Wed Jun 21 10:32:25 PDT 2006 T0340 Make started Wed Jun 21 10:34:13 PDT 2006 Running on camano.cse.ucsc.edu Wed Jun 21 10:39:36 PDT 2006 Kevin Karplus BLAST gets excellent hits in PDB: 1g9oA 61% id over 84 residues E-value 9.5e-25 1i92A 60% id over 88 residues E-value 2.8e-24 1gq5A 61% id over 82 residues E-value 3.6e-24 1gq4A 61% id over 82 residues E-value 3.6e-24 Wed Jun 21 19:22:31 PDT 2006 Kevin Karplus The HMMs also get lots of hits (at least 30) to the b.36.1.1 domains (which include 1g9oA) and at least 67 to b.36.1.* . Fri Jun 30 16:05:48 PDT 2006 Cynthia Hsu, Crissan Harris The best-models.pdb.gz file seems to line up perfectly between most of the models. The main object of concern is the loop from 24-29; it follows the closest model (in terms of sequence identity) but we're not sure if this is accurate. Running the burial script also showed two regions (approx. residues 50 - 54) that were predicted to be buried but still appeared exposed in the model For the future polishing runs, we intend to adjust the constraints follows: The only His in close proximity are His65 and His22, which the rr script does not indicate to be in contact. Our assumption is that they do not participate in metal ion bonds. The only Cysteine, Cys8, is also not in close proximity to any His and Cys and probably not contributing to metal bonds or disulfide bonds. Therefore the constraints on these should be removed. Following the suggestions that Prof. Karplus and George made, we've decided that for try2, we will raise the soft-clashes costs to 40 and the breaks to 150, and delete the constraints. He also suggested to increase the sidechain weight slightly; our guess is to try to set it at around 8. Hopefully this will fix the exposed regions of the target. Under George's supervision, we made the following changes to the try2.under file: All TryAllAlign lines were commented as well as the ReadFragmentAlignment. Because this was a polishing run, lines 109 and 110 ("InfilePrefix decoys/" and "include read-pdb.under") were uncommented. try2 is currently running on lopez. Sat Jul 1 12:47:45 PDT 2006 Cynthia Hsu I can't seem to get the makefile to work properly, so I've tried redoing it. try2 is currently running on orcas. Mon Jul 3 12:14:24 PDT 2006 Crissan Harris, Cynthia Hsu Grant helped us fix our polishing run so try2 is currently running on lopez. Mon Jul 3 16:42:09 PDT 2006 Crissan Harris, Cynthia Hsu All the models are aligning up fairly well; the only concern is the coil from residues 24-28, but since these are, according to the near script primarily exposed, we decided that it was probably not necessary to adjust this region. The red regions labeled by the near script seem slightly exposed, and the protein in general seems "foamy" so we've decided we would attempt to increase the dry weights: namely, "dry5" was increased from 15 to 20, and "dry6.5" was increased from 20 to 23. try3 is currently running on lopez. Tue Jul 4 15:10:45 PDT 2006 Cynthia Hsu There... doesn't seem to be much of a difference between the models as far as I can tell in RasMol, but according to the unconstrained scores, try3 was the best model so far. Under Grant's suggestion, I've decided to up both "dry5" and "dry6.5" to 25, and "dry8" to 17. I thought that sidechains might be improved slightly, so I've decided to raise the cost a little bit to 8.5. Also, according to the unconstrained file, the score for breaks was about 0.3, so I've decided to raise the "breaks" function slightly to 170. Make started Wed Jul 5 07:23:53 PDT 2006 Running on cheep.cse.ucsc.edu Wed Jul 5 07:59:27 PDT 2006 Kevin Karplus There was some damage to the directory (it looked like ehl2 had been overwritten) so I redid "make" to clean things up. A number of the files are not group-readable, so I can't really see what Cynthia and Crissan have been doing. Wed Jul 5 10:42:59 PDT 2006 Cynthia Hsu Looking at the score-all.unconstrained.pretty file, it appears that try4 scores slightly worse than tries 3 and 2, but not by very much. Since the most notable areas of difference were in "soft_clashes" and "n_ca_c", I've raised "soft_clashes" to 60 and "n_ca_c" to 7. The make for try5 was started on whidbey at Wed Jul 5 10:46:37 PDT 2006. I've also ran the fixmode on the