Tue May 16 08:56:39 PDT 2006 T0288 Make started Tue May 16 08:57:10 PDT 2006 Running on cheep.cse.ucsc.edu Make started Tue May 16 09:07:12 PDT 2006 Running on cheep.cse.ucsc.edu Tue May 16 09:26:36 PDT 2006 Kevin Karplus Had to restart, because the a2m file downloaded from the web had a formatting error (an extra space after the ">") This seems to be a comparative model, as the t06 alignment has 99 PDB files. Seems to be a PDZ domain. Tue May 16 09:43:08 PDT 2006 Kevin Karplus There are at least 104 templates available in PDB for this protein. Even BLAST finds 79 hits in PDB. The top hits with blast (probably the closest templates) are 2fne[AC], 1xz9A, 1t2mA, 1wfvA, 1x6dA, 2fcfA, 2cssA, 1ueqA, 1kwa[AB]. 2fneA scores best with the t2k.w0.5, t04.w0.5, and t06.w0.5 HMMs also, so those models have probably not drifted too much. The second highest score is for 1um7A (t2k, t06) or 1rgrA (t04). The 1xz9A sequence that blast scores second best is way down the list on the HMM scoring (though still better than 1.5e-08). At least 2fneA is a crystal structure (lots of the PDZ domains are just NMR structures, which are not as reliable for modeling). The t06 2-track HMMs all prefer 1g90A (probably because the t06 library is far from complete, and hasn't gotten up to 2fneA). The t04 and t2k 2-track HMMs mostly prefer 2fe5A, though 2fcfA sometimes edges it out. 2fneA is not in the template library, though we may want to add it to make better alignments). Tue May 16 11:06:53 PDT 2006 Kevin Karplus There are 23 t04 templates scoring better than E-value 0.01, with 1fc6A, 1v5lA, 1wf7A, 1mfgA, 1g90A all with less than 1e-09. Make started Thu May 18 15:14:58 PDT 2006 Running on lopez.cse.ucsc.edu The first run died in the power failure, and I just got around to restarting it. Thu May 18 17:59:38 PDT 2006 Kevin Karplus The T0288.pdb.blast run I did today reports the following good templates: # Fields: Query id, Subject id, % identity, alignment length, mismatches, gap openings, q. start, q. end, s. start, s. end, e-value, bit score T0288 2fneA 35.80 81 46 3 7 83 29 107 3.3e-06 44.67 T0288 1xz9A 34.52 84 50 2 7 87 15 96 5.6e-06 43.90 T0288 1t2mA 34.52 84 50 2 7 87 9 90 5.6e-06 43.90 T0288 1wfvA 31.88 69 46 1 7 75 15 82 5.6e-06 43.90 T0288 1x6dA 29.58 71 49 1 9 78 21 91 1.6e-05 42.36 T0288 2fcfA 26.67 90 57 1 1 81 2 91 8.1e-05 40.05 T0288 2cssA 32.05 78 50 1 6 80 28 105 2.3e-04 38.51 T0288 1ueqA 34.29 70 45 1 8 77 24 92 3.1e-04 38.12 T0288 1kwaA 27.27 77 52 2 7 82 5 78 4.0e-04 37.74 T0288 1z87A 28.21 78 56 0 6 83 80 157 6.8e-04 36.96 T0288 1z86A 28.21 78 56 0 6 83 3 80 6.8e-04 36.96 T0288 1zubA 30.77 78 51 1 6 80 28 105 6.8e-04 36.96 T0288 2pdzA 28.21 78 56 0 6 83 3 80 6.8e-04 36.96 T0288 1qavA 28.21 78 56 0 6 83 7 84 6.8e-04 36.96 T0288 1x45A 25.00 84 61 1 5 86 7 90 8.9e-04 36.58 This is a bit confusing, as as 2fneA is coming out at E-value 0.5 in best-scores.rdb and 1xz9A at 17.5. Have the HMMs drifted really badly or is BLAST confused? Many of the top BLAST hits don't correspond to decent HMM hits. The try1 run seems to be picking up the 1qavA alignment as its favorite (rank 14 on the BLAST list, rank 11 in the HMM lists). Maybe I should add some of the BLAST hits as MANUAL_TOP_HITS (along with the HMM favorites) and see what comes out on try2. Thu May 18 23:42:34 PDT 2006 Kevin Karplus For the try2 run, I removed the rr constraints and modified the helix constraints to come from try1-opt2. I increased the weights of sidechain, soft_clashes, and breaks considerably. I also modified the initial TryAllAlign to use only the 10 MANUAL_TOP_HITS, a much smaller set of templates than was used for try1-opt2. MANUAL_TOP_HITS:= 2fneA 1xz9A 1t2mA 1wfvA 1x6dA 2fcfA 1g9oA 1ihjA 