Tue Jun 15 08:31:47 PDT 2004 T0203 DUE 5 July 2004 Tue Jun 15 09:59:47 PDT 2004 Kevin Karplus T0203 is a 382-residue protein, so is probably multiple domain. The top hit (1rxxA) seems to be full-length, so we may not have to break into domains. All the strong hits (models 1-5 in T0203.t2k.undertaker-align.pdb.gz) are basically in agreement about the structure, so this becomes a homology modeling (loop building and sidechain packing) problem. There may need to be some modeling done at the N-terminal start of the chain. We'll have to check that undertaker doesn't throw away the good stuff found in the initial alignment. Tue Jun 15 17:47:29 PDT 2004 Kevin Karplus This looks like a single-domain protein, but the good hits are a little bit gappy, so that undertaker blows the protein up a bit in try1. There isn't a lot of good secondary structure in the original alignments, so hanging onto it is hard. We may have to do some packing of beta sheets by hand to get good structures. Wed Jun 16 09:13:08 PDT 2004 Kevin Karplus try2 is not blown up, but over half the constraints are not met, and the secondary structure doesn't match the predictions very well. Wed Jun 16 10:59:40 PDT 2004 Kevin Karplus I rebuilt pairwise alignments this morning, using both t2k and t04, and added more constraints to try3.costfcn. Let's see if a richer starting set and more automatic constraints provide any help at all in producing a better model. Wed Jun 16 21:30:53 PDT 2004 Kevin Karplus try3 may be a little better, but it still has some truly horrendous breaks: E297 and A298 are 30 Angstroms apart! There is something wrong here, as the cost for this break is reported as only 0.114, but it should be huge! Ahh---I see---the break costs reported have already been normalized by the protein length. I'll change undertaker so that we can see the individual break costs instead. Now I can see the bad breaks before P77 and A298 more clearly. I'll try rerunning with the break penalty turned up higher. Thu Jun 17 12:25:45 PDT 2004 Kevin Karplus try4 doesn't seem to have any huge breaks, but still has some big ones and try2 and try3 score better. Perhaps for try5 I need to increase the break penalty even more. Fri Jun 18 16:15:48 PDT 2004 Kevin Karplus try5 has no large breaks and only 6 breaks in total. Most of the structure (which is roughly a 5-fold repeat of hairpin-helix-||strand) seems to be intact, but I think that there is a strand at S290-G296 that should run parallel to G235-W241. I don't seem to have that connection made in any of the decoys. Hbonds are available at R233, G235, V237, and L239. The secondary structure is not coming out at all where we expect. Either the predictions are terrible, or the alignments we're picking are. Let's try again from scratch, without adding any Hbonds. Sat Jun 19 08:04:20 PDT 2004 Kevin Karplus It looked like I accidentally gzipped try6.log before the try6 run was done last night, and the opt1 part of the run wasn't finished until almost 6 this morning---I'll have to reduce the length of the runs in future. Mon Jun 21 13:17:50 PDT 2004 Kevin Karplus The try6 model is looking fairly complete, but there are what look like serious misalignments. The hairpin at D247-D265 is probably not a hairpin, but a helix, with the real hairin that makes a sheet with I209-R213 at L276-T285. Similarly, V60-V64 probably belongs Hbonded parallel to Q24-W29. P77-K103 probably forms an inserted antiparallel sheet. K270-L276 probably belongs as a strand antiparallel to L282-M284. I won't start with all these constraints, but will add more mutual information constraints and a few strand-strand constraints for parallel strand-helix-strand conformations. Wed Jun 23 21:09:23 PDT 2004 Kevin Karplus With the try8 costfcn, try7 scores best (there was no try8 run yet, so that costfcn was probably a placeholder for future work that I did not get back to). With the unconstrained costfcn, try6 scores best. I'm going to have to look at this one some