Fri Jul 14 10:50:46 PDT 2006 T0380 Make started Fri Jul 14 10:51:22 PDT 2006 Running on cheep.cse.ucsc.edu Fri Jul 14 15:45:04 PDT 2006 Kevin Karplus BLAST gets one good hit: 2fhqA 29% over 129 residues (4.7e-11) The HMMS like that best, but also like 1dnlA, 1g79A 1nrgA, ... (all SCOP b.45.1.1) These are all the same fold as what BLAST likes. Wed Jul 26 17:25:38 PDT 2006 Navya swetha Davuluri I checked 2fhq in PDB and it exists as a dimer. There is also a break before residues 78 and 45. I am going to first fix the breaks and then make a dimer. I increased the weight on breaks to 100 and on soft_clashes to 40. I want to pack it more tightly so I increased weight on dry6.5, dry8 and phobic_fit too. try2 started on peep. Wed Jul 26 21:12:03 PDT 2006 Navya swetha Davuluri. try2-opt2 still has breaks. I am going to increase the weight on breaks and soft_clashes further and will start a new run. try3 is an optimization of try2-opt2. try3 started on peep. Mon Jul 31 15:22:12 PDT 2006 Navya swetha Davuluri. The break before residue 45 is still not fixed. I am going to increase the weight on breaks to fix it. I also decreased weight on hbond_geom_beta and hbond_geom_beta_pair. try4 running on shaw. Rosetta likes best try1-opt2.gromacs0.repack-nonPC unconstrained likes best try4, try3, try2, try1 All the models are very similar, and quite close to the first model from alignment. I think I'll try polishing the try1-opt2.gromacs0.repack-nonPC file, which rosetta currently likes best. Tue Aug 1 09:06:35 PDT 2006 Kevin Karplus try5 started on whidbey to polish try1-opt2.gromacs0.repack-nonPC file. Tue Aug 1 10:13:47 PDT 2006 Kevin Karplus try5-opt2 is new best with try5.costfcn, though unconstrained prefers try4, partly from pred_alpha partly from better hbonds. Rosetta now likes best try5-opt2.gromacs0.repack-nonPC All our models are derived from try1-opt2. Perhaps we should do one more from alignments, just to make sure we haven't missed anything. Otherwise, I think we're pretty much done, as the differences between models is very small. Tue Aug 1 10:23:57 PDT 2006 Kevin Karplus try6 started on whidbey with copy of try5 costfcn from alignments. Tue Aug 1 10:59:49 PDT 2006 Kevin Karplus try6 does not seem to be getting great scores---probably the break and clash penalties are too high for optimizing from alignments. Maybe I should try again with a costfcn more appropriate for working with alignments. Tue Aug 1 11:05:50 PDT 2006 Kevin Karplus try7 started on whidbey to try running from alignments with a more appropriate costfcn. Tue Aug 1 11:48:48 PDT 2006 Kevin Karplus Sure enough, try6 did very poorly with both the try6 costfcn and in rosetta's assessment. try7-opt1 already is beating try6-opt2, and it looks like try7 will be almost as good as try1, but not competitive with more polished versions. Tue Aug 1 12:25:51 PDT 2006 Kevin Karplus The try7 costfcn likes try7 almost as much as try1, but Rosetta hates it. Interetingly, try6 provides a C-terminal helix where one is predicted, while all our other predictions have an edge strand. We should probably include this prediction (or a polished version of it) as an alternative. We might want to stick on the N-terminal helix from try4 though, since try6 seems to have messed up there. chimera-4-6 will have M1-I19 from try4-opt2 and V20-Y145 from try6-opt2. Tue Aug 1 12:43:29 PDT 2006 Kevin Karplus try8 started on whidbey to polish chimera-4-6 Tue Aug 1 14:10:03 PDT 2006 Kevin Karplus try8 scores better than try6, but worse than try4, try5, try3, try2, try1, try7. try8 in general looks worse than the rest, with bad breaks and awkward loops. I don't think it is worth doing more polishing though. oops--no wonder it is bad, I should have taken almost everything from try4 and only the c-terminal helix from try6. Let's try again with a chimera-4-8 taking M1-E124 from try4-opt2 and W125-Y145 from try8-opt2. Tue