1q3oA 1mfgA Note that the top hits don't include the 1qavA alignment that undertaker liked for try1-opt2, but I'm hoping that there is enough diversity in the minor flexing of the loops to give us good sampling of the space around this fold. Fri May 19 08:02:04 PDT 2006 Kevin Karplus try2-opt2 scores a bit better than try1-opt2, though it looks to me like it has been pried open to expose the inside. I think the sidechain weight is a bit too large and perhaps the hbond_geom_beta_pair. A little more weight on near_backbone and dry5 might help, perhaps dropping soft_clashes back a little. Fri May 19 09:13:08 PDT 2006 Kevin Karplus try3-opt2 still looks a bit pried open. I'll try increasing phobic_fit and the pred alpha weights. Fri May 19 11:04:07 PDT 2006 Kevin Karplus Both try1 and try4 look very similar to the versions copied from the templates. For try5 I'll just do a polishing run without constraints but with somewhat increased soft_clashes and break. Sat May 20 13:00:52 PDT 2006 Kevin Karplus The differences between the models are so small that I'll just submit try5-opt2 try1-opt2 try2-opt2.repack-nonPC (rosetta's favorite) undertaker-align 1 undertaker-align 2 (I scored the robetta models, and undertaker likes ours better.) Sun May 21 06:58:58 PDT 2006 Kevin Karplus The rr predictions were redone with a newer neural net and now look fairly good (less confusing than before). I also compared our preidctions with the robetta model1 prediction. Our loops are a bit different but I can't really see one as being better than the other. The robetta model does get a bit more sheet at the N-terminal strand---sheet that is, in fact present in try2-opt2, but not in try5-opt1 and try1-opt2. This is mainly a question of whether K87.N hbonds to S1.O (as we have) or V3.O (as robetta-model1 has). How much of a strand breaker is P4? We also disagree on whether there is a tiny C-terminal helix, which we acquired in undertaker (the top 2 alignments don't seem to have it). Mon May 22 11:48:34 PDT 2006 Kevin Karplus I created model4.method and model5.method files by stripping down the generic-method-file from the starter directory (removing the undertaker stuff) and adding comments about what alignment the model came from. I then did "make casp_models" to create the model*.ts files, and edited the PARENT records of the models from alignments (model4.ts and model5.ts). I submitted our 5 models to CASP, to make sure the submission process is working properly. I'm not convinced that we have good loops and terminii for the target, and if anyone wants to work on them, there is still time. Date: Mon, 22 May 2006 12:41:00 -0700 From: Kevin Karplus To: casp7team Subject: first target submitted I submitted 5 models for T0288 today, to make sure that the submission process was working smoothly. It is still possible to resubmit for this target, and we might want to do so. I'm not convinced that we have good loops and terminii for T0288, and if anyone wants to work on them, there is still time. This is an easy fold-recognition/homology modeling target, with 30% identity to the closest template and lots of templates. The core of the model looks fine, and the tiny loops and terminii are hard to do much with. I think we may have a little too much emphasis on H-bonds, particularly in the loop G24-P29, but also the C-terminal Q89-V93, which we have wound into a helix for no good reason. There may also be a bogus H-bond from S1. Thu Jun 15 14:28:27 PDT 2006 Kevin Karplus I picked up the tr288 refinement target (decoys/tr288.pdb.gz, fetched with the fetch_refinement make target). It does not score well with try5.costfcn, even though thtat cost fcn has no penalty for missing_atoms. Doing