more to decide what beta sheets to patch up. Thu Jun 24 12:04:35 PDT 2004 Kevin Karplus There are sheets consistently aligned in several models, and some about which there is disagreement. Looking at the robetta models as well as try6 and try7: robetta-model1: SheetConstraint Y13 M15 Q371 A373 hbond Y14 SheetConstraint A23 I27 P372 T368 hbond W26 SheetConstraint V60 V62 A23 T25 hbond T61 SheetConstraint D119 N123 D130 G126 hbond N121 SheetConstraint G126 G127 P153 R154 hbond G127 SheetConstraint S165 V168 T177 C174 hbond H167 SheetConstraint G172 E178 Q206 P212 hbond C174 SheetConstraint M228 A232 S240 V236 hbond C230 SheetConstraint G235 L239 I271 K275 hbond V237 SheetConstraint Y315 A317 A323 I321 hbond Y315 SheetConstraint G320 A323 H346 V349 hbond I322 robetta-model2: SheetConstraint S4 R5 P372 A373 hbond S4 SheetConstraint Y14 E18 P372 T368 hbond A17 SheetConstraint W120 A124 G131 G127 hbond F122 SheetConstraint G127 A128 P153 R154 hbond A128 SheetConstraint S165 V168 T177 C174 hbond H167 SheetConstraint G172 E178 Q206 P212 hbond C174 SheetConstraint S207 L211 S263 V259 hbond F208 SheetConstraint M228 A232 S240 V236 hbond C230 SheetConstraint G235 L238 D295 A298 hbond V237 SheetConstraint L260 I264 L276 Q272 hbond S263 SheetConstraint F314 A317 P324 I321 hbond Y315 SheetConstraint G320 A323 H346 V349 hbond I322 robetta-model3: SheetConstraint A23 I27 P372 T368 hbond W26 SheetConstraint Q24 G28 T61 S65 hbond I27 SheetConstraint D119 N123 D130 G126 hbond N121 SheetConstraint G126 G127 P153 R154 hbond G127 SheetConstraint S165 V168 T177 C174 hbond H167 SheetConstraint G172 T177 Q206 L211 hbond C174 SheetConstraint S207 L211 S263 V259 hbond F208 SheetConstraint C229 A232 L239 V236 hbond C230 SheetConstraint G235 L238 D295 A298 hbond V237 SheetConstraint L260 I264 L276 Q272 hbond S261 SheetConstraint Y315 A317 A323 I321 hbond Y315 try6-opt2: SheetConstraint Y13 Y14 Q371 P372 hbond Y14 SheetConstraint A23 I27 A72 L76 hbond T25 SheetConstraint Q24 W26 Q371 Q369 hbond W26 SheetConstraint W120 A124 G131 G127 hbond F122 SheetConstraint G127 N129 H157 M159 hbond A128 SheetConstraint G172 V176 N183 E179 hbond C174 SheetConstraint E179 K184 W210 L215 hbond L182 SheetConstraint I209 L215 V254 P248 hbond L211 SheetConstraint M228 A232 S240 V236 hbond C230 SheetConstraint Q249 R252 A266 S263 hbond E251 SheetConstraint F314 A317 A323 G320 hbond Y315 SheetConstraint G320 A323 S347 G350 hbond I321 try7-opt2 SheetConstraint Y13 Y14 Q371 P372 hbond Y14 SheetConstraint Q24 I27 Q371 T368 hbond W26 SheetConstraint N121 A124 D130 G127 hbond F122 SheetConstraint G172 V176 N183 E179 hbond C174 SheetConstraint E178 K184 I209 L215 hbond L182 SheetConstraint I209 Y216 V254 D247 hbond L211 SheetConstraint Q249 R252 A266 S263 hbond E251 SheetConstraint F314 A317 P324 I321 hbond Y315 I can group these into clusters and pick out the agreed-on strand pairs: SheetConstraint S4 R5 P372 A373 hbond S4 SheetConstraint Y13 M15 Q371 A373 hbond Y14 SheetConstraint Y13 Y14 Q371 P372 hbond Y14 SheetConstraint Y13 Y14 Q371 P372 hbond Y14 SheetConstraint Y14 E18 P372 T368 hbond A17 SheetConstraint A23 I27 A72 L76 hbond T25 SheetConstraint Q24 G28 T61 S65 hbond I27 SheetConstraint A23 T25 V60 V62 hbond T61 SheetConstraint A23 I27 P372 T368 hbond W26 SheetConstraint A23 I27 P372 T368 hbond W26 SheetConstraint Q24 W26 Q371 Q369 hbond W26 SheetConstraint Q24 I27 Q371 T368 hbond W26 SheetConstraint D119 N123 D130 G126 hbond N121 SheetConstraint D119 N123 D130 G126 hbond N121 SheetConstraint W120 A124 G131 G127 hbond F122 SheetConstraint W120 A124 G131 G127 hbond F122 SheetConstraint N121 A124 D130 G127 hbond F122 SheetConstraint G126 G127 P153 R154 hbond G127 SheetConstraint G126 G127 P153 R154 hbond G127 SheetConstraint G127 A128 P153 R154 hbond A128 SheetConstraint G127 N129 H157 M159 hbond A128 robetta only: SheetConstraint S165 V168 T177 C174 hbond H167 SheetConstraint S165 V168 T177 C174 hbond H167 SheetConstraint S165 V168 T177 C174 hbond H167 robetta only: