Tue Jun 10 15:41:16 PDT 2008 TR389 This is the refinement target for T0389. I have a few notes at the end of the T0389/README file noting that our servers did poorly, but that our hand prediction was pretty good for this target. The TR389.a2m sequence is shorter than the T0389 one, so I'll redo the predictions to get them to be the right length. The try1 model will be junk, but I'll be able to seed the try optimization with the TR389 model (and possibly some sheet and helix constraints) and see if we can improve it any. Make started Tue Jun 10 15:46:31 PDT 2008 Running on peep.cse.ucsc.edu Wed Jun 11 04:49:14 PDT 2008 Kevin Karplus Looking at the components of the try1 costfcn, where does the initial, carefully-selected model score better than the automatic model, and where does it score worse? plus means that the model selected based on the experimental results scores better try1-opt3 T0389 try10 align_constraints -- =- ehl2_constraints + = align_sheets = + nn1000 - + two_stage - + pred_nb_back - + pred_alpha_back = + pred_n_notor_back - + pred_bys_back = - pred_pb_back - - wet6.5 - = dry5 - - dry6.5 - - dry8 - - dry12 =+ - phobic_fit + + sidechain - - n_ca_c - - bad_peptide + - soft_clashes - - backbone_clashes = - break + - hbond_geom - - hbond_geom_backbone - - hbond_geom_beta - - hbond_geom_beta_pair - - It looks like I'll want to work mainly with the built-in cost functions for this refinement, as the predicted ones are not doing so well. The align_constraints are particularly misleading. I'll also check the T0389_try10 sheets, though not the helices, since they are somewhat different from the TR389 model we were given. Wed Jun 11 05:22:49 PDT 2008 Kevin Karplus I've started try2 starting from just the TR389.init file that we were given to refine. It includes constraints for the sheets and helices of the initial model, but also (with smaller weight) for the sheets of the T0389_try10 model. Those sheets are completely compatible with the initial file, just beginning or ending earlier or later. Eliminating clashes from the init model will probably be the main goal of refinement. I may end up getting some help from gromacs. Wed Jun 11 07:33:34 PDT 2008 Kevin Karplus try2-opt3.gromacs0 scores the best now. The biggest difference is that try2 tucks in the hydrophoic N-terminal tail. It also moves the C-terminal residue a little, which may be incorrect, as this model isnot constrained by the rest of the chain, which may pull the residue out. I started try3 with stiffer break and clash weights. I also put in the try2 sheets and helices and increased dry weights a little to keep the packing tight. There *is* a tendency to "foam" proteins a bit when removing clashes. Wed Jun 11 08:06:38 PDT 2008 Kevin Karplus I'm not sure that the specific fragments are being used often enough now to justify the 400 seconds it takes to read them in and set them up. OK, 7 minutes is not much, but we could get a 15% speedup on the 45-minute optimization by eliminating them. (Question: would an empty SpecificFragmentLibrary cause undertaker to crash?) Wed Jun 11 09:11:22 PDT 2008 Kevin Karplus The closest structures to the try2-opt3 model according to VAST are 2gwfE, 1yt8A, 2oucA, 1gmxA. Interestingly, the combination of 2gwfE and 2oucA would cover the whole chain, since they have gaps in different places. Wed Jun 11 09:28:56 PDT 2008 Kevin Karplus For try4, I'll start from just the gromacs0 and gromacs0.repack-nonPC optimized models, and have soft_clashes and breaks set even higher. We're getting to the point where the different definitions of clashes are changing which models the different programs like best. I probably should go make better tables for undertaker clashes, when I have some time. Wed Jun 11 10:47:04 PDT 2008 Kevin Karplus F30, F99, and F111 seem to have popped out of the core of the protein, leaving voids not quite big enough to squeeze back into. This is a problem on try2, try3, and try4 models, as it comes from the original TR389 file. Even T0389.try10-opt1-scwrl has this problem, so it probably comes from the templates or is otherwise essentially unavoidable. I don't see how, right now, to tuck the PHE in. The try2 through try4 models are within about .5 Angstroms of each other, so further optimization probably won't help much. I am still seeing improvement in the Rosetta energy, though, so I should probably do another run. Wed Jun 11 13:04:25 PDT 2008 Kevin Karplus As expected try5 doesn't move things much. The core stays within about 0.03 Angstroms, and all-atom RMSD within about 0.15 from try4-opt3. The bigger differences are between Rosetta-repacked sidechains and undertaker-chosen sidechains (only the surface sidechains change). The backbones are now changing so little that rosetta gets essentially the same sidechains no matter which of the gromacs-optimized starting points we use. Sat Jun 28 14:51:39 PDT 2008 Kevin Karplus I made some little fixes to undertaker today that mainly affect chains with very small breaks, so I'll do one more polishing run to see if anything moves. I doubt that I'll ever be able to pack the PHE rings in where I'd like to see them, so this will probably be the last run. Sat Jun 28 14:56:53 PDT 2008 Kevin Karplus One more thing to do; D26, R28, Y33 are conserved from 2oucA, so I should measure distances there and add them as constraints! TR389 2oucA D26 D176 R28 R178 E32 E182 Y33 Y183 S36 S186 H37 H187 A41 A191 N43 H193 Y84 Y236 Q86 E238 Constraints: R28.NH1 Y84.OH 3.75 R28.NH1 D26.OD1 2.59 Y84.OH D26.OD1 2.59 R28.NH2 S36.O 2.86 R28.NH2 E32.O 3.11 R28.NE Q86.OE1 5.46 R28.N D26.OD2 2.71 Y33.OH H37.NE2 3.89 Y33.OH H37.ND1 4.56 Y33.OH N43.N 4.13 Y33.OH A41.O 2.85 For try7, I will use the try7 costfcn, but start over from TR389.init Sat Jun 28 16:48:21 PDT 2008 Kevin Karplus try3-opt3 outscores try6-opt3 on the try7 costfcn, but still has clashes. I'll increase the clash and break weights for try8 and run from the try7 models. Sat Jun 28 18:06:29 PDT 2008 Kevin Karplus try8-opt3 looks pretty good, except for the N-terminus. I like the way that try6-opt3 places that better. I'll make a chimera of try8-opt3.gromacs0 and try6-opt3, taking M1-T4 from try6. I'll also add a C-terminal constraint of an Hbond from K135.N to T35.O This has been started as try9. Sat Jun 28 21:11:33 PDT 2008 Kevin Karplus try9-opt3 scores well with the try9 costfcn, and try9-opt3.gromacs0.repack-nonPC.pdb is now the best for rosetta. Sat Jun 28 22:27:25 PDT 2008 Kevin Karplus submitted ReadConformPDB TR389.try10-opt3.pdb # best undertaker ReadConformPDB TR389.try9-opt3.gromacs0.repack-nonPC.pdb # best robetta ReadConformPDB TR389.try8-opt3.gromacs0.pdb ReadConformPDB TR389.try6-opt3.pdb ReadConformPDB TR389.try5-opt3.gromacs0.repack-nonPC.pdb Mon Nov 10 10:41:46 PST 2008 Kevin Karplus By GDT and real_cost, my model4 is the best I submitted. Models 1 and 2 are worse than init by GDT, though better by real_cost.