directory... we've changed the .chsrch file, so hopefully there won't be any more occurrences of this sort, but it's really hard to say. Wed Jul 5 16:59:28 PDT 2006 Crissan Harris, Cynthia Hsu According to Firas, errors with the system at around 10:30 AM prevented try5 from executing all the way. We've rerun the same constraints as before, but under the model of try6. try6 currently running on camano. Thu Jul 6 14:34:18 PDT 2006 Crissan Harris, Cynthia Hsu try6 is worse than tries 3, 2, and 4, when unconstrained. We've added all four models to the superimpose-best.under file. We decided to modify try3.costfcn in order to produce the try7 model. "dry6.5" was raised to 25 and "dry8" was raised to 20. We also raised "soft_clashes" to 60 and "break" to 180. We also commented out the constraints at the end: //include T0340.dssp-ehl2.constraints //include T0340.undertaker-align.sheets //include rr.constraints try7 is currently running on camano. Fri Jul 7 10:48:28 PDT 2006 Crissan Harris, Cynthia Hsu try7-opt2 is the highest scoring model, followed by try3-opt2, try2-opt2, and try4-opt2 in that order. All these were added to the superimpose-best.under file. We've decided to refrain from running more tries at the moment, until someone can tell us what else we should do to imporve the model. Mon Jul 10 16:07:20 PDT 2006 Crissan Harris, Cynthia Hsu Because of the combined results of the Rosetta and the undertaker scores, we listed the best-models as follows: ReadConformPDB T0340.try6-opt2.gromacs0.repack-nonPC.pdb.gz ReadConformPDB T0340.try1-opt2.gromacs0.repack-nonPC.pdb.gz ReadConformPDB T0340.try7-opt2.gromacs0.repack-nonPC.pdb.gz ReadConformPDB T0340.try7-opt2.pdb ReadConformPDB T0340.try3-opt2.pdb Both Rosetta favors the try6 gromacs repacked the most, followed by that for try1 and that for try7. In Undertaker, the gromacs models are not as high-ranked as try7, but between the gromacs repacked models, try1 is not as highly favored. We decided to first attempt to polish the gromacs repacked model for try6. We copied the try6.under file to try8.under, asking it to "ReadConformPDB T0340.try6-opt2.gromacs0.repack-nonPC.pdb.gz." For the try8.costfcn, we copied the try6.costfcn and then examined the Undertaker unconstrained scores to determine where try6-opt2.gromacs0.repack-nonPC.pdb.gz had performed weakly. From this, we decided that it would be best to reduce "soft_clashes" to 40 and raise "bad_peptide" to 15, but since try6-opt2.gromacs0.repack-nonPC.pdb.gz performed better than our best Undertaker model on "n_ca_c", we left this weight as is. We also raised the dry weights as follows: dry5 25 dry6.5 30 dry8 18. Following Professor Karplus's advice, we also brought two residues together, as follows: Constraint V69.CB L21.CB -10. 7.0 14.0 1 try8 is currently running on Whidbey. In deciding whether we wanted to trust Rosetta or Undertaker's scoring of which version of try7 to polish (the gromacs repacked version or the try7-opt2 version) we decided that given that try7-opt2 performed better on breaks, and all of the dry weights, we would continue to polish try7-opt2. It performed worse than the gromacs model on "soft_clashes" and "n_ca_c". "soft_clashes" was already pretty high in the try7 cost function (at 60), so we decided to leave it as is and just raise "n_ca_c" to 10. As before, we followed Professor Karplus's advice and added the constraint between V69 and L21. We also raised "dry5" to 25 and "dry6.5" to 30. try9 is currently running on Lopez. Mon Jul 10 22:19:09 PDT 2006 Cynthia Hsu try9-opt2 is scored as the highest, according to Undertaker. Rosetta still returns try6-opt2 repacked gromacs. try8 performed poorer than try9 in both models. Mon Jul 10 22:47:54 PDT 2006 Cynthia Hsu It just occurred to me that Professor Karplus might have meant for us to use try6-opt2.pdb.gz, not the repacked gromacs version. With this in mind, I have recopied try8, but changed the input file to be T0340.try6-opt2.pdb.gz. the try10.costfcn remains identical to that used in try8. try10 is currently running on camano. Tue Jul 11 12:52:52 PDT 2006 Crissan Harris, Cynthia Hsu try10 failed again, but we decided that it probably wasn't important. The superimpose-best.under file was updated with the top three Undertaker models