Sat Aug 2 13:35:15 PDT 2008 TR476 Sat Aug 2 13:42:21 PDT 2008 Kevin Karplus CASP organizers say REFINEMENT TARGET TR476 (the second best submitted server model according to GDT_TS). GDT_TS=47; RMSD=2.46A. Starting model contains alignment errors. NMR structure and starting model trimmed to res. 2-88. Actually, the starting model was NOT trimmed. Sat Aug 2 14:13:00 PDT 2008 Kevin Karplus I'll rename the original model TR476-untrimmed, and start a make on the trimmed region. Make started Sat Aug 2 14:16:01 PDT 2008 Running on moai26.kilokluster.ucsc.edu Sat Aug 2 19:03:55 PDT 2008 Kevin Karplus The model provided is MUFOLD-MD_TS2, which is not one our MQA functions identified as a likely best. The closest other model is try10-opt3.gromacs0.repack-nonPC (my model2, GDT 48.9%) Since the overall GDT is 47%, is may be that the part of this that agrees with MUFOLD-MD_TS2 is reasonable. Note: they report alignment errors, so I can't put too much confidence in the provided model. Mon Aug 4 12:02:57 PDT 2008 Kevin Karplus The try2 model is based on try1, not on the provided model (which had horrible clashes), so I'll have to run try3 using just the provided model as a starting point. Mon Aug 4 13:56:36 PDT 2008 Kevin Karplus try3 failed initially, because the provided model only has CB atoms, not the rest of the sidechains! I'll run it through SCWRL for try3 and work from that model. Wed Aug 13 10:07:08 PDT 2008 Kevin Karplus try3-opt3 is very close to the initial TR476 model. I think that C4, C7, C47, and C50 should cluster better (like the zinc-binding site I was trying to get in T0476). The starting model is only the second-best server model by GDT, so which one is better? There are no others really close: name GDT to provided model MUFOLD-MD_TS2 -100.0000 PSI_TS4 -45.6897 BAKER-ROBETTA_TS5 -45.4023 MUFOLD-MD_TS4 -45.4023 PSI_TS5 -45.4023 Zhang-Server_TS3 -45.1149 PSI_TS2 -44.5402 PSI_TS3 -43.3908 PSI_TS1 -43.1034 fais-server_TS2 -41.6667 BAKER-ROBETTA_TS1 -40.8046 Zhang-Server_TS1 -39.9425 pro-sp3-TASSER_TS5 -39.3678 BAKER-ROBETTA_TS4 -38.5057 fais-server_TS5 -38.2184 RAPTOR_TS5 -37.9310 Wed Aug 13 10:42:38 PDT 2008 Kevin Karplus I made a small change to undertaker to be able to extract constraints from the current conformation pool (not just the AlignmentLibrary). I wrote a script (constraints-from-conform.under) to extract constraints from the current models. I had it do a short optimization with the try3 costfcn, so that the weights of the models would be set based on the try3 costs. I included mainly the initial and try3 models, but also put in from-try10 and try2-opt3, to get some extra plausible constraints. The results are in conform.constraints conform_rejected_bonus.constraints The noncontact set was empty, probably because the protein is so small. Wed Aug 13 11:10:10 PDT 2008 Kevin Karplus I started a try4 run to optimize the try3, try2, and initial models. The from-try10 model currently scores best (because it almost has the zinc-binding site). I'll try to get the models based on the the initial model to score as well, so that I can start doing optimizations using all models. Wed Aug 13 12:47:52 PDT 2008 Kevin Karplus try4 manages to make the zinc site a little better than from-try10 did, so I'll up the weight of the zinc constraints (and make them tighter) for try5, optimizing from all models. Wed Aug 13 12:53:35 PDT 2008 Kevin Karplus I'm submitting try4-opt3 to VAST http://www.ncbi.nlm.nih.gov/Structure/VAST/vastsearch.html Your structure data has been uploaded. Request ID: 788418673849555845 Wed Aug 13 12:55:57 PDT 2008 Kevin Karplus VAST did not find any structural alignments longer than 24 residues, and that for just the final helix and hairpin, which is not particularly useful for me in trying to match the zinc-binding site. The best hits are 1g3nA, 3coiA, 3d7zA, and 3cekA, none of which were hits with the HMMs. Also scoring fairly well are 3cekA, 2phkA, 2p5aA, 2bujA, 2v70A, 2vn9A, 1yrpA (again, none matched with the HMMs). Wed Aug 13 13:10:48 PDT 2008 Kevin Karplus The try5 run scales up breaks and clashes a bit too fast---it initially works on try4-opt2, but in the final optimization switches over to try4-opt3. I probably want to do the optimization with somewhat looser clash and break weights and slower ramp up. I also need to pack the whole model tighter. Wed Aug 13 17:31:57 PDT 2008 Kevin Karplus try6-opt3 is wierd---it basically pulled P48 and V49 out of the backbone. I'll see if upping the break penalty in try7 can force it back in. I increased the break and clash penalties in try7, but not so much that try5-opt3 moved to the top. Wed Aug 13 18:42:28 PDT 2008 Kevin Karplus try7-opt3 actually makes the breaks worse, not better. Perhaps I should try again, with a cost function that would favor try5-opt2 or some other earlier incarnation. Or, I could try a run from alignments, but using the costfcn of try6. Wed Aug 13 18:46:07 PDT 2008 Kevin Karplus try8 started from alignments, using the try6 costfcn. Wed Aug 13 18:55:31 PDT 