Sat Aug 2 13:35:12 PDT 2008 TR464 Sat Aug 2 13:38:13 PDT 2008 Kevin Karplus CASP organizers say REFINEMENT TARGET TR464 (the best submitted server model according to GDT_TS). GDT_TS=77; RMSD=1.56A. Two problematic regions: around res.41 and 59. NMR structure and starting model trimmed to res. 18-86. Make started Sat Aug 2 13:39:51 PDT 2008 Running on moai01.kilokluster.ucsc.edu Sat Aug 2 18:33:50 PDT 2008 Kevin Karplus The closest server model is BAKER-ROBETTA_TS2 with GDT=74.3%, but my MQAUC7-opt3.gromacs0 is closer at 75%---it is based on Pcons_multi_TS1=BAKER-ROBETTA_TS1. Also close among my models are MQAX1, MQAU1, SAM-T06-server_TS1 (not SAM-TO8_server!), MQAU7, MQAX7 (my model2, 64.9%). try8-opt3 (my model1 64.9%), ... So I need to look at some of my closest previous models and see where it differs from the provided model. Mon Aug 4 11:07:33 PDT 2008 Kevin Karplus try2-opt3, optimized from the provided model, scores well. Rosetta likes try2-opt3.repack-nonPC (actually better than try2-opt3.gromacs0.repack-nonPC, which is very unusual). Mon Aug 4 11:21:55 PDT 2008 Kevin Karplus I picked up a few models from my previous predictions that were close to the provided models: ReadConformPDB from-MQAU1-opt3.pdb.gz ReadConformPDB from-MQAUC7-opt3-gromacs.pdb.gz ReadConformPDB from-MQAX1-opt3.pdb.gz ReadConformPDB from-SAM-T06-server1.pdb.gz I also wrote a short rasmol script to highly the problematic regions identified by the CASP organizers. For the area around A41, I rather like the strand that try1-opt3 is making, though it may not be compatible with any of the rest of the model. The MQAU1 and SAM-T06-server1 models also have a strand near A41. None of these models seem to differ much in the area around R59, except in the placement of the sidechain. Sat Aug 16 13:33:19 PDT 2008 Kevin Karplus I'll make a chimera-try2-MQAU1 that takes G36-N47 from from-MQAU1-opt3 and the rest from try2-opt3. I'm not sure whether try3 will recover completely from this surgery, as try2-opt3 has only one clash and no breaks---it wil take a while for the modified model to get that polished. Sat Aug 16 14:12:49 PDT 2008 Kevin Karplus try3-opt3 doesn't even score as well as the initial model! Oops, for try3 I left in the constraint for the helix in the region I was trying to modify. For try4, I'll do the same optimization without that constraint. Sat Aug 16 14:51:18 PDT 2008 Kevin Karplus OK, once the undesired helix constraint is removed, both try4-opt3 and try3-opt3 score reasonably well (just a little worse than try2-opt3, and that mainly due to the remaining breaks). Both undertaker and rosetta like try2 better than try3 or try4, but I think that there is some hope for improving try4. I'll run try5 from the gromacs-optimized versions of try1, try3, and try4. Sat Aug 16 15:00:34 PDT 2008 Kevin Karplus I should probably run the provided model through VAST, to see if there is a template it is based on---and to see if that template is multimeric. Sat Aug 16 15:06:13 PDT 2008 Kevin Karplus Unfortunately, the VAST server is either very heavily loaded or broken at the moment---it still hasn't parsed the job request to let me start the search after 4 minutes. I may have to wait until Monday to do the VAST jobs. Sat Aug 16 16:15:22 PDT 2008 Kevin Karplus VAST still not accepting the job from before---I tried resubmitting without success. I'm pretty sure the server is down. try5-opt3 improves on try4-opt3.gromacs0 and ends up scoring better than try4-opt3, but rosetta doesn't like it. I don't really like the hairpin in the chimera and try3,4,5. I'd rather have Q22-A24 paired with Q29-S27. Perhaps I should make a costfcn that includes that, and try another run from alignments. Sat Aug 16 17:02:46 PDT 2008 Kevin Karplus try6 does not score very well (but then, neither did try1). No wonder---I named the residues all wrong. I must be more tired than I thought. I wanted Q39-A41 paired with Q46-S44 Trying again in try7, with the correct residue numbers this time! Sat