SheetConstraint G172 E178 Q206 P212 hbond C174 SheetConstraint G172 E178 Q206 P212 hbond C174 SheetConstraint G172 T177 Q206 L211 hbond C174 try6 and try7 only: SheetConstraint G172 V176 N183 E179 hbond C174 SheetConstraint G172 V176 N183 E179 hbond C174 try6 and try7 only: SheetConstraint E178 K184 I209 L215 hbond L182 SheetConstraint E179 K184 W210 L215 hbond L182 SheetConstraint S207 L211 S263 V259 hbond F208 SheetConstraint S207 L211 S263 V259 hbond F208 SheetConstraint I209 L215 V254 P248 hbond L211 SheetConstraint I209 Y216 V254 D247 hbond L211 SheetConstraint M228 A232 S240 V236 hbond C230 SheetConstraint M228 A232 S240 V236 hbond C230 SheetConstraint M228 A232 S240 V236 hbond C230 SheetConstraint C229 A232 L239 V236 hbond C230 SheetConstraint G235 L239 I271 K275 hbond V237 SheetConstraint G235 L238 D295 A298 hbond V237 SheetConstraint G235 L238 D295 A298 hbond V237 SheetConstraint Q249 R252 A266 S263 hbond E251 SheetConstraint Q249 R252 A266 S263 hbond E251 SheetConstraint L260 I264 L276 Q272 hbond S263 SheetConstraint L260 I264 L276 Q272 hbond S261 SheetConstraint F314 A317 P324 I321 hbond Y315 SheetConstraint F314 A317 A323 G320 hbond Y315 SheetConstraint F314 A317 P324 I321 hbond Y315 SheetConstraint Y315 A317 A323 I321 hbond Y315 SheetConstraint Y315 A317 A323 I321 hbond Y315 # try6 is one of from robetta models: SheetConstraint G320 A323 H346 V349 hbond I322 SheetConstraint G320 A323 H346 V349 hbond I322 SheetConstraint G320 A323 S347 G350 hbond I321 Consensus: SheetConstraint Y13 M15 Q371 A373 hbond Y14 # one model made it longer and hbonded with the opposite phase, but # aligned the same residues SheetConstraint A23 I27 P372 T368 hbond W26 SheetConstraint D119 A124 D130 G127 # disagreement about phase of hbonding, but not about pairing SheetConstraint M228 A232 S240 V236 hbond C230 SheetConstraint F314 A317 P324 I321 hbond Y315 Thu Jun 24 13:47:45 PDT 2004 Kevin Karplus For both try7 and try8, the preferred alignment seems to be heavily based on the T0203-1lxyA-t2k-muscle.a2m alignment. Maybe I should add that to show-align.under and see what the alignment-only structure looks like. Thu Jun 24 15:09:03 PDT 2004 Kevin Karplus I didn't get a chance to look at the models (the file server here was dead, so I couldn't run rasmol) before starting try8. I'm pretty sure that both undertaker and robetta are dealing with a bad alignment near the N-terminus, and that Q24-G28 is parallel to P59-A64 22> haQTWIGwp 58> epVTVCAsp Thu Jun 24 17:43:37 PDT 2004 Kevin Karplus I spent 2 or 3 hours trying to hand-edit the alignment to fix up the early strands and modify the cost function to match. I'm still not sure I've got it---there has to be a better way to position elements along an existing template! I did not try to fix the end of the alignment---let's see if this attempt to fix the beginning goes anywhere. Fri Jun 25 07:52:20 PDT 2004 Kevin Karplus try9 managed to form some bits of sheet with the stuff I had unaligned and put sheet constraints on, but they did not get placed in a reasonable position. This protein is too big and complicated for me to build using just sheet constraints, so I'm stuck either believing the alignments (and ignoring the poor secondary structure prediction) or looking for a whole new fold. Since the E-values for the fold are so good (9e-10 for t04 and str2+CB_burial_14_7), I'm a bit reluctant to try such a big job building by hand. The t04 multiple alignment (mainly of "Peptidylarginine deiminase and related enzymes") looks pretty good, and 1rxxA (our top hit) is an arginine deiminase, so I'm really reluctant to change the fold. All the conserved residues of the multiple alignment are identical in the target, so this isn't a case of drifting to a different fold. Maybe I should treat the secondary structure prediction as poor and do optimization from all existing models without the constraints or predicted alpha terms? Fri Jun 25 09:58:38 PDT 2004 Kevin Karplus I looked at how well the str2 