Aug 1 14:24:21 PDT 2006 Kevin Karplus try9 started on cheep to optimize chimera-4-8 (which already outscores try8-opt2) on the try9 costfcn. try9-opt2 comes out scoring well, so I'll move it up to third position. I now plan to submit ReadConformPDB T0380.try4-opt2.pdb ReadConformPDB T0380.try5-opt2.gromacs0.repack-nonPC.pdb ReadConformPDB T0380.try9-opt2.pdb ReadConformPDB T0380.try3-opt2.pdb InFilePrefix ReadConformPDB T0380.undertaker-align.pdb model 1 # t2k 2fhqA Tue Aug 1 15:49:39 PDT 2006 Kevin Karplus Submitted with comment Target T0380 was a fairly easy comparative modeling target. BLAST found an excellent hit to 2fhqA, and our HMMs found several more templates. The only difficulty was in deciding whether to believe the secondary-structure prediction for the C-terminus, which the neural nets believed should be helical, but which all the top alignments had as a strand. For the most part, we believe the alignments more than the neural net predictions, but we did include one model with a C-terminal helix. Model 1 is try4-opt2, the top-scoring model with several of the undertaker cost functions. It was optimized by undertaker from try3-opt2, in turn from try2-opt2, in turn from try1-opt2, which was automatically generated by undertaker from alignments and fragments. Model 2 is try5-opt2.gromacs0.repack-nonPC is the model that rosetta likes best of all the ones we had it repack sidechains for. It was created from try5-opt2 by reoptimizing with gromacs (to remove small clashes) and repacking the sidechains with rosetta. try5-opt2 was optimized by undertaker from try1-opt2.gromacs0.repack-nonPC. Model 3 is try9-opt2, optimized by undertaker from a chimera of try4-opt2 and try8-opt2 (M1-E124 from try4-opt2, W125-Y145 from try8-opt2). try8-opt2 was optimized by undertaker from a chimera of try4-opt2 and try6-opt2, with M1-I19 from try4-opt2 and V20-Y145 from try6-opt2. try6-opt2 was optimized by undertaker from alignments, but using a cost function that had too high a weight on clashes and breaks for optimization from alignments, so the try6 and try8 models were not very good. They were only kept because they had the C-terminal helix. Pating the C-terminal helix onto try4 resulting in a plausible model, though not as likely as the C-terminal strand models. Model 4 is try3-opt2, an earlier stage in the optimization process. Model 5 is sidechain replacement by SCWRL on an alignment to 2fhqA. ---------------------------------------------------------------------- Fri Aug 11 03:49:51 PDT 2006 Kevin Karplus T0380 was chosen as a refinement model. tr380 does indeed have the C-terminal strand, and looks pretty close to our models. It scores about as well as try1-opt2, and repacking it does not seem to change it, leading me to suspect that this is a rosetta model. For some reason, gromacs crashes when trying to optimize the model. The problem seems to be in pdb2gmx: Fatal error: Atom HA in residue MET 1 not found in rtp entry with 9 atoms while sorting atoms. Maybe different protonation state. Remove this hydrogen or choose a different protonation state. Option -ignh will ignore all hydrogens in the input. It looks like gromacs doesn't understand hydrogens very well! After adding the -ignh to the pdb2gmx call in run-gromacs, things run ok, but gromacs makes tr380 score worse. Should we optimize tr380 in a dimer context? We don't seem to have pursued this after an initial setup by Navya, but I don't see any notes in the README file why not. Ah--it seems that Navya never ran a dimer prediction and left the directory not group-writable, so no one could. More importantly, our top hit 2fhqA is a monomer, though 1dnl is a dimer in PQS. Perhaps it is worth pursuing this time. Fri Aug 11 06:55:46 PDT 2006 Kevin Karplus I moved the unwritable dimer directory to navya-dimer, and created a new dimer directory, with the necessary initial files. I have not