Aug 16 21:24:42 PDT 2008 Kevin Karplus try7 gets the hairpin, but not really the sheets and helix. Still, I might be able to do a cut-and-paste from it. Sat Aug 16 21:49:10 PDT 2008 Kevin Karplus I made a chimera-try2-try7 model, though the hairpin I pasted in (T37-N47) doesn't point in the right direction. I'm hoping that undertaker an reorient it and make it fit somehow. Sat Aug 16 22:38:51 PDT 2008 Kevin Karplus try8 almost closes the gaps and presents a reasonable, if slightly different than I expected placement of the hairpin. Since I'm about to go to bed, I'll start a longer run as try9 to polish the try8 model. Sun Aug 17 05:17:40 PDT 2008 Kevin Karplus The try9 run improved on try8 (according to the try9, try5, and try8 costfcns), but rosetta thinks it is worse. I'd like to try to make move the hairpin so that it doesn't bend as sharply at N47. Perhaps if I make a chimera with try5, taking just V38-Q46 from try9?? Sun Aug 17 06:43:16 PDT 2008 Kevin Karplus try10 discarded the hairpin in trying to put it where I wanted---perhaps the "extend_hairpin" needed more weight and the "hold_hairpin" less? Of course, there is another idea on what should be fixed---perhaps I should add SheetConstraint V38 A41 T80 A77 Hbond A77 and not worry about the hairpin for now. Sun Aug 17 09:49:38 PDT 2008 Kevin Karplus try11-opt3 doesn't quite form the sheet I wanted, but is in roughly the right place, I think. It also changes the C-terminal bit of the model, flipping I83 to the inside,but leaving exposed a hydrophobic patch of I83,I55,L21, and L29. The sheet extension I requested and the hairpin are not compatible, as they both need Q39 hbonds, but to different partners. Perhaps if I change the sheet extension, I can make things more compatible: SheetConstraint Q39 D42 T80 A77 Hbond A77 Sun Aug 17 10:06:55 PDT 2008 Kevin Karplus For try12, I included two constraints, both heavily weighted: ConstraintSet extend_sheet SheetConstraint Q39 A41 T80 I78 Hbond A79 ConstraintSet hairpin SheetConstraint Q39 A41 Q46 S44 hbond A41 and am starting from just the three chimeras. Sun Aug 17 11:32:39 PDT 2008 Kevin Karplus try12 gets the hairpins, but does a fair amount of other damage. I think I want to try to copy V38-N47 from try12 to try5 and re-optimize, with a bit less weight on the new sheets. Sun Aug 17 11:41:59 PDT 2008 Kevin Karplus try13 started to optimize chimera-try5-try12. Sun Aug 17 12:25:49 PDT 2008 Kevin Karplus OK, try13 is the most convincing model I've seen so far. I'll do a polishing run for it, starting from just the gromacs optimized models (to try to reduce clashes). Sun Aug 17 15:55:55 PDT 2008 Kevin Karplus try14-opt3 scores well and rosetta likes it reasonably well (though not nearly as well as try2-opt3.gromacs0.repack-nonPC) Sun Aug 17 16:08:30 PDT 2008 Kevin Karplus For try15, I'll polish just try14-opt3.repack-nonPC (with the BackRub oprerators turned up higher). I also played with the weights of the costfcn, to put more emphasis on sidechain probability, burial, and hbonds. Sun Aug 17 16:54:30 PDT 2008 Kevin Karplus try15-opt3 scores better than try14-opt3 with the undertaker try15.costfcn, but rosetta likes try14-opt3.gromacs0.repack-nonPC a little better than try15-opt3.gromacs0.repack-nonPC I just took a look at 2hf1A (the main hit for this protein) and noticed that it is a strand-swapped dimer. I should dimerize my model(s) and reoptimize in a dimer context. Sun Aug 17 17:24:05 PDT 2008 Kevin Karplus I made dimers based on 2hf1A, but I don't think that they are going to be of any use, as the structures are not really similar enough to have any confidnece of them packing the same, and the dimer ends up with two monomers that barely touch, and then not on a hydrophobic patch. Mon Aug 18 10:42:35 PDT 2008 Kevin Karplus The VAST server still seems to be down, so I tried CE (broken, with error message indicating script problems) and DALI: http://ekhidna.biocenter.helsinki.fi/dali_server/results/20080818-031-30ae0814212d2f7910e5475ddc71464c/index.html