Mon Jun 19 08:34:10 PDT 2006 T0334 Make started Mon Jun 19 08:34:49 PDT 2006 Running on shaw.cse.ucsc.edu Mon Jun 19 08:37:47 PDT 2006 Kevin Karplus BLAST finds 2aqjA (55% id over 526 residues, E-value 1e-159) and not much else. This looks like a full-length comparative model. Note: 2aqj is a dimer according to PDB and PQS, but the pdb file has only one monomer in the asymmetric unit so we'll have to use symmetry operators (ours or others) to create the dimer. Mon Jul 3 17:21:09 PDT 2006 Crissan Harris, Cynthia Hsu Selecting the Cis and His residues in Rasmol indicated that there were no metal bonding or disulfide bond sites, so these values were set to zero. A lot of breaks were observed in the model, so the value of this was adjusted from 50 to 120. "soft_clashes" value was increased to 50. At residue 504, we observed strong discrepancies between the predictions and the previous models - t2k predicted it to be a coil, instead of a helix; nor did it appear to be a helix in any of the previous models. We also observed hydrophobic regions from residues 502 to 506, and we removed the helix constraints on these with the intention of folding them closer towards the core of the model, as shown below: HelixConstraint (T0334)M492 (T0334)E501 1 HelixConstraint (T0334)K507 (T0334)T516 1 We were also concerned about the region from residues 482-493, which according to the t2k search should appear as a coil, but in the model exhibit some helical properties. It appears to be composed of primarily hydrophobic residues, which is a concern as it is protruding into the environment. However, since it is far from the rr residues (which we are assuming to be the active site of the protein), we decided to leave it as is in case it is the dimer interface. try2 is currently running on camano. Tue Jul 4 15:48:11 PDT 2006 Cynthia Hsu According to the unconstrained scores, try2 performed worse than try1 did, with the only significant improvements taking place in the level of "breaks". Moreover, removing the helix constraint from residues 502-506 still had no effect on the conformation of the protein, and its helical structure matches that of the model. In view of this, I decided to edit try1.costfcn for try3. "soft_clashes" was raised to 80 and "breaks" was raised to 200. As before, "maybe_metal" and "maybe_ssbond" were set to 0. try3 is currently running on orcas. Wed Jul 5 09:52:31 PDT 2006 Cynthia Hsu It appeared that just by changing those two conditions, the protein molecule is made much much worse (random coils floating separate from the protein and a couple of alpha helices stretched very far away). For try4, try1.costfcn was copied. As before, "maybe_metal" and "maybe_ssbond" costs were both removed. Breaks were set to 140; hopefully lowing this cost value will prevent the strange configurations that were produced in trys 2 and 3. In an attempt to further close the breaks, I've decreased "dry8" to 12 and "dry12" to 3, hoping that this will give the molecule a little bit more flexibility. Also, "n_ca_c" has been raised from 5 to 10. Hopefully, this will close the breaks in the backbone as well. I considered adding individual constraints to try to seal the backbone, but decided that it was probably better off leaving it as is, since there were far too many to account for all and I don't know how much this sort of meddling will mess things up. try4 is currently running on lopez. Wed Jul 5 16:53:22 PDT 2006 Crissan Harris, Cynthia Hsu According to Firas, there were errors with the computers this morning at around 10:30 AM that would explain why our try4 did nothing except produce a bus error. George suggested that we repeat the run as try5. try5 is currently running on lopez. Wed Jul 5 17:41:48 PDT 2006 Crissan Harris, Cynthia Hsu After discussing our progress w/ Prof. Karplus, we killed try5, since we made several errors in modifying our costfcn. Namely, it would have been preferable to copy the costfcn from the try2 model, since its unconstrained score was significantly better in the criteria that would be more important in a polishing run, such as soft_clashes and breaks. We made the following changes to the try6.under file, since it was a polishing run: We commented out all the TryAllAlign, NameComform, and the Include lines, as well as the ReadFragmentAlignment. We also included the following lines: // Switch to the following (commenting out the TryAllAlign) // to polish existing models InfilePrefix decoys/ include read-pdb.under Now that try6.under has been set up to appropriately execute a polishing run, we made the desired changes to the try6.costfcn file (which was copied from try2.costfcn): "soft_clashes" was raised to 80 and "breaks" were raised to 200. Examining the scores, we also decided to raise "n_ca_c" from 5 to 10. We also raised "sidechain" from 5 to 7. try6 is currently running on lopez. Thu Jul 6 10:23:54 PDT 2006 Crissan Harris, Cynthia Hsu try6 performed the best of the previous tries, but we are still concerned about the level of breaks. To fix an especially large one prior to K83, we lowered the SheetConstraint on the beta sheets in that region as follows in hopes that this would allow undertaker to focus on closing the breaks, instead of forming beta sheets: SheetConstraint (T0334)A76 (T0334)Y78 (T0334)F171 (T0334)W169 hbond (T0334)S77 0.2 SheetConstraint (T0334)I82 (T0334)I85 (T0334)T273 (T0334)I276 hbond (T0334)K83 0.5 There were other breaks, namely before F171, T258, K378, and V256,but we hoped that these would be closed by merely raising the "break" cost to 240. try7 currently running on Whidbey. Thu Jul 6 15:50:22 PDT 2006 Crissan Harris, Cynthia Hsu Grant helped us make a dimer using the 2aqj alignment and our try6-opt2 model. The dimer looks a little bit foamy, but visualizing with the near and burial scripts in RasMol do not show any huge discrepancies between what should be and what is or is not exposed. This was performed using the make-dimer.under file; subsequent make-dimer1.under files for the different models will be numbered. We had to wait for try7 to finish running on Whidbey, but after that's done, we'll try dimerizing try7 if it scores higher on unconstrained than try6 does. Thu Jul 6 21:06:08 PDT 2006 Crissan Harris, Cynthia Hsu try7 is definitely the best overall of our monomer models, according to the unconstrained scores (with try 3 being close behind). We created the make-dimer2.under file for our second try using try7 as the monomer. Fri Jul 7 10:33:28 PDT 2006 Crissan Harris, Cynthia Hsu The protein still looks very foamy, so we've decided to raise the dry weights as follows: dry5 20 dry6.5 25 dry8 20 dry12 10 try8 was started on lopez on Fri Jul 7 10:44:59 PDT 2006 Fri Jul 7 16:36:25 PDT 2006 Crissan Harris, Cynthia Hsu try8-opt2 and try8-opt1 both score very well in comparison to the previous models. We've decided that it would be in our best interest to begin optimizing it as a dimer. make-dimer3.under is dimerzing try8-opt2. mult1 (try1) is running on lopez. Sat Jul 8 13:41:19 PDT 2006 Cynthia Hsu the dimerized try1 appears completely disintegrated in RasMol. I'm not entirely sure what happened. I've set the "TweakMultimer" value in try2.under to equal 0, in hopes that that will keep the dimer clustered together. I also removed the "multimer 2" parameter to OptConform; our dimer probably wasn't a cyclic one and that may have been the reason it unwound itself as it did. I also reduced the "num_gen" to 40, and the second "super_num_gen" to 250. mult2 is currently running on camano. Sat Jul 8 18:01:20 PDT 2006 Kevin Karplus Looking at the 2aqj "dimer", I'm not surprised that the optimization as a dimer did not seem to work well---the interface betwen the who chains is really quite minimal. Since we don't have to do a dimer submission for this model, and dimeric constraints are not likely to have much effect on the monomer (given how little the chains appear to interact), we are probably best off just doing monomeric optimization on T0334. What models do we have? best with unconstrained: try8-opt2 rosetta's favorite: decoys/T0334.try7-opt2.gromacs0.repack-nonPC.pdb.gz fully automatic: try1-opt2 alignments align1 (to 2aqjA) align2 (to 1pn0A) Are there any different, better models? I'd be glad to drop the 1pn0A alignment to add a different model. Sat Jul 8 18:13:22 PDT 2006 Kevin Karplus I created a new "align1.costfcn" for evaluating models. It uses high break and clash penalties (like try8) but uses constraints from T0334.dssp-ehl2.constraints and SheetConstraints from alignment T0334-2aqjA-t2k-local-str2+near-backbone-11-0.8+0.6+0.8-adpstyle5.a2m The align1 costfcn scores try8-opt2 best, but try3, with much worse breaks and clashes, fits