Mon Jul 10 09:33:06 PDT 2006 T0368 Make started Mon Jul 10 09:34:43 PDT 2006 Running on lopez.cse.ucsc.edu Mon Jul 10 10:45:58 PDT 2006 Kevin Karplus BLAST gets no strong hits in PDB (best is 2dbaA, 28% id over 112 residues, E-value 0.126). The HMMs get strong hits to a.118.8.1 (TPR=Tetratricopeptide repeat) and k.38.1.1 (designs based on TPR domain). Oligomerization not specified by crystallographers. Mon Jul 10 17:45:39 PDT 2006 Sylvia Do I've been looking at the secondary structure predictions (T0368.t06 dssp-eh12) and the try1.opt2 model. There is a discrepancy at about Y166 to S174. T0368.t06.dssp-ehl2-logo has a strong prediction for a strand there, but the model shows a helix. The t04 logo sequence agrees with t06, but t2k has no strong hits to that area. 1NA3A, the top in best hits for t06, is all helix though, so probably 166-174 really is a helix. Thu Jul 20 14:16:27 PDT 2006 Sylvia Do I took the best models from the server. The ones that scored well besides our own, and that I am putting into the superimpose file are: SAM_T06_server_TS1 and ROBETTA_TS5. general agreements: A10-M37, 40-55 for 1&2; 3 is in the area, 55-70, 70-87 ( these last two - robetta is following it, but is a little short), 100-118, 115-130, 134-end, all the models are different. notes: model 3 (Robetta) seems to be straying the most from the other two - try1-opt2 and Sam I'll try again with some more servers. Adding ROBETTA_TS4 and ROBETTA_TS1, I see that that area is very different for the models. models 3,4,5 (the Robettas) agree at about 136-145, though. No major breaks to worry about. Mon Jul 24 13:27:02 PDT 2006 Sylvia Do Trp170 wants to be buried. The other side has two sides that are buried, and correctly, so some other helix sohuld cover it.PHE57 also neeeds to be buried more. Leu69 and Leu92 needs to be buried as well. Perhaps move helixes from Met1 - thr42 up a little? There are two cysteines that if of one of that helix moves up a little, may be close enough for a disulfide bond. It looks like it can be some type of bundle. #I'm trying to bring these two closer, hopefully bringing the entire helix up #a little. Constraint C33.CB C58.CB -10.0 2.5 4.0 1 I will use ProteinShop (when I can get to a computer that can use it) to try and rotate and move some of the other ones. try2 is running on lopez. Mon Jul 24 14:28:48 PDT 2006 Grant Thiltgen I looked at the alignments from the t2k and t06 alignments, and the t06 alignments appear to be better than the t2k alignments as more stuff is being conserved (taking out the crazy conserved tryptophans at the end of the protein). It does appear that the t2k alignments are the ones that are finding the closest hits for the all alpha proteins, but if the t2k alignments are not picking up the right residues, perhaps it's not an all alpha protein after all? The t2k alignment has 950 sequences in it, but the t04 and t06 alignments only have 10. What's up with that? Mon Jul 24 15:31:09 PDT 2006 Sylvia Do Oops. Just looking at try2-opt1, and the distance is father now, instead of closer. Fixed it: Constraint C33.CB C58.CB -1.0 2.5 3.0 1 try3 is running on furi. Mon Jul 24 15:40:15 PDT 2006 Kevin Karplus It is unusual, but not unheard of, for t2k to find more than t06 and t04. In this case, I think it was that the t04 and t06 alignments were looking for the conserved TRP residues, which the t2k alignment did not care as much about. They all find the same templates, so you may want to use the more specific alignments in this case. 2.5 Angstroms seems too close for a pair of CYS.CB atoms. In fact, a clash will be reported for anything closer than 3.13 Angstroms. If you are trying to get a disulfide, use SSBOND C33 C58 (and add known_ssbond 1 to the costfcn). If you just want the residues close, but not a disulfide, use a distance constraint that is reasonable (like the maybe_metal distance of 3.6 Angstroms between SG atoms). Mon Jul 24 16:33:39 PDT 2006 Sylvia Do Ohh. Okay. So I put the SSBOND c33 c58 in rr.constraints, and added known_ssbon 1 to the costfcn. ProteinShop is NOT working correctly, so I don't think I will be using