Mon Aug 18 12:35:15 PDT 2008 Kevin Karplus For try16, I'll try adding some contraints involving residues around R59, to get the tip of the hairpin to flex the other way. This may be difficult, as all the models so far have the same structure there. Mon Aug 18 13:35:25 PDT 2008 Kevin Karplus The VAST server is still broken, but DALI reports No: Chain Z rmsd lali nres %id Description 1: 2hf1-B 3.7 1.4 39 59 28 MOLECULE: TETRAACYLDISACCHARIDE-1-P 4-KINASE; 2: 2hf1-A 3.7 1.5 39 61 28 MOLECULE: TETRAACYLDISACCHARIDE-1-P 4-KINASE; 3: 2pk7-B 3.6 1.2 38 60 29 MOLECULE: UNCHARACTERIZED PROTEIN; 4: 2pk7-A 3.5 1.1 38 64 29 MOLECULE: UNCHARACTERIZED PROTEIN; 5: 2jr6-A 2.6 1.6 39 68 26 MOLECULE: UPF0434 PROTEIN NMA0874; 6: 2jny-A 2.5 2.1 38 61 37 MOLECULE: UNCHARACTERIZED BCR; I saved the full results in dali_results_for_T464.html Mon Aug 18 17:18:18 PDT 2008 Kevin Karplus try16 did not really accomplish anything in terms of moving the residues near R59. Although it might be possible to back up to a much lower-quality model that would be easier to modify, I think I'll give up on fixing that region instead. Mon Aug 18 17:22:27 PDT 2008 Kevin Karplus Strange: 2pk7A was the second-highest hit in TR464.t06-w0.5-scores.rdb, but was not included in the best-scores results---probably because it isn't in our template library. 2jnyA was the top hit, but also wasn't included. 2jr6A was 11th. I'll put these in MANUAL_TOP_HITS, and do an optimization from alignments, to see if anything comes of it (like a different R59 loop). I'll also try dimerizing with them (only 2hf1A and 2pk7A are dimers). Mon Aug 18 17:38:30 PDT 2008 Kevin Karplus try17 started from alignments with the try16 costfcn (which may not be balanced well for optimization from alignments). Mon Aug 18 18:00:46 PDT 2008 Kevin Karplus I tried making dimers from TR464, try2, and try16, but they came out quite different. I suspect that there is something not quite right with the PrintMultimerPDB command Mon Aug 18 20:51:22 PDT 2008 Kevin Karplus try17 is not a great model, but it *does* get a different loop around R59. I'll try making a chimera of try16 and try17, then polishing the result. Mon Aug 18 21:55:13 PDT 2008 Kevin Karplus try18-opt3 outscores try16-opt3! rosetta does not like it that much, so maybe I should try polishing try18-opt3.gromacs0.repack-nonPC. Mon Aug 18 22:02:48 PDT 2008 Kevin Karplus try19 started with sidechain costs turned up (from try14.costfcn as a base) and with only try18-opt3.gromacs0.repack-nonPC as an initial model. Mon Aug 18 22:46:48 PDT 2008 Kevin Karplus try19 slightly edges out try18 with the try19.costfcn, but with other costfcns, try18 consistently beats try19. Ever rosetta prefers try18, so I think that this may be as far as I can go with using undertaker to optimize a gromacs+rosetta optimized model. Tue Aug 19 10:20:40 PDT 2008 Kevin Karplus try20 improves on try18 according to undertaker, though rosetta doesn't think so. There are no significant clashes, but things still look a little bit foamy to me (less foamy than try2-opt3.gromacs0.repack-nonPC or the original model, though). Tue Aug 19 11:26:31 PDT 2008 Kevin Karplus I've been trying to figure out whe the dimers for try2 and try20 come out so different when based on 2pk7. I made a change to undertaker to allow printing out the alignment model that the multimer is based on, and it is the same for try2 and try20, so the difference is in the superposition. I think that the problem is that the part of the alignment from A34 to S44 has very different structure and placement in the two models. I'll try unaligning that in TR464-2pk7A.dimer-shrunk-a2m and see if that makes the multimers consistent. Tue Aug 19 11:36:06 PDT 2008 Kevin Karplus That seems to have fixed the problem, so that dimers are now consistent, for both 2hf1 and 2pk7 superpositions. I'll make dimers for all my current favorite models (try2, try5, try11, try16, and try20) and see if I can make a good dimer out