the constraints better. Unfortunately, try3 is pretty useless to us, as the helices after A427 have been ripped away from the main body of the model. For try7 and try8, the biggest problem seems to be the loop G91-G102, which is flopping out away from the rest of the model. Other than A97, there are no residues in this region predicted to be buried, so perhaps it *does* just flop around in solution. It is nowhere near the conserved residues, so is probably not functionally essential. Cynthia and Crissan, do you have any ideas for other models to include? Sat Jul 8 20:49:50 PDT 2006 Cynthia Hsu I can't really think of any other models to include. try6, try7, and try8 were the only ones that appeared anything like the way a protein should be, and the unconstrained scores seem to indicate that try8-opt2 is probably the best one. I don't have much confidence in anything else that we've produced. Sun Jul 9 13:40:51 PDT 2006 Kevin Karplus OK, I'll do a preliminary submission of best with unconstrained: try8-opt2 rosetta's favorite: decoys/T0334.try7-opt2.gromacs0.repack-nonPC.pdb.gz fully automatic: try1-opt2 alignments align1 (to 2aqjA) align2 (to 1pn0A) I'll also start an optimization run from rosetta's favorite, to see if we can pack it any better. Sun Jul 9 13:53:12 PDT 2006 Kevin Karplus Preliminary submission done. try9 started on cheep. Sun Jul 9 14:14:48 PDT 2006 Cynthia Hsu I thought that as a last attempt to try to optimize the model, I'd raise "dry5" to 25, "dry6.5" to 30, and "dry8" to 22. I'm not sure if this will have the desired effect or not, but we'll see. try10 currently running on shaw. Sun Jul 9 17:58:24 PDT 2006 Kevin Karplus try9-opt2 new best score, and rosetta now likes decoys/T0334.try9-opt2.gromacs0.repack-nonPC.pdb.gz best. Still waiting to see how try10 does. Sun Jul 9 21:08:41 PDT 2006 Kevin Karplus try10-opt2 slightly beats try9-opt2 on the try10 costfcn, due to better sidechains (though it is slightly worse on breaks and clashes). Mon Jul 10 11:12:20 PDT 2006 Cynthia Hsu I've decided to increase the weight of "break" to 260, and raise "soft_clashes" to 100. "dry6.5" was lowered to 28, to hopefully allow some more freedom of movement, and "dry8" was returned to the value of 20. I also set "n_ca_c" to 8, to give the target a little bit more flexibility. [Tue Jul 11 12:44:27 PDT 2006 Kevin Karplus Making n_ca_c smaller is a bad idea when polishing---it is there to fix clearly erroneous bond angles introduced by undertaker or gromacs. ] try11 is currently running on shaw. Mon Jul 10 16:46:52 PDT 2006 Crissan Harris, Cynthia Hsu Rosetta lists the following as the best-scoring models: decoys/T0334.try9-opt2.gromacs0.repack-nonPC.pdb.gz decoys/T0334.try10-opt2.gromacs0.repack-nonPC.pdb.gz decoys/T0334.try7-opt2.gromacs0.repack-nonPC.pdb.gz decoys/T0334.try11-opt2.gromacs0.repack-nonPC.pdb.gz decoys/T0334.try8-opt2.gromacs0.repack-nonPC.pdb.gz In comparing try9-opt2.gromacs0.repack-nonPC.pdb.gz to try11-opt2 (which scored the highest in undertaker), try11-opt2 performed better in everything with the exception of "soft_clashes". The second favorite of Rosetta's, try10-opt2.gromacs0.repack-nonPC.pdb.gz, did better on "n_ca_c" and "soft_clashes". In view of this, we've decided to attempt to polish it by raising the dry weights, and "bad_peptide". try12.under was set to ReadConformPDB T0334.try10-opt2.gromacs0.repack-nonPC.pdb.gz. In the costfcn, we raised the dry weights as follows: dry5 30 dry6.5 33 dry8 25. "bad_peptide" was raised to 15. try12 is currently running on shaw. We've also decided to polish try11-opt2 by raising "n_ca_c" to 12 and lowering "soft_clashes" to 80. We also changed the dry weights as followed: dry5 30 dry6.5 32 dry8 22 Mon Jul 10 22:25:18 PDT 2006 Cynthia Hsu I just examined try12, and it appears that it doesn't really seem to perform better than try11. I also noted that apparently we never ran try13, so that is currently running on lopez. Tue Jul 11 10:05:21 PDT 2006 try13 performed the best out of the unconstrained scores, while Rosetta favors try12-opt2.gromacs0.repack-nonPC. I've modified the superimpose-best.under file to read the following: // include read-pdb.under ReadConformPDB T0334.try13-opt2.pdb #best unconstrained ReadConformPDB T0334.try11-opt2.pdb #2nd best, undonstrained ReadConformPDB T0334.try10-opt2.pdb # best unconstrained