that anytime soon. try4, based off try1, is running on abyss. Tue Jul 25 14:42:22 PDT 2006 Grant Thiltgen I am going to start a run based on just the t04 and t06 alignments in order to get more specific things from those instead of the diverse alignments that the t2k alignments picked up. This may make something new or it might not. It is interesting to note that there is a strand predicted in the t04 and t06 alignments that gets swamped out by the t2k alignments. I also removed all the original constraints and included the T0368.t06.str2.constraints file to build off of. Try5 started on whidbey. Tue Jul 25 15:03:24 PDT 2006 Grant Thiltgen Since the two cysteines in the protein don't seem to be conserved, I wouldn't expect a disulfide bond in this protein. Tue Jul 25 14:29:01 PDT 2006 Sylvia Do The disulfide bond seem to mess everything up, anyway... I'm going to do try6 with the following constraints: Constraint Y26.CB L62.CB -10.0 7.0 14.0 1 Constraint F57.CB F124.CB -10.0 7.0 14.0 1 Constraint C58.CB A108.CB -10.0 7.0 14.0 1 Constraint L66.CB N159.CB -10.0 7.0 14.0 1 trying to get some areas less exposed. I'm not sure this will work. One of the constraints is an attempt to flip a strand over, but I don't know how well a single constraints for that will do. At the very least, I hope that some of the exposed area will rotate a little, towards the center. try6 is running on orcas. Tue Jul 25 16:17:51 PDT 2006 Grant Thiltgen Try5 didn't work out that well. Wed Jul 26 16:36:18 PDT 2006 Sylvia Do I used Cynthia's ProteinShop to flip a helix around. I am now running try7 on shaw. Thu Jul 27 11:56:20 PDT 2006 Grant Thiltgen I think I am going to take two approaches: 1) Base things off of the t04 and t06 models, which have two predicted antiparallel strands. I am going to attempt to make a sheet out of those two strands and the rest helices. 2) Base things off of the t2k alignments, which are all helical. I am going to attempt to base things off of the helices in the secondary structure predictions. Thu Jul 27 12:28:53 PDT 2006 Grant Thiltgen I set up try8 and try9. Try8 is using the t06 and t04 alignments and using the predicted secondary structure as constraints for the costfcn. I used str2 and the notor and sep alphabets to get hydrogen bonding. I kind of arbitrarily paired the two predicted strands in an anti-parallel fashion. Try9 is using the t2k alignments and using the predicted secondary strucutre from t2k as constraints. This is all helices and hydrogen bonds. Both try8 and try9 started on whidbey. Thu Jul 27 14:31:25 PDT 2006 Grant Thiltgen So far try8 and try9 don't seem to be inproving much in way of the unconstrained costfcn. Try8 didn't really make the sheet I wanted and try9 added one in. I am going to try running them again with higher constraints. Try10 and try11 started on whidbey. Rosetta seems like to like try2 and try4 better than try1. Of these two, try2 seems more believable. Thu Jul 27 15:33:58 PDT 2006 Sylvia Do I seem to be making all my mistakes this week. -_-; I accidently wrote over Grant's try11.costfcn (but not try11.under). I changed it as to how I think Grant had it (based off try9, but with contraints raised like try12). try12 will be based of the proteinshoped try7-movearound. It's a helical bundle, but I don't really know if it will hold for undertaker. I've rotated the helices so that the hydrophobic regions are facing in. try12 will also have increased costs for phobic_fit (to 5) and breaks (to 75). Oh yea. ProteinShop works on neigh, regardless of user, so if anyone needs it, use neigh. try12 is running on lopez. Thu Jul 27 16:57:23 PDT 2006 Sylvia Do I looked at opt1 for try12. It looks like most things are in place, except for the last two helices. These two (well, now three) have been giving me a hard time, so I'm going to try to combine the two into one long helix and see what happens. For try13, I too try12's .under and costfcn. I added this to the costfcn: HelixConstraint (T0368)G140 (T0368)S174 2 and made the costs for breaks higher, 125. try13 is running on abyss. Fri Jul 28 11:14:53 PDT 2006 Grant Thiltgen