of any of them. Tue Aug 19 11:45:16 PDT 2008 Kevin Karplus dimer/try21 started to try to pack dimer tighter. Tue Aug 19 13:04:48 PDT 2008 Kevin Karplus dimer/try21 does pack the dimer a bit better, but I wonder if I should try to make the sheets line up between the monomers. I'll try that for dimer/try22 I noticed that the "unpack" operation for dimers that did not start at 1 was numbering residues wrong, so I fixed the unpack-multimer script and Make.main to use START_COL. Tue Aug 19 14:45:19 PDT 2008 Kevin Karplus dimer/try22 made only slight tweaks to dimer/try21 (the biggest probably being to change the rotamer for P70). The monomers (from-dimer-try22-opt3 and from-dimer-try21-opt3) do not score as well as try20, try18, try16, and try15 with the try20 costfcn). Perhaps I should try making a dimer of try20, but based on the placement in try22. Tue Aug 19 16:28:23 PDT 2008 Kevin Karplus I optimized dimer-try20-try22, and the try23 model scores much better than try21 or try22, with both undertaker and rosetta. Tue Aug 19 17:01:43 PDT 2008 Kevin Karplus Submitting with comment: Although the provided base model is not one of the server models, it appears to be based on the BAKER-ROBETTA models (TS1 and TS2) and has the usual problem of being a bit foamy. I used some of my previously generated models that were close to the provided model to produce other plausible models for the suspect regions around residues A41 and R59. None of my close models had a different loop for R59, so I concentrated mainly on the loop around A41. I pasted in a hairpin from one of my models aroudn A41, but I didn't like it, because it had a rare separation=4 Hbond, yet had the canonical DG residues for a separation=3 Hbond. I added constraints to force the more conventional hairpin and generated a new model (try7) from alignments, pasting the loop from this model into the model optimized from the provided model (try2), optimizing to close breaks. This ended up moving the loop (try9), so I added still more constraints to hold the hairpin close to the sheet---indeed extending the sheet. When I was satisfied with that loop, I tried tweaking the R59 loop without success (I was too deeply into the local minimum to move around). So I tried generating a new model from alignments with the same constraints that I'd been using to try to tweak the R59 loop, and generated a slightly different loop. I pasted this in an reoptimized the model. Finally, I tried supreposing my monomer on the dimers in 2hf1 and 2pk7. This probably doesn't help, since the experimental model is an NMR model, and the thus more likely to be monomeric. The resulting changes for optimizing in a dimeric context were fairly small, though, so I decided to keep them. Model 1 from-dimer-try23-repack.pdb.gz # < dimer-try20-try22 dimer-try20-try22 has the try20-opt3 monomers arranged as in the try22-opt3 dimer dimer/try22-opt3 < dimer/try21-opt3 < dimer-try20-2pk7 try20-opt3.pdb < try18-opt3 < chimera-try16-try17 chimera-try16-try17 is mostly try16, but L54-Q61 from try17-opt3 try17-opt3 < align(2pk7A) 2 TR464.try16-opt3.pdb # < try15-opt3 < try14-opt3.gromacs0.repack-nonPC < try13-opt3.gromacs0 < chimera-try5-try12 chimera-try5-try12 is mostly try5-opt3, but V38-N47 from try12-opt3 try12-opt3.pdb < chimera-try5-try9 chimera-try5-try9 is mostly try5-opt3, but v38-Q46 from try9-opt3 3 TR464.try11-opt3.pdb # < try10-opt3 < chimera-try5-try9 chimera-try5-try9 is mostly try5-opt3, but V38-Q46 are copied from try9-opt3 try9-opt3 < try8-opt3 < chimera-try2-try7 chimera-try2-try7 is mostly try2-opt3, but T37-N47 from try7-opt3 try7-opt3 < align(2hf1A) 4 TR464.try5-opt3.pdb # < try4-opt3.gromacs0 < chimera-try2-MQAU1 G36-N47 copied from MQAU1-opt3, which was optimized from the SAM-T06-server_TS1 model 5 TR464.try2-opt3.gromacs0.repack-nonPC.pdb # < TR464 # best rosetta score optimized from the initial model with undertaker, followed by gromacs, followed by sidechain repacking with Rosetta. Still quite foamy.