refinement on an incomplete chain is not currently possible with undertaker, so I'll have to make a complete conformation of it first. superimpose-refine.under superimposes it as model 1 with the models we submitted in refine-models.pdb Thu Jun 15 14:36:15 PDT 2006 Kevin Karplus I can make a chimera by copying 1-5 and 87-end from one of the models it is superposed with. Thu Jun 15 14:44:33 PDT 2006 Kevin Karplus Made as decoys/refine-chimera1.pdb.gz, taking 1-5 from T0288.try2-opt2.repack-nonPC.pdb and 87-end from the first undertaker-align model, as these were the most "open" conformations we had for the ends. This chimera has bad breaks (duh!), so needs to be optimized before it can be reasonably compared with other models. try6 started on cheep to optimize refine-chimera1.pdb Thu Jun 15 17:15:58 PDT 2006 Kevin Karplus try6-opt2 does improve the score, but doesn't quite get it to the level even of try1-opt2. We may want to add SheetConstraints to extend the N-terminal/C-terminal strand pair, and to hang onto the hairpin around Y27. Thu Jun 15 18:19:03 PDT 2006 Kevin Karplus In try7, the "refinement" of the chimera was improved mainly by alignment insertion, which is probably not the intent of the exercise, but I might submit it anyway. I'll also try a similar run, with the same score function, but no alignment insertion as try8. Thu Jun 15 20:44:17 PDT 2006 Kevin Karplus try8-opt2 ended up better than try7-opt2, without doing alignment insertion. In fact, try7 and try8 resulted in almost identical models, except at the termini. Rosetta doesn't like repacking them much though. Thu Jun 15 20:56:08 PDT 2006 Kevin Karplus I started a try9 run on cheep to polish up the refinement models, with no constraints and with breaks and soft-clashes turned way up. Thu Jun 15 21:34:12 PDT 2006 Kevin Karplus try9 and try8 look almost identical in rasmol, but try9 scores best of any of our models, and rosetta likes it best for repacking. I'm probably about done with "refining" this comparative model: try9-opt2 try9-opt2.repack-nonPC try7-opt2 try7-opt2.repack-nonPC try6-opt2 Thu Jun 15 22:07:53 PDT 2006 Kevin Karplus Refinement models submitted with the following "method": The supplied conformation was superimposed on our previous complete models, and the missing residues were grafted on. Residues 1-5 were taken from T0288.try2-opt2.repack-nonPC.pdb and 87-end from the first undertaker-align model (an alignment with 1g9oA). The resulting chimera was optimized with undertaker to produce try6-opt2 (Model 5). The model was further optimized by undertaker for try7-opt2 (Model 3) which also had th sidechains repacked by rosetta (Model 4). This optimization included alignment insertion, which is almost the same as starting over with a different initial model, which is not in the spirit of the refinement experiment. However, try8-opt2 came from an optimization run with the same starting points and cost function, but with alignment insertion turned off. It arrived at a very similar model to try7-opt2. try9-opt2 is the optimization of try8-opt2 with break and clash costs turned up high. It is probably about as good as we can get with the crude cost functions of undertaker. It is submitted as model 1, and the same backboen with sidechains repacked by rosetta is submitted as model 2. Sun Sep 24 10:37:20 PDT 2006 Kevin Karplus We submitted try5-opt2 try1-opt2 try2-opt2.repack-nonPC align1 align2 Model 4 (align1) is the best of these. All of our refinements are WORSE than tr288. The Frankenstein server had 3 models that (after scwrling) were better than tr288. Our best refinement model was try6-opt2.gromacs0, which was slightly better than try6-opt2 (refine model 5).