and the top two Rosetta models. We've decided to attempt to pack it tighter, so we've changed the dry weights to the following: dry5 30 dry6.5 35 dry8 23 dry12 5 try11 started on lopez Tue Jul 11 13:45:16 PDT 2006 Crissan Harris, Cynthia Hsu try11 scored higher than try9. We've modified superimpose-best.under to include a variety of top scoring models. ReadConformPDB T0340.try11-opt2.pdb //highest scoring in Undertaker #ReadConformPDB T0340.try9-opt2.pdb //not as polished as try11 #ReadConformPDB T0340.try7-opt2.pdb //not as polished as try9 ReadConformPDB T0340.try3-opt2.pdb ReadConformPDB T0340.try2-opt2.pdb ReadConformPDB T0340.try6-opt2.gromacs0.repack-nonPC.pdb.gz //Rosetta's favorite ReadConformPDB T0340.try1-opt2.gromacs0.repack-nonPC.pdb.gz //Rosetta's second favorite ReadConformPDB T0340.try8-opt2.pdb ReadConformPDB T0340.try4-opt2.pdb Tue Jul 11 14:24:46 PDT 2006 Kevin Karplus These models are all *extremely* close. There are a few small differences in rotamers for the surface residues, and that's about it. Let's look at try11-opt2 best with most recent cost fcn try6-opt2.gromacs0.repack-nonPC Rosetta's favorite try1-opt2 fully automatic I'll compare them with a few of the top server models: ROBETTA_TS3 ROBETTA_TS4 ROBETTA_TS1 The ROBETTA_TS3 model has a different loop at D24-G29 from the rest. All the Robetta models have different termini (before P5 or after G85). Other than that, I see only small differences (rotamers of surface residues). I think we're ready submit this one try11-opt2 best with most recent cost fcn try6-opt2.gromacs0.repack-nonPC Rosetta's favorite try1-opt2 fully automatic align1 (1g9oA) align2 (1tp5A) Is there another model that differs in any significant way (say different termini)? If so we could put that in 4th position, sliding 1tp5A off the end. Tue Jul 11 14:47:31 PDT 2006 Kevin Karplus I did the submission with the following comment. We got very little diversity of models for this target. It might have been useful to polish up a model from the second-best template, instead of concentrating all our effort on the best template, just ot hedge our bets a bit. We're actually fairly confident of P5-G85, but we did not attempt to sample different conformations for the termini, nor for the loops. Model 1 is try11-opt2, our best-scoring model. Model 2 is try6-opt2.gromacs0.repack-nonPC, Rosetta's favorite sidechain repacking of one of our backbones. Model 3 is try1-opt2, our fully automatic run. Model 4 is sidechain replacement by SCWRL on an alignment to 1g9oA, our best template. Model 5 is sidechain replacement by SCWRL on an alignment to 1tp5A, our 2nd best template. Please send me e-mail if a new model is developed, or if you feel one of the other models should be submitted. Wed Jul 12 18:09:47 PDT 2006 Crissan Harris, Cynthia Hsu Following Prof. Karplus's advice, try12.under was modified to use the second-best undertaker alignment, 1tp5A. The cost function of try12 was the same as it was in try1, but with the addition of the undertaker-align sheets from 1tp5a and the helical constraints from try11-opt2. try12 is currently running on vashon. Thu Jul 13 14:41:05 PDT 2006 Kevin Karplus I did put in try12-opt2 as model 4, pushing the alignment to 1tp5A off the end, and resubmitted the prediction. try12-opt2 is not great in terms of breaks and clashes, but it is not worth the trouble to do further polishing, I think. Thu Jul 13 14:44:00 PDT 2006 Kevin Karplus OOPS. It is too late for this target. Crissan and Cynthia did not tell me that they had a new model, and I was not looking for it this morning. Oh, well---not much harm done, other than that the model4 and model5 no longer match what CASP will have for them. Wed Sep 13 14:03:26 PDT 2006 Kevin Karplus The model4 (try12-opt2) that we submitted too late was better than the align2 that we pushed off the end, but not as good as align1, which was better than any of our more complete models, the best of which was T0340.try4-opt2.repack-nonPC (not submitted). ROBETTA_TS5 was the best of the ROBETTA models, and would have been a bit better than align1 as a starting point---certainly better than anything undertaker generated. The SAM_T06_server_TS1 did better than we did by hand (and better than any of ROBETTA models).