Try1 still scores the best overall, so I feel like I should at least do a polishing run of model one. Try2 and try3 are our next two high scoring by the undertaker costfcn. Try2 opt2 is the best model by the rosetta scoring function. I'm also going to look at the server models to see which of these score the best and look at them to see what they did. try14 is a polish of try1 and try15 is a polish of try2-opt2.gromacs0.repack-nonPC. Both started on whidbey. Interestingly, the try1-opt2 model is the best model in the unconstrained costfcn with the servers scored as well. SAM_T06_server_TS1 is the best server model. ROBETTA_TS5 and Pmodeller6_TS2 are the next best servers. Most of the Robetta models score well. The Zhang-Server_TS3 model also scores in the top server models. Fri Jul 28 11:37:43 PDT 2006 Grant Thiltgen Hmm. It appears that almost all of the alignmnets do not have the helices in a bundle, but in fact have them in more of an arch. This makes the protein long and curved, and not bundled together. This could be why the undertaker scoring doesn't like the models that do this as much as the ones that stack it all together. Looking at the alignments, I'm more prone to believe some of the server models that do the same thing instead of our models. I think I may try to build from the beginning again, but pull out the top alignments that only have a few helices that probably caused the model to clump the way it did. Fri Jul 28 12:54:26 PDT 2006 Grant Thiltgen I have the alignments for the top 8 undertaker alignments that had the long bundle instead of the close bundle. Try16 is using the try9 costfcn but only using the alignments from the eight models from the undertaker alignments: 1hz4A 1wm5A 1xnfA 1e96B 1ihgA 1fchA 2fbnA 1elwA Also the polishing runs score slightly better on the unconstrained costfcn than the original models. They also score slightly better on the rosetta scoring function. Try16 started on whidbey. Fri Jul 28 15:52:58 PDT 2006 Sylvia Do try16 didn't do too well with rosetta or undertaker's scoring. in undertaker, it scored very badly on the phobic fit, and the drys. I've increased these for try17: dry6.5 30 dry8 25 dry12 7 \ phobic_fit 6 \ try17 is running on shaw. Sat Jul 29 13:28:16 PDT 2006 Grant Thiltgen I think the problem arises that undertaker likes to try to make everything fall into a big glob instead of trying to let the helices pack nicely along each other like the alignments have. I think the answer is to drop phobic fit, not to increase it, but I could be wrong. I am going to drop phobic fit from the try16 costfcn and try to run it as try18. try18 started on shaw. Sat Jul 29 14:25:27 PDT 2006 Sylvia Do try18 looks ...intersting. 117-174 is now one long helix. So is 45-114. As this does not look like the alignments, I don't think this is a good one. Unfortunately, I can't think of anything else to do for this target right now, either. Sat Jul 29 17:25:23 PDT 2006 Grant Thiltgen It appears that completely removing phobic fit from the model is a bad idea. Good to know. I could try reducing it to one but I'm not sure it will do what I want either. I may just have to look at some of the better server models and see which of them do best and attempt to optimize them in undertaker (although I am worried that it is going to be a problem and undertaker is going to try to colapse the helices.) Sat Jul 29 17:50:43 PDT 2006 Grant Thiltgen Ironically, try18, which looks horrible scores best by the Rosetta scoring function. Try19 is going to be a polish of try14, and try20 is going to be a polish of try15. These are our best two models currently, and try15 is favored by rosetta. Both try19 and try20 started on shaw. Try21 and try22 are working on polishing models sent out by the servers. Try21 is polishing the top three scoring robetta models, which look more like the models aligned from than our models do. Try22 is polishing Zhang-Server_TS3, which is the top model from another server scored on our scoring function. Both try21 and try22 started on whidbey. Sat Jul 29 22:06:33 PDT 2006 Grant Thiltgen Hmmm. Both of the