Tue Jul 8 13:46:43 PDT 2008 TR435 Make started Tue Jul 8 14:11:00 PDT 2008 Running on moai21.kilokluster.ucsc.edu Tue Jul 8 14:11:09 PDT 2008 Kevin Karplus The CASP organizers say REFINEMENT TARGET TR435 (one of the best submitted models). Starting MODEL GDT_TS=75. 14 N-term residues cut out as not participating in the crystal contacts. Residues 62-72 absent in the experimental structure. This was a server-only model. Tue Jul 8 15:06:21 PDT 2008 Kevin Karplus OK! The top model is ours: SAM-T08-server_TS1. This means that I can proceed with confidence that what I know to do is reasonable. Actually, ours is the closest server model to the provided TR435 model, but it has a GDT of only 97%, and about a 1 Angstrom RMSD_CA, so this must be a human-produced model, not a server model. Tue Jul 8 18:59:11 PDT 2008 Kevin Karplus try1-opt3 fits the align_constraints better than TR435.pdb, and seems to have more beta sheet. I may need to extract sheets (and possibly other constraints) from the TR435.pdb file to keep the good parts. It looks like TR435 needs break and clash reduction, mainly. There may be some CYS clustering to examine. Tue Jul 8 19:13:05 PDT 2008 Kevin Karplus C47-C122 seem to be forming a non-conserved disulfide. I should add that as a constraint. The main difference bewteen try1 and TR435 seems to be the position of the C-terminal helix. Tue Jul 8 22:51:01 PDT 2008 Kevin Karplus try2 done, but more gap closing/break reduction needed. Tue Jul 8 23:03:15 PDT 2008 Kevin Karplus I'm starting try3 from all models, but I should probably also do an optimization from just the try1 models, to provide an alternative model with the different position of the C-terminal helix. Tue Jul 8 23:06:25 PDT 2008 Kevin Karplus I started try4 to optimize try1. It may not work well, as breaks and clashes may have taken too big a jump from the try1 costfcn to the try3 costfcn. Wed Jul 9 08:10:54 PDT 2008 Kevin Karplus try4-opt3 actually ends up with fewer clashes than try3-opt3 and more Hbonds, but it doesn't fit the initial sheets and helices quite as well, nor does it make the disulfide quite as perfectly. Some of the biggest differences between try3 and try4 are in the experimentally unresolved loop for 62-72, which may have bearing on the placement of the final helix. The end of the hairpin from R26 to G35 is also slightly different. Rosetta likes best try3-opt3.gromacs0.repack-nonPC For try5, I'll use all models and up the clash penalties, and optimize from all models. For try6, I'll use the same costfcn, but exclude try2 and try3 models, so as to polish the try4 model. ------------------------------------------------------------ From casp@predictioncenter.org Wed Jul 9 10:14:46 2008 Date: Wed, 09 Jul 2008 10:14:44 -0700 From: CASP Organization: Protein Structure Prediction Center To: Kevin Karplus Subject: Re: query about refinement model TR435 In-Reply-To: <200807082226.m68MQfWB003766@cheep.cse.ucsc.edu> Yes, TR435 is a human group model. Your SAM-T08-server_TS1 model is also very good (73.8 GDT_TS). Andriy Kevin Karplus wrote: > I tried looking through the server models for TR429, TR432, and TR435. > > For TR429, I found no server models close to the refinement model, at > least not for the full-length. I gather that this was from one of the > human-aided groups. > > For TR432, I easily found that the model was MUProt_TS5. > > For TR435, I did not find any server model that matched, though > SAM-T08-server_TS1 seemed to be the closest. This was a server-only > target, so I had expected to find the model in the server tarball. > Was it a human-submitted model? (I realize that human-submitted models > were allowed for the server-only targets, I just didn't expect to see > one.) > > I'm quite pleased at the SAM-T08-server doing well on one of the > "easy" targets, because we have historically paid no attention to > those and done rather poorly on them. I put in some modifications > this year to avoid throwing away good properties of models, and it > seems to have done some good. > > > -- Andriy Kryshtafovych for CASP organizers ------------------------------------------------------------ Wed Jul 9 14:30:36 PDT 2008 Kevin Karplus try5 and try6 look pretty good, with the main difference being the C-terminal helix. I think I'll submit: ReadConformPDB TR435.try5-opt3.pdb # < try3-opt3.gromacs0 < try2-opt3 < TR435 ReadConformPDB TR435.try5-opt3.gromacs0.repack-nonPC.pdb # < try3-opt3.gromacs0 < try2-opt3 < TR435 ReadConformPDB TR435.try6-opt3.pdb # < try4-opt3.gromacs0 < try1-opt3 < align(2qpwA) ReadConformPDB TR435.try3-opt3.pdb # < try2-opt3 < TR435 ReadConformPDB TR435.try4-opt3.pdb # < try1-opt3 < align(2qpwA) Wed Jul 9 14:57:14 PDT 2008 Kevin Karplus DAMN! I goofed up the A2M file yesterday---the chain is supposed to start with SRARL..., not ARLSL... So almost all the work on this target was wasted!!! I'll move everything to a subdirectory (GOOF/) and try again. Make started Wed Jul 9 15:01:52 PDT 2008 Running on moai26.kilokluster.ucsc.edu Wed Jul 9 19:57:12 PDT 2008 Kevin Karplus OK, I've redone the predictions with the correct sequence. Now I'll try to patch 2 N-terminal residues onto my former models and polish them. Actually, I'll patch in S15-L21 from the initial TR435, and optimize with the try7 costfcn (copied from the GOOF try6 costfcn). OOPS, that won't work, because the residue numbering was all wrong in the GOOF costfcns. Instead, I'll extract sheets and helices from the try5 and try6 models (as interpreted in this target, not the GOOF numbering, and use those as constraints in try7. try7 will optimize the from-GOOF-try5 model, and try8 will optimize the from-GOOF-try6 model, both using the same costfcn. Mon Nov 10 10:50:11 PST 2008 Kevin Karplus Models 3 and 1 improved GDT and real_cost, though not by much.