Mon May 15 09:07:41 PDT 2006 T0286 Make started Mon May 15 09:12:30 PDT 2006 Running on lopez.cse.ucsc.edu Mon May 15 10:42:07 PDT 2006 Kevin Karplus It looks like T0286 is a comparative modeling target, with # Class: Alpha and beta proteins (a/b) Mainly parallel beta sheets (beta-alpha-beta units) # Fold: Flavodoxin-like 3 layers, a/b/a; parallel beta-sheet of 5 strand, order 21345 # Superfamily: SGNH hydrolase as its top hits Mon May 15 12:01:37 PDT 2006 Kevin Karplus Top hits with t06 target, t04 target, t2k target: 1jrlA 1ivnA 1j00A 1yzfA 1wab 1es9A 1bwr 1bwq 1bwp 1fxwF 1vjgA 1k7cA 1esc 1zmbA Top hits with t06 template 1es9A 1esc 1k7cA Top hits with t04 template 1wab 1ivnA 1es9A 1yzfA 1deoA 1vjgA 1k7cA 1esc Top hits with t2k template 1es9A 1wab 1yzfA 1ivnA 1vjgA 1k7cA 1esc [Note sequence 1deoA=1k7cA] Mon May 15 14:37:24 PDT 2006 Kevin Karplus Merging the top hits with all methods gives the ordering 1es9A 1ivnA 1yzfA 1esc 1k7cA 1wab 1vjgA 1deoA 1zmbA 1j00A 1bwr 1fxwF 1bwq 1bwp (all with SCOP classifications are hits to the same superfamily). Make started Mon May 15 17:06:00 PDT 2006 Running on lopez.cse.ucsc.edu The rr constraints were not properly made because of a typo in Make.main, so I killed the try1 run, erased the try1.costfcn file, and started the make again. Mon May 15 18:32:43 PDT 2006 Kevin Karplus I looked at try1-opt1, and it seems that D59-H82 is completely messed up. I wonder why. Mon May 15 19:24:32 PDT 2006 Kevin Karplus There is a big insertion in the alignment to 1ivnA, with S47 of T0284 aligning to A40 of 1ivnA and G62 aligning to S41. We have a potential alignment for this extra bit as an additional strand in model 8 of T0286.undertaker-align.pdb from T0286-1hdoA-t04-local-str2+near-backbone-11-0.8+0.6+0.8-adpstyle5.a2m Now 1hdoA is from a different fold family (c.2.1.2, instead of c.23.10.*), but it might be worth grabbing this one sheet constraint and adding it to the constraints for rest of the model---the hydrogen bond is even in the correct phase and the alternating hydrophobicity matches up. The only strange thing is the conserved G62 in the middle of the strand. Let's try these sheet constraints for try2 (on lopez) SheetConstraint (T0286)D43 (T0286)G46 (T0286)E61 (T0286)S64 hbond (T0286)F44 SheetConstraint (T0286)K7 (T0286)D13 (T0286)D43 (T0286)R49 hbond (T0286)I8 SheetConstraint (T0286)I8 (T0286)G12 (T0286)V86 (T0286)W90 hbond (T0286)M9 SheetConstraint (T0286)D85 (T0286)L89 (T0286)V121 (T0286)V125 hbond (T0286)V87 SheetConstraint (T0286)F124 (T0286)A126 (T0286)V156 (T0286)F158 hbond (T0286)V125 For try3, I'll try shifting the G62 to line up with G46: (on shaw) SheetConstraint (T0286)D43 (T0286)G46 (T0286)D59 (T0286)G62 hbond (T0286)F44 Tue May 16 07:40:55 PDT 2006 Kevin Karplus Neither try2 nor try2 managed to form the edge strand. Perhaps I need to force that first (inserting 1hdoA alignments), then allow the rest of the structure to be inserted. Neither try2 nor try3 has done a particularly good job of packing the helix at T67-N77 against the sheet. I think that I68, I71, and I75 probably need to be near M9 and V11. In try3, L107, L110, and I114 are improperly exposed---their helix probably needs a quarter turn. Thu May 25 14:53:35 PDT 2006 Kevin Karplus I don't like my previous analysis of what to add in try2 and try3. I need to rethink what to do with S50-S64. Sun May 28 17:52:38 PDT 2006 Kevin Karplus I'm trying to score all the server results with try3.costfcn. I also need to make an unconstrained costfcn to score with. The server files crash undertaker reading back in one of the SCWRLed files. The unconstrained costfcn likes try1-opt2 best, but try2-opt2 is almost as good. Sun May 28 21:51:38 PDT 2006 Kevin Karplus I fixed undertaker enough to run the server files through the costfcn, though it still gets some misnaming problems because the MIG_FROST_AL1.pdb file has no atoms in it. Interestingly, some servers do better than us with the try3 costfcn. try3-opt2 and try2-opt2 (and their repackings) score best, but then come Pmodeller6_TS2 and ROBETTA_TS2. Sun May 28 21:59:08 PDT 2006 Kevin Karplus I also scored everything with unconstrained.costfcn Now try1-opt2, try2-opt2, SAM_T06_server_TS1, try3-opt2 come out on top, followed again by Pmodeller6_TS2 and ROBETTA_TS2. Mon