polishing runs of the server models pulled in some of the helices on the outside. I don't think there's a way to get undertaker to allow the helices to stay out and allow the protein to be longer and thinner instead of making the protein more globular. Some of the PDB files of the models we aligned to had some ligands bound in the cleft that the helices made. I think that's one reason why undertaker wants to collapse it, because there's nothing there in these models. For submission should we submit the two best models we have (the one that was polished from the first undertaker run and the best rosetta scoring model), one model that is the bundle that Sylvia proteinshopped, and two alignments that match the shape of the protein that we like? I updated the best-models.pdb.gz file to show these models. So the best models are try19, which is the best scoring undertaker model based off of the first run, try15, which is the best scoring rosetta model that is based off of try2, try12, which is a bundle that sylvia proteinshopped, and two alignments that show the helical packing I'd like to get out of undertaker. Sun Jul 30 16:10:01 PDT 2006 Kevin Karplus I modified superimpose-best.under slightly, to try to make the E74-L83 helix superimpose, so that differences in structure between models would be more evident. Sun Jul 30 16:13:55 PDT 2006 Kevin Karplus I see the differences between the models. I think I agree with Grant that the alignments look more promising than the other models. Getting undertaker to hold onto a helix arrangment usually requires creating distance constraints measured off the alignment-based model. We may want to automate this in future. Sun Jul 30 19:17:27 PDT 2006 Kevin Karplus # Let's try constraints from align2 (1elwA): Constraint A10.CB I39.CB -10 5.4 8 1 Constraint A10.CB S40.CB -10 4.6 8 1 Constraint A13.CB A36.CB -10 3.8 8 1 Constraint Y26.CB L69.CB -10 4.8 8 1 Constraint A30.CB A65.CB -10 5.2 8 1 Constraint C33.CB L62.CB -10 4.0 8 1 Constraint A56.CB F86.CB -10 3.9 8 1 Constraint A56.CB Y85.CB -10 4.1 8 1 Constraint H59.CB K81.CB -10 4.4 8 1 Constraint F72.CB L114.CB -10 5.0 8 1 Constraint L76.CB A110.CB -10 3.6 8 1 Constraint V104.CB V127.CB -10 3.5 8 1 Constraint A108.CB E123.CB -10 4.0 8 1 Constraint A108.CB F124.CB -10 5.2 8 1 Constraint M121.CB M144.CB -10 3.6 8 1 Constraint F124.CB K141.CB -10 4.1 8 1 Sun Jul 30 19:32:36 PDT 2006 Kevin Karplus Optimization from alignments with these constraints (scaled down to 0.2) and dssp-ehl2 constraints started as try23 on abyss. Sun Jul 30 20:53:26 PDT 2006 Kevin Karplus I'm not real excited about the model in try32-opt1, but it might be worth including in the mix to replace one of the alignment-only models. Sun Jul 30 21:24:25 PDT 2006 Kevin Karplus try23-opt2 does not score all that well even with the try23 costfcn (try19, try14, try15, try1, try21, try22, try20, try2 all score better). I guess I'll go with Grant's ordering, as I don't have time to come up with more sophisticated scaffolding to keep the ring of helices from collapsing. The reason for choosing the alignment to 1elwA rather than 1na3A is simply that it is longer---they are otherwise in pretty good agreement. Sun Jul 30 21:50:55 PDT 2006 Kevin Karplus Submitted with comment For T0368, we had good hits to the Tetratricopeptide repeat family (alpha/alpha superhelix fold). Unfortunately, this is not a compact globular fold, and undertaker kept trying to collapse it into something more spherical. Because of this unfortunate tendency, we have included a couple of models from alignments without optimization, to show what we believe the core of the protein looks like. Model 1 is try19-opt2, the best scoring model with the unconstrained costfcn. try19-opt2 was optimized by undertaker from try14-opt2 in turn from try1-opt2, our fully automatic model. Model 2 is try15-opt2.gromacs0.repack-nonPC, the model that Rosetta likes best to repack sidechains for (other than a ridiculous model that has very long helices and no hydrophobic core). This model was generated from try15-opt2 by reoptimizing with gromacs and repackinghte