network predicted the templates using the template alignment. Rather than re-run the predictor, I just looked for the templates in the existing str2 by-chain files. The 1h70A predictions seemed fine (1.0 bits/position), and the 1jdw ones were poor but not terrible (about 0.5 bits/position). Sat Jun 26 12:40:59 PDT 2004 Kevin Karplus In try10 (our best scorer when ignoring 2ry-structure prediction) looking at conservation: D169, H224, and N187 cluster tightly (good). C366 is nearby and maybe should be part of the cluster. W91, D94, R93, and D89 are all on the surface. That's good for the polars, but I'd like to see the TRP buried. Perhaps the helix should be moved so that W91 is close to W29, which is also exposed. E18 and D226 seem to be buried. Although try10-opt2 is our best-scoring, and try10-opt2.repack-nonPC is our second best, Rosetta likes try8-opt2.repack-nonPC better. Fri Jul 2 1:15pm Jenny Draper I found a paper from a group who has cloned, expressed, and purified this protein: Janowitz T, Kneifel H, Piotrowski M. Identification and characterization of plant agmatine iminohydrolase, the last missing link in polyamine biosynthesis of plants. FEBS Lett. 2003 Jun 5;544(1-3):258-61. Highlights are: * it's an agmatine deiminase (iminohydrolase) * it's a homodimer * two cystine residues (180 & 366) are in/near the catalytic pocket, but do not perform the catalysis (mutation reduces, but does not eliminate functionality) * two other cystine residues (229 & 230) are conserved but not necessary for function So... the top hit (1rxxA, arginine deiminase) is definately the right template.... ------------------------------------------------------------ Fri Jul 2 15:01:27 PDT 2004 Kevin Karplus OK, so we're sure it is a pentein---now we have to get the right alignment. In our conservation measures, C366 is strongly conserved (it is the one that drastically reduced enzyme function), but C180 is only weakly conserved. C229 and C230 are not conserved at all. For try11, I'm running with very strong constraints, in the hope of catching an alignment that is compatible with our secondary structure predictions. I've also included the 4 constraints George predicted, which really amount to predicting two close approaches: S90-A124, S90-W120 S90 is not conserved but is wedged between conserved D89 and W91, A124 is not conserved, but W125 is, and W120 is weakly conserved. This may be part of a hairpin-helix-strand unit of the pentein. V237-I271, V236-I274 V236 is not conserved, V237 is weakly conserved. I271 is weakly conserved and I274 is not conserved. This looks like a probable strand-helix-strand pairing -------------------------------------------------- Fri Jul 2 17:57:38 PDT 2004 Kevin Karplus Try11, with the strong constraints, does seem to have aligned strands and helices a bit better, though there are few correct strand pairings. Still a number of misalignments---probably didn't have an alignment in the set of different alignments that did the right thing. Another possibility is that the "TryAllAlign" at the beginning got trapped into this alignment and moving out was too difficult. I wonder if I can do a run from fragments and occasional alignment insertions, rather than from TryAllAlign, so that there are more initial starting points. Perhaps we need an operator for shifting an alignment on a template---that would involve massive changes to undertaker though, as well as quite a different philosophy about how alignments are handled. What I could do is to generate more different alignments by tweaking the alignment-generation, but I don't know if there is time for that on T0203. I'll think about it. Fri Jul 2 22:49:40 PDT 2004 Kevin Karplus FOO! I forgot to change try11 to try12 in try12.under, so the outputs of the try12 run completely overwrote the try11 run! This is bad, because the try12 results were ab initio ones, and so not complete models! In fact, the try12 run does not manage to get back even the sheet constraints that