May 29 14:33:22 PDT 2006 Kevin Karplus The try1-opt2 model has a helix yanked out of place (T67-W78) and is probably *not* the one we want to submit. Actually, all four of our models have problems with that helix. The Pmodeller and Robetta models seem fine there, as do the alignments. I wonder what is going wrong. The problem may be an alignment problem---our models are putting in two helices S53-H60 and W66-W78, and aligning the earlier helix to the one in the template, allowing the later (more strongly predicted to be helical) helix to stick out whereever it can. Pmodeller6 and ROBETTA both modeled the first region as a loop. Looking at the rr constraints that might be useful to constrain the helix: It looks like Constraint I71.CB I91.CB 3.0 4.5 6 might be useful. So might Hbond S64.N E17.OE1 Constraint T16.CG2 G62.CA 3.0 3.9 5 Constraint G106.CA I68.CG2 3.0 3.3 5 These constraints were taken from various of the first 5 alignments in undertaker-align, selecting RR constraints between residues that were satisfied in the alignments. Mon May 29 16:15:29 PDT 2006 Kevin Karplus Interestingly, with the try4 costfcn (using these constraints), a different server model moves to the top (just behind try3-opt2): PROTINFO-AB_TS1-scwrl. Actually, all the PROTINFO_AB models score very similarly. Mon May 29 17:19:31 PDT 2006 Kevin Karplus The try4 optimization run seems to be losing track of breaks, just like the run I did from T0295. I must have introduced a bug this weekend into undertaker, that is causing it to lose sight of breaks. Probably the bug is in find_breaks(), since that is what I mainly worked on. Mon May 29 20:09:16 PDT 2006 Kevin Karplus try4-opt improved a bit on the constraints, but has rather poor breaks, probably because of the undertaker bug that is causing the loss of information about the breaks. Tue May 30 17:02:04 PDT 2006 Kevin Karplus I think I fixed undertaker, and am running try5 to see if I can really reduce the breaks. (The cost function is unchanged from try4.) Tue May 30 20:02:46 PDT 2006 Kevin Karplus Although try5 has done something reasonable with the helix that was formerly misaligned, and has satisfied most of the rr constraints, it has done some damage in the process, rotating helix N103-V117 the wrong way out, so its hydrophobic face is exposed, and strand L40-G46 is detached from the sheet. Thu Jun 1 18:20:35 PDT 2006 Kevin Karplus I tried playing around with a definition for try6, based on what I saw in some of the top-scoring server models. I'm in agreement with several server models about most of the beta sheets. THe big question is what happens around D43-G62. I still favor adding an extra edge strand to the sheet: SheetConstraint D43 G46 D59 G62 hbond F44 5. which try5 did not quite manage to do. I've taken out a few of the constraints that tried to get the helix in the right place, as I think they overconstrained residues G62 and S64. Thu Jun 1 20:39:20 PDT 2006 Kevin Karplus try6-opt2 scores better than previous runs on the try6 costfcn, but actually meets the constraints less well than try5. I'll have to look at it to see what it is doing, but I suspect that I need to greatly increase the constraint weight, and make sure that the new constraint has a much higher weight than the easily satisfied ones. try6-opt2 has gone back to putting the insertion after G62, which is not where I think it should go. Fri Jun 2 10:55:43 PDT 2006 Kevin Karplus try7-opt2 has a decent alignment for I42-G46, parallel to I6-P10, but it is still trying to pair G62 with G12. rather than with G46 Fri Jun 2 18:03:09 PDT 2006 Kevin Karplus try8-opt2 has pretty much convinced me that I can't get what I want with G62. I'm going to remove all constraints on the segment near G62 and reoptimize. Sat Jun 3 07:01:54 PDT 2006 Kevin Karplus try9-opt2 hardly moved from try8-opt2. I think I need to start over from alignments, as the optimization has gotten stuck in a local minimum. Sat Jun 3 13:52:54 PDT 2006 Kevin Karplus try10 scores almost as well as try9 (despite having gone through less optimization). I still don't really like S50-W66, but giving up on the extra strand has helped. I'll do a similar optimization from alignments, but upping dry terms and phobic_fit a bit for try11 (on lopez). Sat Jun 3 