Sat Aug 2 13:35:17 PDT 2008 TR488 Sat Aug 2 13:47:23 PDT 2008 Kevin Karplus CASP organizers say REFINEMENT TARGET TR488 (the best submitted server model according to GDT_TS). Quite accurate model: GDT_TS=88; RMSD=1.43A. We suggest to pay the main attention to refinement of the loop around res.14. Make started Sat Aug 2 13:48:05 PDT 2008 Running on moai08.kilokluster.ucsc.edu Sat Aug 2 14:06:07 PDT 2008 Kevin Karplus The template is probably 2qg1A, 2he2A, or 1iu0A. Sat Aug 2 19:40:10 PDT 2008 Kevin Karplus The server model provided is MULTICOM-CLUSTER_TS5. SAM-T08-server_TS2 is fairly close (GDT 91.8%, but RAPTOR_TS1, MULTICOM-CLUSTR_TS1, MULTICM-REFINE_TS1,4,2, and Pcons_multi_TS2 are closer). Mon Aug 4 10:35:47 PDT 2008 Kevin Karplus TR488 has bad clashes, so optimization in try2 starts from just the TR488 model (not including try1 models). Mon Aug 4 13:47:20 PDT 2008 Kevin Karplus try2-opt3 looks pretty good--there are still some breaks and clashes to polish out. Based on the comments by the CASP organizers, I'll try replacing K11-L18 by the loop from try1-opt3, and optimizing that for try3. When it is done, I should probably do a polishing run from all models. Fri Aug 15 15:07:36 PDT 2008 Kevin Karplus try3-opt3 scores better than try2-opt3, but still has clashes. Rosetta still likes try2-opt3.gromacs0.repack-nonPC much better. Fri Aug 15 15:23:08 PDT 2008 Kevin Karplus I started a polishing run as try4, starting from just the gromacs-optimized models. It will probably polish try3-opt3.gromacs0 Fri Aug 15 17:09:51 PDT 2008 Kevin Karplus try4-opt3 is new best, and try4-opt3.gromacs0.repack-nonPC is new rosetta favorite. The try4-opt3 and try4-opt3.gromacs0.repack-nonPC models differ mainly in the rotamers of surface residues. I'll do another run with sidechain cost turned up, starting only from the gromacs0.repack-nonPC models. Fri Aug 15 19:14:39 PDT 2008 Kevin Karplus try5-opt3 scores much better, and try5-opt3.gromacs0.repack-nonPC is rosetta's favorite, though try5-opt3.repack-nonPC is close behind. Fri Aug 15 19:22:55 PDT 2008 Kevin Karplus I started another polishing round from all the gromcas0.repack-nonPC models, but I think I may have gotten as far as I'm going to with this refinement. Sat Aug 16 09:19:02 PDT 2008 Kevin Karplus Both undertaker and rosetta like try6 better. I'll do a similar polishing run, turning up breaks and clashes even higher as well as dry5 and dry6 to try for tighter packing. This run will again start with just the gromacs0.repack-nonPC models. Sat Aug 16 10:21:07 PDT 2008 Kevin Karplus try7 beats try5, but does not beat try6 (either in undertaker scoring with try7.costfcn or in rosetta). So I think I've converged to about the best I'm going to get with these methods. The biggest differences from the original model are in the targeted loop around residue 14 and the C-terminus. In both cases, I like my model better than the original, though the helix that I form at the C terminus is not predicted by the neural nets. I'll submit ReadConformPDB TR488.try6-opt3.pdb # < try5-opt3.gromacs0.repack-nonPC < try3-opt3.gromacs0 < chimera-try2-try1 ReadConformPDB TR488.try6-opt3.gromacs0.repack-nonPC.pdb # < try5-opt3.gromacs0.repack-nonPC < try3-opt3.gromacs0 < chimera-try2-try1 ReadConformPDB TR488.try7-opt3.pdb # < try6-opt3.gromacs0.repack-nonPC < try5-opt3.gromacs0.repack-nonPC < try3-opt3.gromacs0 < chimera-try2-try1 ReadConformPDB TR488.try4-opt3.pdb # < try3-opt3.gromacs0 < chimera-try2-try1 ReadConformPDB TR488.try2-opt3.gromacs0.repack-nonPC.pdb # < TR488 Sat Aug 16 10:55:09 PDT 2008 Kevin Karplus Submitted with comment The provided model has bad clashes, so I did an optimization run with undertaker to clean it up. I then copied residues K11-L18 from my initial model into the cleaned up model and tried optimizing that. In addition to the targeted loop, undertaker changed the C-terminus, trying to pack I93 and L95 tighter, adding a helix which was not predicted by the neural nets nor part of the provided model. Model 1 TR488.try6-opt3.pdb # < try5-opt3.gromacs0.repack-nonPC < try3-opt3.gromacs0 < chimera-try2-try1 This was the best scoring with various undertaker cost functions, with no detected clashes and only tiny chain breaks. 2 TR488.try6-opt3.gromacs0.repack-nonPC.pdb # < try5-opt3.gromacs0.repack-nonPC < try3-opt3.gromacs0 < chimera-try2-try1 This was the best scoring with the rosetta cost function. It is derived from the first model by doing a quick energy minimization with gromacs followed by repacking sidechains (except PRO and CYS) with rosetta. 3 TR488.try7-opt3.pdb # < try6-opt3.gromacs0.repack-nonPC < try5-opt3.gromacs0.repack-nonPC < try3-opt3.gromacs0 < chimera-try2-try1 This is a further optimization of model 2, but did not score as well as model 1. The failure to improve probably indicates approximate convergence of the optimization strategy being used---the local minimum for model 1 is difficult to beat. 4 TR488.try4-opt3.pdb # < try3-opt3.gromacs0 < chimera-try2-try1 This is an earlier optimization that was not selected in subsequent optimization runs. It has a slightly different C-terminus. 5 TR488.try2-opt3.gromacs0.repack-nonPC.pdb # < TR488 This was optimized directly from the provided model, without replacing the targeted loop. It has already changed the C-terminal conformation. Mon Nov 10 11:38:33 PST 2008 Kevin Karplus I made no improvement in GDT, but model5 made a slight improvement in real_cost.