -10. 7.0 14.0 0.7 try11 is now running on shaw. Thu Jun 29 17:27:45 PDT 2006 Firas Khatib try12 will be very similar to try3, but with maybe_ssbond off and the rr constraints given equal weight. try12 is running on lopez. Thu Jul 13 14:50:19 PDT 2006 Sylvia Do I looked at try10, try11, and try12. None of them did very well on the unconstrained costfcn or on their own costfcn, for that matter. They all have some bad breaks, but that seems to be the trouble with all the trys. try9-opt2, try9-opt2.repack, and try9.opt1 did well on the unconstrained costfcn, even better than try1-opt2. While try9 did better on the scores, I think try10 makes a better looking TIM barrel. I'll see if I can get try9 to look more like try10 by increasing the break costs from try9 from 50 to 75. This will be try13. Just as a note: try13 <- try9 <- try1-opt2 with try7's costfcn try13 is now running on bark. Thu Jul 13 16:24:37 PDT 2006 Sylvia Do I don't see the scores for try10 on grep-best Rosestta or the unconstrained costfcn. I think this is because when I made try10 I did not have my umask set correctly. I did fixmode before (and again), but it do not help. I'll just redo it, as try14. try14 is running on whidbey. I'm just playing in ProteinShop right now, trying to see if I can coax try9-opt2 into a more TIM-barral-y shape. The file I'm working on is called T0325.try9-opt2-PShop.pdb under the ProteinShop subfolder. If I can get it to look good, I'll change the file name. Thu Jul 13 23:42:30 PDT 2006 Sylvia Do I'm looking at try13 and try14 right now, and it strange, but try14, which was supposed to be a copy of try10, does not look like it. Gah, did I do something wrong? I copied try10's .under and costfcn to try14, then used make. try13-opt2 looks more promising. Fri Jul 14 15:21:03 PDT 2006 Sylvia Do ProteinShop is a bit wonky. The model that opens in ProteinShop does not look like the actual model. So far, I like how try9 and try13 look. I think there is a a hinge created around K72-V100. I think Grant suggested this a while ago. In any case, I'll see if I can bring it a little closer to the core from try13. I added this constraint to try13's costfcn to bring the "arm" closer to the core in try15: Constraint K72.CB V100.CN -10 7.0 10.0 1 try15 is running on orcas. Mon Jul 17 12:49:42 PDT 2006 Sylvia Do Try15-opt2 is doing well in the score-all.unconstrained.pretty. It is losing to try9 mostly due to the soft clashes. I will try to fix this by increasing the penalties for soft clashes. I've changed the penalties for the soft_clashes from 20 to 30, and backbone_clashes from 2 to 5 from try15. try16 is running on orcas. Mon Jul 17 17:24:31 PDT 2006 Sylvia Do Try16-opt2 is currently the best scoring on the score-all.unconstrained.pretty. Burial for all of them are pretty bad though. Also, in reducing the soft clashes, try16 now isn't so well packed and has a few holes that I don't like.I don't know about the breaks though -they're all pretty bad as well. I'll try increasing the constraints for breaks and dry 6.5 and 5 to see if it helps. try17 will be try16 with break increased from 75 to 125. In addition, try16's dry6.5 has been increased from 20 to 25, and dry 5 from 15 to 20 for try17. try17 is running on bacchus. Tue Jul 18 14:39:02 PDT 2006 Firas Khatib all our tries, including try1 have the helix from 102-117 straying away from where the alignments have it. I put in some constraints to keep this one in place and am running this as try18 on squawk Tue Jul 18 14:43:08 PDT 2006 Sylvia Do try16 may be our best scoring one, but it broke a helix at D110. N101-K118 is strongly predicted to be a single helix, so I will add this constraint to the costfcn of try16. The Helix constraint will follow try15 - helix from N101 to K118. The helix constraints are added from try15-opt2: HelixConstraint N101 S115 HelixConstraint I113 K118 try19 is running on orcas. Wed Jul 19 00:09:51 PDT 2006 Firas Khatib try18 is garbage, I increased the breaks constraint and am running this as try20 on shaw. Wed Jul 19 00:33:59 PDT 2006 Sylvia Do try19 did not work; there are still two helixes from 101-118, divided at 110. Well, it seems like 19 is losing points on soft clashes, phobic fit, and breaks. I'll try what I did for 19, but this time with the constraint: HelixConstraint N101 K118 instead on what I had for try19. In addition, I'm going to increase the penalties for soft clashes from 30 to 35, and backbone_clashes from 5 to 8. I will also increase break penalties from 75 to 100. I also noticed a typo that I had made in the earlier files. I had meant to type Constraint K72.CB V100.CB -10 7.0 10.0 1 but instead, I had Constraint K72.CB V100.CN -10 7.0 10.0 1 So trys15? I believe it is and its successors have that. I've changed it to .CB for this version. I don't know if that made any difference though. try21 is running on Orcas. Wed Jul 19 12:55:43 PDT 2006 Sylvia Do Increased the weight for the helixconstraint to 10 and also the penalty for breaks to 200. Hopefully this should work. Modified: HelixConstraint N101 S118 10 try22 is running on lopez Wed Jul 19 16:46:57 PDT 2006 