we put in. I'll run try13 as an approximate repeat of the try11 run that got overwritten. The try13 cost function, with constraints turned way up, likes best the ab initio model from try12 (called try11-opt2), probably because the strands are extended and the helices coiled, even though almost nothing pairs up right. I'm hoping that the strong constraints with TryAllAlign might find a better alignment, though this is a pretty forlorn hope unless Jenny (or someone) has added more alignments and updated the read_alignments lists with "make read_alignments". Sat Jul 3 07:07:50 PDT 2004 Kevin Karplus I think I made the run for try13 a little too long---it still hasn't finished. The try13-opt1 model scores much better than the try11-opt2 one (from try12), and it is the probably the best alignment we've made so far, though it is rather gappy. I'm building some new 3-track alignments with str2 and burial weights turned way up. I thought that high weights on secondary structure tracks would only be useful for distant homologs, but it looks like we need it even for this close one. It looks like the hand alignment I started making to 1lxyA was being useful in this process, so maybe I should try hand editing the alignment some more. Foo---that is tedious work! Sat Jul 3 07:22:14 PDT 2004 Kevin Karplus OK, try13 done and try13-opt2 really does look better than most of the attempts so far. For try14, I'll do a similar thing to try13, but with fewer, shorter runs and with the new alignments included in the mix. I also turned up breaks and hbond_geom_beta_pair slightly to reshape the cost surface a bit and perhaps start from a different intial alignment. For try15, I'll start with a number of models and try getting some crossover with constraints not turned up quite so high. I've added a number of distance constraints from the t04-thin*.mi.constraints file to the try15 costfcn. It still scores try13-opt2 the best, so they probably are not hurting the cost function too much. The distance constraints were chosen to be the highest-weight constraints in each file, taking only ones with weight above about 0.7 using the new -0.5*Z_score weighting scheme. When these runs are done, I'll probably try to add sheet constraints to make sheets where strands appear to be trying to come together, and remove any mutual-information constraints that appear to be bogus. Sat July 3 1:45pm Jenny Draper I'm trying a VAST alignment from try13-opt2, hoping to get an alignment to play with in cn3d. So far, I haven't found any other decent alignment editors that will work on my windows machines. Is there a way I can view our current alignments in cn3d? Sat July 3 2:00pm Jenny Draper The VAST alignment/search is done: ID VS59447, pwd T0203try13 It likes 1lxyA best, 1rxxC next. Looking at the results now... Sat Jul 3 14:14:16 PDT 2004 Kevin Karplus Try14 is done, try15 half done. The try15 cost function likes try15-opt1 best, try14-opt2 next, then try13-opt2. The unconstrained cost function likes try10-opt2 best, then try6-opt2, then try5-opt2. It thinks that try14-opt2 is the best of the models created with strong constraints. Sat July 3 3:30pm Jenny Draper I'm not sure how to incorporate the VAST results into the T203 directory/undertaker runs? Kevin, I like some of the structure I'm seeing in the top alignment, to 1lxyA. 1lxyA is a homodimer crystalized with a ligand, and the active site and positioning of C366 seem pretty damn good, except for the breaks around residues 106-124, and the gap at 39-40 may be pretty shaky, since there's a loop there in lxyA that folds into the acive site. Take a look.... I'm gonna give up for now, but I'll check back now and then; email me if you think I can still help! -Jenny Sat Jul 3 18:09:13 PDT 2004 Kevin Karplus I picked up the alignments from VAST that Jenny created and put them in the 1lxyA and 1rxxA directories as 1lxyA-T0203-try13-vast.a2m and 1rxxA-T0203-try13-vast.a2m. Maybe I should rerun VAST with try15 Sat