ReadConformPDB T0334.try12-opt2.gromacs0.repack-nonPC.pdb # rosetta's favorite ReadConformPDB T0334.try9-opt2.gromacs0.repack-nonPC.pdb # rosetta's second favorite # ReadConformPDB T0334.try6-opt2.pdb # ReadConformPDB T0334.try3-opt2.pdb # C-terminal helices ripped out #ReadConformPDB T0334.try1-opt2.pdb # automatic Of the two large (>1) breaks that still exist in try13 and try11, both of try11's breaks are approximately 1.2, whereas try13 still has one break of about 1.5. In view of this, we've decided to run one last try, using the try13 costfcn but with a higher value for breaks. "breaks" was increased to 290 and "soft_clashes" were raised to 90. try14 is currently running on shaw. Tue Jul 11 12:47:39 PDT 2006 Kevin Karplus I would like to get rid of the 2 large breaks. I'd also like to compare our models with the best-scoring server models (ROBETTA_TS3 and ROBETTA_TS1, for example). Tue Jul 11 13:01:02 PDT 2006 Kevin Karplus Our models are all almost identical, while the ROBETTA models try different variants on the loops: W87-F107 G133-A142 L318-R327 G440-D450 It may be worthwhile to do another run from the alignments (like try1, but with just the constraints from try13-opt2.sheets and try13-opt2.helices), to sample different loop configurations. Tue Jul 11 13:14:39 PDT 2006 Kevin Karplus try15 started on cheep. This may not really have been necessary, as try2 and try3 also started from aligments, so have already sampled different loops. Where do the current best (either try12-opt2.gromacs.repack-nonPC or try13-opt2) come from? try13-opt2 < try11-opt2 < try10-opt2 try12-opt2 < try10-opt2.gromacs0.repack-nonPC try10-opt2 < try8-opt2 < try7-opt2 < try6-opt2 < try2-opt2.repack-nonPC So it looks like polishing try1-opt2 and try3-opt2 might provide slightly different models. I was going to make try2-opt2.gromacs0.repack-nonPC with make T0334.do2, but Cynthia has edited try2.under since try2-opt2 was made, so make wants to remake the pdb file (which I do *not* want). I can do make decoys/T0334.try2-opt2.gromacs0.repack-nonPC.pdb.gz \ decoys/grep-best-rosetta -W decoys/T0334.try2-opt2.pdb.gz to have the desired effect. Tue Jul 11 13:37:22 PDT 2006 Kevin Karplus It looks like try2-opt2 was the best of the three starting models we had (after gromacs0.repack-nonPC), so just doing a single polishing run for each of try1-opt2.gromacs0.repack-nonPC and try3-opt2.gromacs0.repack-nonPC should be enough to make some alternative models. I'll set up try16 and try17 to do that polishing. Tue Jul 11 13:46:48 PDT 2006 Kevin Karplus try16 and try17 started on abyss. Tue Jul 11 19:55:59 PDT 2006 Kevin Karplus try16 scores better than try17, and I suspect both will be better than try15 (which hasn't quite finished yet). Tue Jul 11 20:14:51 PDT 2006 Kevin Karplus Actually, try15 does better on the unconstrained costfcn than either try16 or try17, though try16 does better on the try16=try17 costfn. Tue Jul 11 20:19:37 PDT 2006 Kevin Karplus Oops, try17 is a mess, with helices ripped out of the core and stuck out. I had in my notes about that try3-opt2 had this problem, but I didn't re-read the notes---so I wasted several hours of computer time. I guess I can do either try1-opt2 or an alignment for model5. Perhaps the alignment would be most informative. Tue Jul 11 20:27:52 PDT 2006 Kevin Karplus Submitted with comment Most of the models we generated were extremely similar to each other, so on the last day before the deadline, we made an attempt to generate some different models by going back to the alignments and building models from scartch again. These models did not get the extensive polishing of the first two models we are submitting, but are included to attempt to cover the space of possiblities a bit more. Model 1 is try14-opt2, the best with the unconstrained costfcn. Model 2 is try12-opt2.gromacs0.repack-nonPC, rosetta's favorite after repacking sidechains. It is try12-opt2 run through gromacs with forcefield 0, then through rosetta, choosing rotamers for all residues except PRO and CYS. Model 3 is try16-opt2, a further optimization of try1-opt2, our fully automatic run. Model 4 is try15-opt2, an independent run from the alignments with constraints taken from try13-opt2, but without extensive optimization to close gaps or remove clashes. Model 5 is sidechain replacement by SCWRL on an alignment to 2aqjA