Firas Khatib try23 uses try16 as input but with the constraints I had for try20 try23 is running on shaw. I think one problem with most of these models is that the strand from 60-70 is in the wrong place. I will try to fix this. Wed Jul 19 17:47:38 PDT 2006 Firas Khatib I am running try24 with try9 as input, but with constraints to try to slide the sheet I60-T65 down. try24 is running on shaw. Wed Jul 19 17:54:47 PDT 2006 Firas Khatib try23 is done and scores well. it almost copied over try16 because I messed up try23.under, but I noticed this as try23-opt1 had finished (and was called try16-opt1.pdb) and scored better than try16-opt2.pdb.gz and try16-opt1.pdb.gz so I quickly copied all the try16 files and renamed them .original in case try23 stepped over them when it gzipped stuff at the end. Anyway, try23-opt2 scores best with the unconstrained costfcn as well as with the try24 costfcn. I am going to do one more run: try25 will input try23 and use try24's costfcn. Wed Jul 19 18:10:28 PDT 2006 Firas Khatib try23.under has been renamed to try23-fixed.under so that I can run the gromacs and rosetta-repack runs. Wed Jul 19 18:27:53 PDT 2006 Firas Khatib try25 is running on camano. it uses try24's costfcn (to slide the sheet I60-T65 down) with the added constraint of keeping the helix from 102-117 closer to the core. Wed Jul 19 18:57:43 PDT 2006 Firas Khatib as soon as try24 and try25 finish I will post my ranking recapping where tries came from: try25 -> try23 -> try16 -> try1 try24 -> try9 -> try1 try23 -> try16 -> try1 try22 -> try1 try21 -> try1 try19 -> try1 try17 -> try1 try16 -> try1 try15 -> try1 try13 -> try1 try9 -> try1 everything else did not use ReadConformPDB Wed Jul 19 19:17:29 PDT 2006 Sylvia Do try23-opt2 < try16-opt2 < try1-opt2 < alignments (2c1gA too, but uses 2c1iA last) try22-opt2 < try1-opt2 < alignments (2c1iA) try9-opt2 < try1-opt2 < alignments (2c1iA) try17-opt2 < try1-opt2 < alignments (2c1iA) try1-opt2 automatic model < alignments (2c1iA) try23 is scoring really well, but I don't like how there is a kink or something at ASP110, a leftover from try16. I like try22 better because it fixes this, even if it didn't beat 23 or 16 in the score-all.unconstrained. Wed Jul 19 21:41:55 PDT 2006 Firas Khatib try24 and try25 are done... I am going to check them out and suggest a ranking Wed Jul 19 22:07:17 PDT 2006 Firas Khatib try25 has too many breaks, I like how try24 was able to shift the I60-T65 sheet down, though... but there are also quite a few breaks. Wed Jul 19 22:28:40 PDT 2006 Firas Khatib I think at this point, trying to get a diverse 5 models as possible is best. The only region where the high scoring models seem to disagree, however, is around I74-G120. All the other models that have different conformations: try14,try18,try20 don't score very well at all, mostly because they are very ugly! I think in terms of ranking, I'm leaning towards: try24 -> try9 -> try1 based off 2clg, 2cli, and 1yax i wish it had less breaks though try22 -> try1 based off 2clg, 2cli, and 1yax because Sylvia prefers this one try23 -> try16 -> try1 based off 2clg, 2cli, and 1yax because it scores best with unconstrained try15 try12 because they are different time to close this file so Kevin can read it (since I am 3 hours late already) Wed Jul 19 23:14:11 PDT 2006 Kevin Karplus I tried looking at grep-best-rosetta, but it was corrupted. It turns out that someone had added ".original" *AFTER* the .gz on several of the files. I renamed *.gz.original to *.original.gz so that the files could be looked at and evaluated. Now I can see that the Rosetta hates all our models, but the least bad (from its view) is try24-opt2.gromacs0.repack-nonPC. Since try24-opt2 is Firas's favorite, I'll substitute rosetta's favorite---it *does* have fewer breaks, as Firas would want. Wed Jul 19 23:24:43 PDT 2006 Kevin Karplus Submission done with comment Model 1 is try24-opt2.gromacs0.repack-nonPC, which was optimized from try9, in turn from try1. This is the model that rosetta hates least of the models that it repacked sidechains for. Model 2 is try22-opt2, the model the Sylvia likes best of the models created. It was reoptimized from try1-opt2, the automatic model. Model 3 is try23-opt2, the model that scores best with the unconstrained cost function. Model 4 is try15-opt2, included to increase the diversity of our submitted models. Model 5 is try12-opt2, included to increase the diversity of our submitted models. Wed Jul 19 23:48:58 PDT 2006 Firas Khatib just as a last chance, I'm running try26, which is try24 with a high break penalty on shaw and try27, which is try25 with a high break penalty on camano we might get lucky and fix those breaks! Thu Jul 20 08:08:00 PDT 2006 Firas Khatib sadly, I do not think that try26 or try27 look any better than their previous models. Fri Mar 23 17:46:22 PDT 2007 Kevin Karplus Our best model was try6-opt1-scwrl, but it wasn't nearly as good as Robetta_TS3, which we had pulled out as the best server model (it wasn't the best, but it was better than anything *we* did, so we still would have been better off being a meta-server).