July 3 6:30pm Jenny Draper Why not? I just started a new VAST run from try15-opt1. ID: VS59452 PWD: T0203try15 Sat Jul 3 18:31:57 PDT 2004 Kevin Karplus I've submitted try15-opt2 to VAST also as job VS59451 password T0203try15 When I get those alignments back, I'll try adding them to the mix and run try16. The try16 costfcn will be the try15 one, with mutual information distance constraints that were badly violated commented out. I'm trying to decide whether to include sheet constraints or not---they could improve the quality of the sheets, but would lock in the current alignment. Maybe I should pick up sheet constraints from the 1lxyA-T0203-try15-vast alignment when it is done. Sat Jul 3 21:15:25 PDT 2004 Kevin Karplus I picked up the top three alignments (1lxyA, 1h70A, 1rxxA) for the two try15 VAST runs. They are slightly different between the opt1 and opt2 alignments. For try16, I'll not add sheet constraints, but I will toss the new alignments into the mix (but not the old models). Sun Jul 4 08:27:08 PDT 2004 Kevin Karplus The try16 run, starting from alignments, scores almost as well as the try15 run, which had started from models. There are still problems around 77-135. We should probably eliminate the strand pair SheetConstraint (T0203)W120 (T0203)A124 (T0203)G131 (T0203)G127 hbond (T0203)F122 instead trying for something like R81 M84 D94 W91 W91 S95 T98 R102 (length mismatch need to trim back) Sun Jul 4 10:08:03 PDT 2004 Kevin Karplus I did a lot of constraint editing for try17.costfcn, picking up sheet constraints from try16-opt2 and tweaking them (or eliminating them). I'll do a run for try17 and see if I can improve it at all. I also want to create a new operator for undertaker: ShiftSegment (and ShiftSubtree) that tries shifting a portion of the model over 1 or 2 residues. This would be particularly useful to do for remodeling beta strands. Sun Jul 4 12:11:49 PDT 2004 Kevin Karplus try17 hasn't finished yet, but I'm starting try18 with the new ShiftSegment and ShiftSubtree operators, using the same cost function as for try17. When try17 finishes, I'll look at it, make up a new cost function, perhaps get a new VAST alignment, and try again for try19. I think that we'll have to hand-define more of the sheet constraints we want, because I don't really like the looks of try17-opt1. try18 with the new operators is not blowing up, and the operators do seem to make some big improvements occasionally. Sun Jul 4 13:34:29 PDT 2004 Kevin Karplus try17-opt2 submitted to VAST: job VS59474 password T0203try17 I'm not real fond of sheets formed in try17---there are still a lot of secondary-structure conflicts. I wonder if I should try making a hand alignment in which P248 is aligned where R269 is currently aligning. In the try15 alignments, this region is unaligned. The new shift in try18 don't seem to work once the model has started closing up the gaps---probably the clashes and breaks caused at the ends of the moved segment or subtree are too big a penalty. I suspect that these operators will be best for ab-initio prediction, not for tweaking beta sheets in comparative models, as I had initially hoped. Sun Jul 4 14:36:09 PDT 2004 Kevin Karplus I superimposed all the VAST alignment predictions with show-vast-align.under into T0203.vast-align.pdb. Some of the alignments have parts I don't like (models 1,2,3 and 6,7,8 have bad helices that should be strands included). Those are 1lxyA-try13, 1lxyA-try15-opt1, 1lxyA-try15-opt2, and 1rxxA-try13, 1rxxA-try15-opt1, 1rxxA-try15-opt2. The remaining ones (various try17 alignments) all look better, though only models 11 and 12 avoid some of the other errors, mainly by being much shorter. Model 11= 1h70A-T0203-try15-opt1-vast, Model 12=1h70A-T0203-try17-vast. I could start run try19 from just those two alignments, and try to build up the sheets from fragments. I'm not very hopeful that will work, but it is probably worth trying. Hmmm, the initial models are terrible (way too much still