Fri Jul 18 13:38:17 PDT 2008 TR453 Fri Jul 18 13:39:03 PDT 2008 Kevin Karplus from CASP: REFINEMENT TARGET TR453 (one of the best submitted models). Quite accurate starting model: GDT_TS=89; RMSD=1.45. Concentrate on refining area 35-44. There is another area of possible improvement. Residues 1-4 are missing from the experimental structure and cut out from the refinement target. Make started Fri Jul 18 13:42:38 PDT 2008 Running on moai25.kilokluster.ucsc.edu Fri Jul 18 14:03:38 PDT 2008 Kevin Karplus The target has >45% sequence identity to 2oaiA, so the high GDT_TS is not too surprising. This was a server-only model, so I have no prior knowledge of the target. The MQAC quality assessment predicted GDT of 85% for LEE-SERVER_TS4 and over 80% for the top 148 models. The MQAU quality assessment predicted GDT of 79% for SAM-T08-server_TS1 and over 76% for the top 159 models. The models are all very close, and I'm unlikely to be able to improve things with my cost functions. Fri Jul 18 20:54:44 PDT 2008 Kevin Karplus TR453=MULICOM-CLUSTER_TS4=MULTICOM-RANK_TS1 The SAM-T08-server is down at GDT=87.6%, worse than SAM-T06-server_TS2 at 89.9%. It looks like I want to move closer to alignments, not further away! Make started Fri Jul 18 21:03:08 PDT 2008 Running on cheep.cse.ucsc.edu Sat Jul 19 05:45:45 PDT 2008 Kevin Karplus I had to re-run the make, because I had left T0453 instead of TR453 as the sequence id in the TR453.a2m file. All the alignments and the initial TR453 model are very similar. The biggest difference seems to be the A34-I44 loop, though there is some variation in the D78-R81 loop. Sat Jul 19 09:57:40 PDT 2008 Kevin Karplus try2, although based on TR453, has moved all the loops closer to the templates. The hard thing for me is figuring out why TR453 was chosen as the best---everything is quite close, and I'm not sure which of the small differences made TR453 better than other server models. try2-opt3 fixes the clashes and the bad N-CA-C bond angles, and is better on almost every costfcn (not align_constraints and hbond_geom). For try3, I'll start from just gromacs-optimized models, and turn up align_constraints, soft_clashes, and hbond_geom. Sat Jul 19 14:09:45 PDT 2008 Kevin Karplus Rosetta still prefers try2-opt3.gromacs0.repack-nonPC Sat Jul 19 14:14:47 PDT 2008 Kevin Karplus Perhaps the thing to try next on this refinement target is to replace loops of try2-opt3.gromacs0.repack-nonPC with loops of try3-opt3 one at a time, and see if we can optimize a chimera that is better than both parents. Sun Jul 20 09:07:24 PDT 2008 Kevin Karplus chimera-D76-D83 will be try2-opt3.gromacs0.repack-nonPC for most of the model, but try3-opt3 for D76-D83. chimera-V8-W15 will be mostly try2-opt3.gromacs0.repack-nonPC, but try3-opt3 for V8-W15 chimera-L32-I44 will be mostly try2-opt3.gromacs0.repack-nonPC, but try3-opt3 for L32-I44 try4 will optimize chimera-D76-D83 try5 will optimize chimera-L32-I44 try6 will optimize chimera-V8-W15 All will use the same cost function, try3.costfcn. Sun Jul 20 10:49:56 PDT 2008 Kevin Karplus The try4-5-6 costfcn prefers try5-opt3, try3-opt3, try4-opt3, try6-opt3, try2-opt3.gromacs0. Rosetta likes best try6-opt3.gromacs0.repack-nonPC Sun Jul 20 13:11:29 PDT 2008 Kevin Karplus For try7, I'll do a polishing run starting from all the gromacs optimized models. I think that will probably favor try5, in which case I'll do another run starting just from the models that don't include that variant of the L32-I44 region. Sun Jul 20 13:50:12 PDT 2008 Kevin Karplus try7-opt3 does indeed polish try5-opt3.gromacs0. Things improve, but rosetta still likes try6-opt3.gromacs0.repack-nonPC better. For try8, I'll try polishing the try6 models. As in try7, I'll use only the gromacs-optimized models, but exclude try3,5,7. ReadConformPDB TR453.pdb.gz ReadConformPDB TR453.try2-opt3.gromacs0.pdb.gz ReadConformPDB TR453.try2-opt3.gromacs0.repack-nonPC.pdb.gz ReadConformPDB TR453.try4-opt3.gromacs0.pdb.gz ReadConformPDB TR453.try4-opt3.gromacs0.repack-nonPC.pdb.gz ReadConformPDB TR453.try6-opt3.gromacs0.pdb.gz ReadConformPDB TR453.try6-opt3.gromacs0.repack-nonPC.pdb.gz Sun Jul 20 15:34:11 PDT 2008 Kevin Karplus try8-opt3, optimized from try6-opt3.gromacs0, scores almost as well as try7-opt3, but rosetta doesn't care much for try8-opt3.gromacs0.repack-nonPC, prefering try6, try7, and try2. Sun Jul 20 15:46:12 PDT 2008 Kevin Karplus I think I've reached diminishing returns here. I'll submit 5 models: ReadConformPDB TR453.try6-opt3.gromacs0.repack-nonPC.pdb # < chimera-V8-W15 # best Rosetta energy, close to provided model # loop V8-W14 taken from try3-opt3 ReadConformPDB TR453.try7-opt3.pdb # < try5-opt3.gromacs0 < chimera-L32_I44 # very different in L32-I33 loop (