sidechains with rosetta. try15-opt2 was optimized by undertaker from try2-opt2.gromacs0.repack-nonPC, and try2 was optimized from alignments by undertaker. Model 3 is try12-opt2, was optimized by undertaker from a model hand-tweaked with ProteinShop, starting with try7-opt2, which was optimized by undertaker from an earlier ProteinShopped model: try12-opt2 < try7-movearound < try7-opt2 < try6-fliphelix < try6-opt2 < alignments Model 4 is sidechain replacement by SCWRL on an alignment to 1elwA. Model 5 is sidechain replacement by SCWRL on an alignment to 2fbnA. ------------------------------------------------------------ Fri Aug 11 03:30:08 PDT 2006 Kevin Karplus T0368 was selected as a refinement target, though residues 158-174 are missing in the crystal structure. tr368 scores adequately, but its score can be improved by running it through gromacs and repacking the sidechains. The problem for refinement, as for our own modeling, will be keeping undertaker from collapsing the model into something more globular, particularly near the C-terminus. Fri Aug 11 06:58:26 PDT 2006 Kevin Karplus I wonder if oligomerization will help. Our top template (1na3A) is a monomer, but the next hit (1elwA) dimerizes. 2fbn is not in PQS, and 1fchA is octameric. We shouold probably look at these tempaltes and see if the multimer structure helps stabilize the monomers. Mon Aug 14 14:45:19 PDT 2006 Firas Khatib After looking at the different templates with Kevin, and trying to figure out whether to dimerize this model, we noticed that almost all the templates have ligands, so we will try to introduce one to act as a scaffold. Mon Aug 14 15:18:31 PDT 2006 Firas Khatib I ran VAST on tr368 and got the following hits: 1hz4 - 158 aligned residues: from 5-163 (second domain in cavity) 1zb1 - 148 aligned residues: from 121-318 (N-terminus in cavity) 1wm5 - 139 aligned residues: from 2-157 (sheet and C-terminus in cavity) 2br9 - 137 aligned residues: from 4-200 (two ligands in cavity) 1e96 - 135 aligned residues: from 2-152 (sheet and C-terminus in cavity) 1xnf - 135 aligned residues: from 34-170 (three helices in cavity) 2fbn - 135 aligned residues: from 30-170 (dimerizes at the cavity) 2ffn - 134 aligned residues: from 106-257 (other domain in cavity) 1w3b - 133 aligned residues: from 167-301 (dimerizes on back of cavity!) 2fo7 - 133 aligned residues: from 1-134 (monomer, looks very similar) 1u6g - 129 aligned residues: from 227-407 (multimer in cavity) so far, the only ones with ligands are 1elw, 2br9 and 1fch Mon Aug 14 17:10:20 PDT 2006 Firas Khatib I have inserted the ligand (using Proteinshop) that I took from 1elw. There are no clashes according to Proteinshop and the spacefill looks decent. I called it: tr368.with.ligand.renum.pdb I am going to input it for try24 try24.costfcn starts easy: SetCost wet6.5 15 near_backbone 5 way_back 5 dry5 15 dry6.5 20 dry8 15 dry12 5 \ phobic_fit 2 \ sidechain 5 \ n_ca_c 5 bad_peptide 10 \ bystroff 5 \ soft_clashes 20 backbone_clashes 2 \ break 100 \ pred_alpha2k 2 \ pred_alpha04 2 \ pred_alpha06 2 \ hbond_geom 7 \ hbond_geom_backbone 10 \ hbond_geom_beta 50 \ hbond_geom_beta_pair 100 \ missing_atoms 1 \ maybe_metal 0.5 \ maybe_ssbond 1 try25 is more of a polishing run: SetCost wet6.5 15 near_backbone 5 way_back 5 dry5 15 dry6.5 30 dry8 25 dry12 5 \ phobic_fit 2 \ sidechain 5 \ n_ca_c 5 bad_peptide 10 \ bystroff 5 \ soft_clashes 40 backbone_clashes 2 \ break 100 \ pred_alpha2k 2 \ pred_alpha04 2 \ pred_alpha06 2 \ hbond_geom 7 \ hbond_geom_backbone 10 \ missing_atoms 1 both try24 and try25 are running on shaw Tue Aug 15 10:31:29 PDT 2006 Firas Khatib well... that didn't work! The ligand didn't hold anything in place. time for some constraints! Tue Aug 15 15:15:34 PDT 2006 Firas Khatib so I realized that I didn't have KnownBreak in the try.costfcn file so I added that to try26. That of course did not work since the .a2m file didn't have the residues in the ligand. Wed Aug 16 15:56:01 PDT 2006 Firas Khatib so I created a 'ligand' directory and am running try27 there. try1-opt1 seems to have moved the ligand, so it seems that constraints will