17:15:27 PDT 2006 Kevin Karplus The try11 run did not do as well as try8, try9, and try10, even on the try11.costfcn Sun Jun 4 07:16:46 PDT 2006 Kevin Karplus The best scores with the unconstrained costfcn are try6-opt2, try9-opt2, try11-opt2, try10-opt2, ... try6-opt2 does poorly with the constraints for the helix that I think was misaligned in the early models. Rosetta like repacking try10-opt2 best, though it really hates them all. The residues that Rosetta really doesn't like in try10-opt2.repack-nonPC are P69, T67, P143, N139, F158, P56, P52, A150, P10, V45 L102, W66, W131, ... Note that rosetta was not allowed to change the prolines, even if it didn't like them. According to undertaker the bad clashes in try10-opt2.repack.nonPC are other-bump: 1.50866 Ang (T0286)T67.O and (T0286)P69.CD threshold= 2.82818 cost= 0.976959 other-bump: 1.59209 Ang (T0286)N139.O and (T0286)P143.CD threshold= 2.82818 cost= 0.968175 other-bump: 1.7875 Ang (T0286)P52.O and (T0286)P56.CD threshold= 2.82818 cost= 0.936257 other-bump: 1.98717 Ang (T0286)T67.O and (T0286)P69.N threshold= 2.78226 cost= 0.867255 other-bump: 2.29836 Ang (T0286)A150.CB and (T0286)F158.CZ threshold= 3.13955 cost= 0.846078 other-bump: 2.10952 Ang (T0286)P10.CD and (T0286)V45.O threshold= 2.82818 cost= 0.827788 other-bump: 1.92748 Ang (T0286)W66.O and (T0286)Q70.N threshold= 2.47381 cost= 0.776263 other-bump: 2.55998 Ang (T0286)W131.NE1 and (T0286)L177.CD1 threshold= 3.23077 cost= 0.752494 other-bump: 2.6174 Ang (T0286)I187.CG2 and (T0286)W191.NE1 threshold= 3.19526 cost= 0.697875 neighbor-bump: 1.75538 Ang (T0286)T67.O and (T0286)I68.N threshold= 2.10044 cost= 0.659307 other-bump: 2.47493 Ang (T0286)K155.O and (T0286)F158.CE1 threshold= 2.95009 cost= 0.651369 other-bump: 2.08771 Ang (T0286)E163.O and (T0286)Q165.N threshold= 2.47381 cost= 0.638742 other-bump: 2.27908 Ang (T0286)N139.O and (T0286)P143.CG threshold= 2.69748 cost= 0.636251 At least Rosetta and undertaker pretty much agree where the clashes are, and even that rosetta reduced the clashes. Mon Jun 5 07:58:36 PDT 2006 Kevin Karplus I have to make preliminary predictions this morning, but this target needs more work---perhaps even generating new models from the alignments. I need to pick 5 models for preliminary predictions: Here are the best with various cost functions: unconstrained: try6, try9, try11, try10 try11: try9, try10, try8, try11, try6, try5 try10: try9, try10, try8, try7, try5 try6: try9, try8, try6, try7, try11 rosetta: try10, try2, try6, try1, try5 So let's submit try9-opt2 try10-opt2.repack-nonPC try8 try6 undertaker-align model 1 (from 1es9A) Wed Jun 21 17:27 PDT 2006 Zack Sanborn Chris and I took a long look at the best models made for T0286. It was our first time looking at this structure. The best scoring model (try9-opt2) looked overall pretty good, but had a pretty bad break in it. We did a superposition with the ROBETTA models (from the score-all.try6+servers) it looked like ROBETTA was doing the best out of most of them. In general, there was good agreement between the undertaker and ROBETTA models, which we thought was good news. However, they differed most in how they dealt with the insertion region, where we have little conservation. Kevin thinks this region will likely be disordered, so it might not really matter what we do with it. Or, rather, there's little we can do about it except for making sure it doesn't get in the way of the good things about the structure (strands and helix) So, what we are trying to do is get rid of the bad gap around residue number G62 and S64. It looks like it was trying to make a sheet (there's a Sheet constraint for it), but didn't and lined up the wrong residue. This alignment seems to have caused the break, so we've upped the right sheet constraint, and increased the gap penalties and are trying a run starting from all models to see if we get a better model out. We hope this doesn't screw up the "good" parts of the structure. try12 was started on camano at 17:30 using the following command: ( make -k T0286.do12 >& do12.log; gzip -9f do12.log ) Wed Jun 21 23:42 PDT 2006 Zack Sanborn Well, try12 seemed to do very little, but it is the current best scoring model (although, only by a bit). For try13, I've upped the constraints cost as well as