yet made a dimer. Wed Aug 16 15:52:06 PDT 2006 Cynthia Hsu For the purposes of polishing, I've copied the try1.costfcn to the try10.costfcn, then commented out the dssp-ehl2 constraints, the undertaker-align.sheets, and the rr constraints. "bad_peptide" was raised to 15, "dry5" to 20, "dry6.5" to 25, and "dry8" to 20. Other than that, there were limited improvements to be made in the scores. try10.under was a polishing run of tr380.pdb.gz try10 is currently running on orcas. Wed Aug 16 17:18:47 PDT 2006 Cynthia Hsu try10 seems to have a decent packing. In fact, it now scores much better than all of our previous models. Interestingly enough, it seems to benefit in terms of soft_clashes, and to a lesser degree, phobic_fit, from repacking. Irregardless, I've decided to still try and optimize try10-opt2.pdb, as it had a much better score overall in things that generally are not modified (such as "wet6.5, "hbond_geom_beta_pair"). In addition, there was no change to soft_clashes in the try10 cost function, so I think that raising these values would be useful. I raised "n_ca_c" to 7, just to keep it from being raised as the cost of "bad_peptide" was raised to 20. Because there still seems to be slight foaminess in the model, I raised "dry6.5" to 30 and "dry8" to 23. Looking at the refinement model, I found, despite the presence of three Cys residues, no disulfide bonds, so I removed "maybe_ssbond". I considered removing the "maybe_metal" cost, as the His did not appear to be interacting with each other or the Cys residues, but I found that H90 and C120 were in close enough proximity that if H90 were oriented the other way, it could potentially form a metal bond interaction. As a result, I left in the weight. I also raised "soft_clashes" to 40 and "break" to 100. try11.under polished try10-opt2.pdb. try11 is currently running on camano. Wed Aug 16 17:51:15 PDT 2006 Cynthia Hsu Just received the email from Prof. Karplus regarding that the answer for T0380 had already been released. I've killed the gmake/undertaker run on camano for try11. Sun Aug 20 20:12:51 PDT 2006 Kevin Karplus try10 *was* an improvement over tr380, though not by much. Our best submitted model was model1 (try4-opt2) though try1-opt1 was slightly better. Tue Aug 22 11:30:48 PDT 2006 Cynthia Hsu Looking at T0380.model1-real.pdb in RasMol, our model 1 resembled the actual result fairly closely, except for the fact that we tried to pack the end loops of the beta sheets a little bit too close to the rest of the barrel. In the actual model, there was some distance, which in retrospect makes sense because of the exposed residues as labeled in the near script. In examining the cost function for Model 1 (try4.costfcn), the only major difference I can see is the lowered hbond_geom_beta constraints, the raised dry weights on dry6.5 and dry8, and the very very high soft_clashes and break score. It seems that there was a single very very large break before A78, and another slightly smaller one before K45, that persisted throughout all the models. was very persistent through most of the models. Looking at try1, the phase appeared to be slightly off from Y30 to D42, although this is hard to tell. In addition, Rosetta5, which scored extremely well, did not rank as well with our unconstrained cost function. This appears to be a general trend across all of our models. Fri Sep 1 14:42:06 PDT 2006 Kevin Karplus Our best model (other than refinement models) was try1-opt1, which is fairly similar to our best submitted (model1=try4-opt2). We do seem to have picked the best of our alignments (align1) to work on. Sun Sep 10 08:06:41 PDT 2006 Kevin Karplus Our best model was try10-opt1, which was a refinement model and which did improve on try380 and tr380.gromacs0. Sat Jan 20 07:06:00 PST 2007 Kevin Karplus For some reason this refinement model does not seem to have been submitted---ah, early release.