definately be necessary. Wed Aug 16 17:33:40 PDT 2006 Firas Khatib I am adding in constraints for try29, but something is wrong with the numbers. It thinks that V181 is the last residue (D182) for some reason. Maybe the knownBreak at 175 messes stuff up? Thu Aug 17 14:02:29 PDT 2006 Firas Khatib well, try28 kept the ligand in place, but all the helices basically wrap around it. Looks like many constraints between the helices will be needed. I finally got the constraints for try29 to work (I had the incorrect fasta sequence, only 1 GLU where there should have be 2!) so I will see if those constraints help! Thu Aug 17 15:04:55 PDT 2006 Firas Khatib I can't believe it, but it seems to have worked for try29-opt1! only the N-terminal helices moved, but not by much! Hooray for scaffolding! Hopefully opt2 will be even better! Thu Aug 17 15:56:36 PDT 2006 Firas Khatib try29-opt2 came out well also! You can see all 3 (the original and the try29's) in best-models.pdb. I am running try30 on the original with the same constraints as try29 but with polishing run weights set. Fri Aug 18 12:57:01 PDT 2006 Firas Khatib try30 worked as well! try29-opt2 scores a tiny bit better, but I will do more polishing runs. I will try doing polishing runs off of try29 and try30's models. try31 will use try29-opt2 as input try32 will use try30-opt2 as input both are running on vashon using try30's costfcn Fri Aug 18 15:40:52 PDT 2006 Firas Khatib all right! try32 and try31 both score even better than try30 and try29! Tue Aug 22 11:55:00 PDT 2006 Kevin Karplus What happens if we try polishing try32-opt2.gromacs0.repack-nonPC (rosetta's favorite after try27 and try28, which we think are messed up---undertaker says they are incomplete) *without* constraints? Do the helices stay in place or do they collapse? There are no sheets, so we should turn of the hbond_geom_beta parts of the cost fcn. C33 and C58 seem to come quite close: are they a disulfide or a metal-binding site? There *are* several HIS residues in the vicinity. The CYS residues are not conserved (the multiple alignments have YFLAC for C33 and ALV for C58), so a disulfide bond seems unlikely. The top hit (1na3A) has no cysteine. Let's turn off maybe_ssbond, but leave maybe_metal in place. Sigh---I wish we'd thought of putting in the ligand on the original optimization! Tue Aug 22 12:33:56 PDT 2006 Kevin Karplus ligand/try33 is optimizing the gromacs0.repack-nonPC models for try29-try32, but seems to be preferring to work mainly with try30-opt2, rather than try32-opt2. Breaks and clashes are reducing a lot in try33, but I don't know yet whether the helices are staying in place or collapsing. If the model is collapsing, then I'll need to try a polishing run with constraints, but less strong than in try32. When I put together the submission of the refinement models, I'll have to remove the ligands, which is probably easiest done with emacs, rather than creating a new script or adding new features to undertaker. Tue Aug 22 12:44:28 PDT 2006 Kevin Karplus try33-opt1 seems to have been optimized from try32-opt2.gromacs0.repack-nonPC, rather than try30-opt2.gromacs0.repack-nonPC. Tue Aug 22 12:53:48 PDT 2006 Kevin Karplus Weird! try33-opt1 has attached the ligand to the end of the chain, despite the KnownBreak G175 command. Despite the loss of the ligand from the binding pocket, try33-opt1 has not moved the helices around much. Perhaps we should pick up the unliganded try30, try32, and try33 models, and try optimizing them. (I'll wait until try33-opt2 is finished.) Tue Aug 22 15:20:37 PDT 2006 Firas Khatib That is indeed very odd about try33 attaching the ligand to the C-terminus! You'd think it would change the rest of the fold, but surprisingly it doesn't. I don't understand how that could happen when the helices clearly colapsed on themselves before we used the ligand, but I guess if it seems to be working, then we should try the previous runs without the ligand. Tue Aug 22 16:32:49 PDT 2006 Kevin Karplus I copied try30-opt2, try31-opt2, try32-opt2 and try33-opt2 from ligand/decoys to decoys, and removed the extraneous G175 