2008 Kevin Karplus try9 started from existing models, using a costfcn with a very high break cost, which favors try5-opt3. Wed Aug 13 19:59:01 PDT 2008 Kevin Karplus try8 has bad breaks---worse even than try7. try9 starts out trying to optimize try7-opt3, but switches over to optimizing try5-opt3---unfortunately, without improving the zinc site by much. Wed Aug 13 20:24:09 PDT 2008 Kevin Karplus For try10, I'll try closing the gaps in try8. This will most likely fail, but if I can get a halfway-decent zinc site, I could try cutting and pasting it into try9-opt3. Wed Aug 13 21:12:13 PDT 2008 Kevin Karplus Although try10 closes breaks somewhat, there are still bad breaks next to and in the zinc site: Conformation[15] TR476.try10-opt3.pdb.gz has 3 breaks TR476.try10-opt3.pdb.gz breaks before (TR476)C47 with cost 11.1649 TR476.try10-opt3.pdb.gz breaks before (TR476)C50 with cost 3.21439 TR476.try10-opt3.pdb.gz breaks before (TR476)R61 with cost 1.06381 Perhaps I should try a run optimizing from all server models? Thu Aug 14 08:45:27 PDT 2008 Kevin Karplus MQAX10-opt3 has small breaks, but gets the zinc site. It seems to be based mainly on HHpred5_TS1, which only has a GDT of 27% with the provided model to polish. I'll try making a chimera of try9-opt3 and MQAX10-opt3, taking A2-P10 from MQAX10-opt3. Thu Aug 14 09:38:03 PDT 2008 Kevin Karplus try11 will attempt to close the break and optimize chimera-MQAX10-try9 The biggest problem with the chimera is clashes, which may turn out to be difficult to resolve. Thu Aug 14 10:18:08 PDT 2008 Kevin Karplus try11 closed the break and got an adequate zinc site, but has a few too many clashes. Thu Aug 14 10:39:34 PDT 2008 Kevin Karplus try12 will try to squeeze try11 and remove clashes at the same time. So far, I have not produced a model that is really compact, and I have not addressed the comment by the CASP organizers "Starting model contains alignment errors." Thu Aug 14 12:11:51 PDT 2008 Kevin Karplus try12 is not noticeably more compact, but it does have fewer clashes (though not as low as try9-opt3). Thu Aug 14 12:15:54 PDT 2008 Kevin Karplus try13 started from all models except try7 and try9, to attempt to polish try12. Thu Aug 14 13:17:18 PDT 2008 Kevin Karplus try13 improves slightly on try12 (better zinc site, fewer clashes). Interestingly, the try13 run started trying to improve try6-opt3 (good zinc site, bad breaks and clashes) before switching over to improving try12-opt3. try9 and try7 score better than try13, despite their bad breaks and clashes---maybe time to tone down the zinc weight. Thu Aug 14 14:56:41 PDT 2008 Kevin Karplus The try14 run starts out trying to improve try7-opt3, then switches to try13-opt3. The packing constraints I added seem not to be helping, so I'll try polishing (without try5 and try9) using essentially the same costfcn, but without the packing constraints. Thu Aug 14 15:58:26 PDT 2008 Kevin Karplus Because of a typo, I just changed the dates of several of the files in this directory, but I don't think any serious damage was done (the problem was an extra "-": "make best-models.pdb.gz" became "make -best-models.pdb.gz"). try15-opt3 from try14-opt2 (not opt3) seems to be about as good as I'm going to get with the current approach. The model is still very foamy, nothing like a real protein. Tue Aug 19 17:18:18 PDT 2008 Kevin Karplus try16 will try to squeeze the existing models to reduce the foaminess. I turned down the clash weights, but I had to leave the break weights high to avoid picking up the models that have P48 and V49 pulled out. try17 will try to do the same thing, but starting from just hte models with small breaks and turning down the break weights. Tue Aug 19 17:24:42 PDT 2008 Kevin Karplus try17 started on peep. Tue Aug 19 22:00:21 PDT 2008 Kevin Karplus try17-opt3 is based on try13-opt1, based on try6-opt3, which still has the nasty removal of P48-V49. I need to remove a lot more from the list. I'll try again, with the same costfcn, but keeping only those models with less than 0.05 breaks. try16 optimizes try9-opt1. It closes up the breaks at the cost of some pretty bad clashes, but packs things more tightly. I think that the clashes may be too terrible to be fixable. For the zinc site, maybe I should add Hbond C7.N C4.O Hbond C50.N C47.O Tue Aug 19 23:02:51 PDT 2008 Kevin Karplus Rats, try18 also pulled out the P48, V49 pair, though I'm pretty sure there were no instances of it in the initial models---it looks like the break was create by an OptSubtree or JiggleSegment command! For try19, I'll just add the Hbonds to the zinc site and try upping the dry weights to try for tighter packing. Wed Aug 20 08:31:12 PDT 2008 Kevin Karplus try19 improves slightly on try9 I should try doing the same thing to improve try15. Wed Aug 20 09:51:58 PDT 2008 Kevin Karplus try20-opt3, optimized from try15-opt3, scores almost as well as try19-opt3 (10 points worse out of 830), and rosetta likes try20-opt3.gromacs0.repack-nonPC almost as well as try3-opt3.gromacs0.repack-nonPC. try20-opt3 