from the initial random conformation), and InsertAlignment gets called a lot, eventually inserting the bad alignments that we were trying to avoid. Sun Jul 4 16:41:18 PDT 2004 Kevin Karplus try18 is done, and scores better with the try17/try18 costfcn than try17-opt2 did. It still has the (probably incorrect) pair of strands at Q249-A266, but scores better with the try19 cost function than try17 did. Try19 is unlikely to catch up, since the try19-opt1 costs are worse than the try18-opt1 and try17-opt1 costs, with the try19 costfcn. Maybe I should try polishing up the existing models since I have to submit tomorrow morning, even though I don't think we have the strands quite right yet. Sun Jul 4 20:14:34 PDT 2004 Kevin Karplus I made a "just-in-case" submission of 1 T0203.try18-opt2.repack-nonPC.pdb (best scoring with try20.costfcn) 2 T0203.try17-opt2.pdb 3 T0203.try10-opt2.pdb (best with unconstrained.costfcn) 4 T0203-1rxxA-t2k-local-str2+CB_burial_14_7-0.4+0.4-adpstyle5.a2m (full auto) 5 1h70A-T0203-try15-opt1-vast.a2m (fewest conflicts with 2ry prediction) Meanwhile, I'm running try20 to see if I can get a little gap closure. Mon Jul 5 08:12:44 PDT 2004 Kevin Karplus try20-opt2 and try20-opt.repack-nonPC do score better and don't look any worse in rasmol, so I've replaced model1 with try20-opt2.repack-nonPC and resubmitted. The gaps and misalignment are still pretty horrible. I'll try a quick try21 run with the weights of the individual constraints adjusted and with the weights of the different cost components adjusted. The adjustments were not very principled (increased bystroff and pred_alpha weights, to favor structures that fit 2ry prediction better; increased breaks and clashes, to try to close things up more). The individual constraints put more weight on the sheet constraints and less on the strand constraints. The idea was partly to try to get some of the pairings that still weren't forming and partly just to shake things up out of the local minimum. Sun Sep 19 10:29:26 PDT 2004 Kevin Karplus I added REAL_PDB:=1vkpA FINAL_COSTFCN:=try21 to the Makefile and did a whole-chain rmsd evaluation. The ranking of our submitted models and the robetta models were model5, model4, robetta1, model3, model1, robetta3, model2, robetta5, robetta2, robetta4. Both model5 and model4 are from alignments and may be artificially good through being incomplete. Thus it may be that robetta beat us on this target, at least on whole-chain rmsd, which is not very meaningful when the models are this far off. model3 was our best unconstrained model, so it looks like the constraints were hurting us a bit here. Wed Sep 22 10:42:51 PDT 2004 Kevin Karplus Using GDT score, the order is model4 32.83% robetta3 30.63% robetta1 29.40% robetta5 28.02% robetta2 26.03% try9-opt2 25.34% our best full model model5 22.87% model1 22.32% model3 22.32% model2 21.22% robetta4 17.10% The first initial alignments was better than our best full model. The constraints were not really hurting, since model1 (strong contraints) beat model3 (unconstrained) on GDT. Fri Sep 24 13:41:11 PDT 2004 Kevin Karplus name length missing_atoms cost rmsd rmsd_ca GDT smooth_GDT model4.ts-submitted 382 769 12.1656 11.4823 -30.4945 -28.6902 robetta-model3.pdb.gz 382 0.0000 15.9462 15.4521 -28.7775 -27.5240 robetta-model1.pdb.gz 382 0.0000 14.3181 13.5407 -26.4423 -25.4314 robetta-model5.pdb.gz 382 0.0000 16.0568 15.5224 -26.0989 -24.8650 T0203.try9-opt1.pdb.gz 382 0.0000 15.6742 15.0825 -24.2445 -23.3033 robetta-model2.pdb.gz 382 0.0000 16.9639 16.5287 -23.0769 -22.3313 model5.ts-submitted 382 1302 10.3708 9.6268 -22.3901 -21.0907 model3.ts-submitted 382 0.0000 15.5470 14.9441 -20.9478 -20.8814 model1.ts-submitted 382 0.0000 15.7744 15.1379 -20.7418 -20.4986 model2.ts-submitted 382 0.0000 15.9652 15.5556 -19.5055 -18.8917 robetta-model4.pdb.gz 382 0.0000 18.4218 17.6569 -15.5907 -15.2503 Our model4, which was an alignment is our best model. Our best complete model was try9-opt1.