on. I also copied over try30.costfcn through try33.costfcn, removing all the references to residues of the ligand. I actually have an explanation for why it was necessary to put in the ligand initially, but not later. Once residues have packed together in a local minimum, it is harder for movement to improve the score. Closing gaps and optimizing sidechains with the ligand in place got into a local minimum that no longer needed the ligand to prevent large-scale movement. (I don't know that this is the *right* explanation, but it is a plausible one.) Of the ligand models, rosetta likes try33 the best by a lot, probably because it does not realize that the ligand is supposed to be a separate chain. By specifying the MONOMER_LENGTH=175 and doing make T0368.mult30 ... T0368.mult33, I can unpack the long chain into two chains and optimize them with gromacs and repack them with rosetta. Since Rosetta now knows they are two chains, the scores are much more reasonable. With the ligand, rosetta likes best try33-opt2.unpack.gromacs0.repack-nonPC, try30-opt2.unpack.gromacs0.repack-nonPC, try31-opt2.unpack.gromacs0.repack-nonPC, try32-opt2.unpack.gromacs0.repack-nonPC. Without the ligand, rosettal likes best try31-opt2.gromacs0.repack-nonPC, try33-opt2.gromacs0.repack-nonPC, try30-opt2.gromacs0.repack-nonPC, try32-opt2.gromacs0.repack-nonPC, Let's try polishing these 4 without the ligand. Tue Aug 22 16:48:08 PDT 2006 Firas Khatib Hopefully your theory is correct, because it would mean that we have found a nice local minimum compared to the refinement structure that we began with (since Undertaker collapsed the helices when starting with the refinement model, but no longer does). This is hopefully a good sign! Tue Aug 22 17:00:48 PDT 2006 Kevin Karplus I started the try34 optimization run with no ligand---we'll see if it behaves itself. Unfortunately, I left InsertSpecificFragment with a high pseudocount (I probably should have upped CrossOver instead). This won't ruin the run, but may make it a little less efficient about picking up good stuff from the multiple models. Tue Aug 22 17:34:57 PDT 2006 Kevin Karplus try34 is based mainly on try30-opt2.gromacs0.repack-nonPC try34-opt1 has moved very little from try30, and is fairly close to the original tr368 model. Wed Aug 23 00:05:34 PDT 2006 Firas Khatib try34-opt2 seems very close to try33, and doesn't have that deviation that try30 has around G156-K167. it does score better using the unconstrained costfcn as well as try34's costfcn. Wed Aug 23 10:05:05 PDT 2006 Kevin Karplus rosetta still likes try31-opt2.gromacs0.repack-nonPC best of the refined models, though try34-opt2.gromacs0.repack-nonPC is not far behind. try34 is much closer to try30 than to try33. try31-opt2.gromacs0.repack-nonPC seems more densely packed than try34. It also has smaller breaks, but larger clashes. Perhaps I should do a polishing run with breaks and dry terms turned up to try to get tighter packing. Wed Aug 23 10:19:43 PDT 2006 Kevin Karplus try35 started on cheep to do one more polishing. I don't expect much change this time, and will probably submit the top 5 runs after this. Wed Aug 23 11:40:10 PDT 2006 Kevin Karplus undertaker does not like try35 quite as much as try34 (even with the try35 costfcn), but rosetta likes try35-opt2.gromacs0.repack-nonPC best of any of the refinement models. I will submit ReadConformPDB T0368.try34-opt2.pdb # best try35.costfcn ReadConformPDB T0368.try35-opt2.gromacs0.repack-nonPC.pdb # best robetta ReadConformPDB T0368.try31-opt2.gromacs0.pdb # best gromacs with try35.costfcn (not try34 or try35) ReadConformPDB T0368.try33-opt2.repack-nonPC.pdb # best rosetta without gromacs (not try35) ReadConformPDB T0368.try32-opt2.pdb # other decent choice Wed Aug 23 12:07:38 PDT 2006 Kevin Karplus Refinement models submitted. Wed Aug 23 12:49:21 PDT 2006 Firas Khatib my mistake... I was looking at best-models.pdb and thought that the red chain (the last chain) was try30, but didn't realize that tr368.with.ligand.renum.pdb was the last chain in superimpose-best.under! try34 indeed looks a lot like try30.... my bad.