the costs for breaks and gaps in the (vain) hope that Undertaker will try to do something sensible with the coil that is causing the break. try13 started on camano at 23:40. Fri Jun 23 11:44 PDT 2006 Zack Sanborn try13 had the same result as try12. I still don't know why that sheet isn't getting made. I think we need to talk to Kevin about it at some point, but in the meantime I'll make a try14 run that jacks the sheet constraint we want WAY up (to 50, all other constraints are about 5). So, we'll see how that goes. try14 started on camano at 11:49. Fri Jun 23 14:35 PDT 2006 Zack Sanborn Kevin spoke to me and told me that the last 4 try's had nearly zero chance of doing what we wanted. We kept trying to fix existing models by playing with the constraints. However, that's not going to work, since Undertaker will not break an existing structure up enough to try to satisfy the constraints. The right way to do this is starting from the best scoring alignments using the constraints we want. Since try9-opt2 was the model was the model we were trying to fix, I copied the try9.under and try9.costfcn over (since they started from a good set of alignments). Then copied in the try9-opt2 sheet constraints: SheetConstraint (T0286)T5 (T0286)K7 (T0286)S41 (T0286)D43 hbond (T0286)I6 1 SheetConstraint (T0286)K7 (T0286)G12 (T0286)D85 (T0286)W90 hbond (T0286)M9 1 SheetConstraint (T0286)G12 (T0286)D13 (T0286)G62 (T0286)H63 hbond (T0286)G12 1 SheetConstraint (T0286)D85 (T0286)L89 (T0286)V121 (T0286)V125 hbond (T0286)V87 1 SheetConstraint (T0286)F124 (T0286)D127 (T0286)V156 (T0286)V159 hbond (T0286)V125 1 The first sheet constraint above was extended to form the sheet we wanted. The third constraint above is the one that caused that funny sheet that broke our structure, so it was removed. The same helix constraints and residue constraints from try9-opt2 were used. The final constraints can be found in try15.costfcn. try15 was started on camano at 14:30. Fri Jun 23 17:30:28 PDT 2006 Kevin Karplus Why weren't helix constraints taken from try9-opt2.helices and residue-residue constraints removed? Sat Jun 24 09:42:40 PDT 2006 Kevin Karplus try15-opt2 does score best with the try15 costfcn, also with the unconstrained costfcn, but it still has the funky mid-strand insertion. Well-no wonder---it isn't starting from alignments, but just polishing up try13. I'll make a try16 that starts over from alignments, rather than trying to polish existing models. (try16 started on shaw) Sat Jun 24 11:10:28 PDT 2006 Kevin Karplus try16-opt1 seems to be doing an OK job on the sheet, but helix G62-H82 has drifted away from the sheet. Perhaps I should try again with the extra constraints that were in try15.costfcn. Or, at least some of them. try17 started on cheep. Sat Jun 24 13:18:25 PDT 2006 Kevin Karplus try16 and try17 seem to have both done ok on the sheet, but both had trouble with T67-W78. Sat Jun 24 13:40:33 PDT 2006 Kevin Karplus I think I still like try10-opt2.repack-nonPC better, though try17-opt2 gets the hbonds of the sheet a little better. Sat Jun 24 13:53:28 PDT 2006 Kevin Karplus I tried making two chimeras by taking try15-opt2 and pasting in S41-T67 from either try10-opt2.repack-nonPC (chimera-15-10) or try17-opt2 (chimera-15-17). The chimera-15-17 has very bad clashes, but may be rescuable. I should start two polishing runs one for each chimera. Sat Jun 24 14:17:10 PDT 2006 Kevin Karplus try18 started on cheep, polishing chimera-15-10 try19 started on shaw, polishing chimera-15-17 Sat Jun 24 17:07:19 PDT 2006 Kevin Karplus try18 is a new best score on try18.costfcn and unconstrained.costfcn. Sat Jun 24 17:18:24 PDT 2006 Kevin Karplus I think that I've done about all I have time for on this one. I'll submit what we have: try18 try15 try17 try16 align1 (from 1es9A) Probably none of the insertions are quite right, but guessing the structure is tough. If Zack or Chris does more work on this, they should e-mail me to update the submission. Sat Jun 24 17:33:04 PDT 2006 Kevin Karplus oops, I forgot to look at try19, which scores almost as well as try18. I don't like the breaks in try19-opt2. I think that try18-opt2 could be improved by trying to force L55 near L97, but getting that to happen would be tough. I'll stick with what I submitted.