and try20-opt3.gromacs0.repack-nonPC have very similar backbones, but the packing of the aromatics in the middle is quite different. Since this is the area with the most sever foaminess, I think I'll try squeezing harder, starting from just try20-opt3.gromacs0.repack-nonPC Wed Aug 20 11:07:52 PDT 2008 Kevin Karplus Although try21-opt3 is a tiny bit more compact, it still has not really done much to close up the hole in the middle. I'll do one more optimization run for it (to get rid of the small breaks and clashes it introduced), but I don't expect significant improvement. I suspect that someone with a more aggressive optimizer (or FoldIt) should be able to squeeze this down to more compact model. I have not attempted to fix whatever misalignment the CASP organizers were reporting. Basically, I think I've wasted a lot of time on this target without much to show for it---just the improvement in what I believe is a zinc-binding site. Wed Aug 20 17:22:57 PDT 2008 Kevin Karplus I'll try one other thing---trying to lengthen the first helix back to P10, so that the initial hairpin fits more comfortably over the other cys residues. I'll start from just the provided model, so that there is plenty of room for improvement. If I get anything reasonable, I'll try doing a cut-and-paste into one of the polished models. Wed Aug 20 19:33:25 PDT 2008 Kevin Karplus I put so much weight on the new helix constraints that try23 had no trouble forming the helix, though the N-terminal hairpin now faces the wrong way to make the zinc-binding site. I'll try making a try22-try23 chimera, with L11-R33 taken from try23. Wed Aug 20 19:44:23 PDT 2008 Kevin Karplus On second thought, I don't think that will work. Maybe I should try forcing the helix to start at S9 or even G8. That will wrap the N-terminus around until it faces the other two CYS residues. I try those longer helices in try24 (starting at S9) and try25 (starting at G8). Wed Aug 20 21:30:58 PDT 2008 Kevin Karplus try24 is not so bad---the the helix is formed and the zinc site is almost intact. I should do an optimization from try24 with the helix turned down to a more normal level and breaks and clashes turned up---perhaps just the try21 costfcn with a longer helix. Wed Aug 20 21:48:19 PDT 2008 Kevin Karplus I'll try polishing try25 the same way, though it does not start out quite as promising. Wed Aug 20 22:04:45 PDT 2008 Kevin Karplus I think that the misalignment being referred to is th initial helix being off by 1. if E14 is moved where E15 is currently, and E18 where E19 is currently, then the predicted burial looks a good deal more convincing. I'm not quite sure how to achieve this change, though. Wed Aug 20 22:35:34 PDT 2008 Kevin Karplus I could do a self-alignment which is off by 1, but some residue would have to end up unaligned, resulting in a missing residue from the resulting model. Probably I should have I23 missing, then add it in manually in a sort of aribtrary position, then try to reoptimize. I might need some constraints to keep the helix from rotating back. I should probably try using try22, try23, try24, try25, and try26 as templates. Thu Aug 21 00:17:59 PDT 2008 Kevin Karplus OK, I made the alignments, extracted constraints from them, then started an optimization from alignments (including the fake alignments to try22 through try27) as try28. try28-opt2 is looking pretty good, except for a bad break between E24 and N25. When try28 finishes I'll try a polishing run on it. Oh, and burying K12 while exposing L11 is also bad. Thu Aug 21 02:00:59 PDT 2008 Kevin Karplus try29-opt3 has 2 bad breaks: before C4 and before N25 gromacs thinks it can fix the one before N25. I still don't like the buried K12. I think that rotating the whole helix may have been a mistake---there may be a bulge so that only part of the helix should have been rotated. Or maybe rotation is just wrong. I'm getting too tired to think and this target is due in a few hours, so I'll just submit a few of the models I've created, without worrying too much about whether they are reasonable---none of them really are. If I'd had time to work on this model alone, and not all the others, perhaps I could have done more with it, but I'm pretty burned out after 4 months straight of CASP. Thu Aug 21 02:38:58 PDT 2008 Kevin Karplus Submitting with comment: I'm too burned out after 4 months of CASP with insufficient sleep to provide a coherent comment on this refinement model. The main thing I tried to do was to cluster the 4 cysteine residues into a zinc-binding site: C4, C7, C47, C50. Initial attempts resulted in a big void in the middle of the protein (there in the provided model also) and no change in alignment, so I also tried extending the initial helix back further and rotating it. Much of this work was done in the last day before the target was due, and so is far from being completed. The resulting models all have serious problems, but I'm not likely to make it better without some sleep. Model 1 TR476.try22-opt3.pdb #