Mon Jun 24 09:16:16 PDT 2002 25 June 2002 Kevin Karplus The T0138.try1-opt.pdb structure does not look too horrible, except that the last strand seems to have been peeled off the sheet. The t2k-2track-undertaker alignment with 1dbwA only has 3 of the 4 strands. The template itself has more strands (1 5-strand parallel sheet), so we can extend the alignment to line up the 4th strand. The T0138-1dbwA-2track-global-adpstyle5.pw.a2m.gz alignment tries to get all 5 strands, but to do so it has to align one of the strands with a region predicted to be helix. The 1dbwA-T0138-fssp-global-adpstyle5.pw.a2m.gz alignment gets three of the strands well, then makes the next strand out of something predicted to be helix and makes a helix out of the region predicted to be strand. It also makes the first strand out of a region predicted to be helix. 1hf2A-T0138-fssp-global-adpstyle5.pw.a2m.gz has the 4th strand, but has a terrible alignment before that, with a sheet formed from pieces predicted to be helix. I think we may have to play with hand alignment (perhaps to 1dbwA) to get the alignments we want. I'm also not certain that a parallel sheet is what is called for, as the first and last strands are predicted to be anti-parallel, while strands 2 and 3 are predicted to be parallel. I'm not sure how much we should believe the STR predictions on direction though. Hmm--the doc file says that chemical shifts are on file in the BMRB database. Supposedly those numbers can be used to make a secondary structure (backbone geometry) prediction that is more accurate than neural-net predictions from sequence data. I wish I knew how to use chemical shift info. (This protein will probably be a big win for the Rosetta group, who DO have a method for using chemical shift info.) 12 July 2002 Kevin Karplus Reran undertaker with new fragments from improved fragfinder. In try2-opt, three of the strands form a nice sheet, but the 4th strand is way out in space. Tried again with more alignments to sample---now the 4th strand winds up into a helix! The strand is L98-G107, and it probably pairs with the third strand, though the direction is not obvious (probably parallel, since the rest of the sheet is). It looks like I'll have to guess the alignment of the strand and add some constraints. 17 July 2002 Kevin Karplus Noticed that the try3 run DID NOT sample more alignments---it had a mistake in the commands that were supposed to include the read-alignments.under scripts. 18 July 2002 Kevin Karplus The try4-opt run looks pretty good. There is a five-strand sheet, one strand of which is predicted to be a helix. It also looks like there are two disulphide bridges: C20-C53 and C8-C34 I'll add the disulphide bridges to the scoring function and try again. In try5-opt, the disulphide bridges are nicely formed, but the C-terminal strand Y99-H105 has been peeled off the sheet. It isn't involved in the disulphides, so it should be fairly independent. The Hbonds for this strand in try4-opt are 636 G83 N 819 H105 O 627 V81 O 811 H105 N 622 V81 N 801 E103 O The strand is then broken in try4-opt, but the pattern would continue 612 A79 O 794 E103 N 609 A79 N 787 S101 O Let's try adding these constraints. In try6-opt, the SSbonds are again very nicely formed, but the two edge strands have peeled off a bit from the sheet. Perhaps I need to increase the weight of the constraints relatve to the known_ssbond weight. In try7-opt, the disulphides again look good, but the strands are stil peeled off, looking about the same. I think I need to add some constraints (based on H-bonds in try4-opt) to keep the sheet together: Constraint 53 406 2.0 2.7 3.2 // C8 N I51 O Constraint 57 416 2.0 2.7 3.2 // C8 O C53 N Constraint 67 420 2.0 2.7 3.2 // W10 N C53 O I'll rerun with these added and also see whether the new "scwrl_each" option to OptCommands in undertaker works. In try8-opt, we have a 5-strand sheet, but the middle strand of the sheet seems to have slipped out of position a bit, and the edges strand L98-L106 has peeled away from the sheet for half the strand. In try9-opt part of the strand around L98-S101 has curled up into a helix, and the strand at A6-V11 is close to where it ought to be, but not close enough to Hbond. For this run I was trying to slide the middle strand back 2 to match the 2ry predictions: Constraint 438 620 2.0 2.7 3.2 // L56 N I80 O Constraint 444 629 2.0 2.7 3.2 // L56 O V82 N Constraint 53 420 2.0 2.7 3.2 // C8 N C53 O Constraint 57 430 2.0 2.7 3.2 // C8 O I55 N Constraint 67 436 2.0 2.7 3.2 // W10 N I55 O but it does not seem to have resulted in a good structure. The problem may be that I started from a conformation that was already wedged and the search didn't move it far enough. Perhaps I should try again without any seed conformations. 19 July 2002 Kevin Karplus try10-opt-scwrl scores the best of any structure with the new scoring function, and the predicted strands line up fairly well. In try10-opt the beta strands barely form a sheet---they are just a little too far from each other, and there is a bad break between A102 and E103. Perhaps I should add some more constraints to hold the sheet together??? It is painful to have to add so many constraints to represent the missing H-bond function. I could also try including some of the old structures as starting points, though not the try9 ones. None of the structures so far have a good set of Hbonds for L98 through L106---that strand is almost always peeled off in some way. Perhaps I need to add more Hbonds---the obvious next ones are 801 O E103 N V81 622 already in score 794 N E103 O A79 612 already in score 787 O S101 N A79 609 already in score 783 N S101 O V77 600 already in score Maybe I should tighten these a bit to make them more important? Changing from 2.0 2.7 3.2 to 2.3 2.7 3.0 as the min,desired,max paramters for the Hbonds on this strand doesn't affect the ranking of the top decoys much. I tried passing the try10-opt structure to VAST, and it got good hits to 2minB, 1dcfA, and 1jkmB, which did not score all that well in either target or template scoring, nor did those contribute fragments to fragfinder. The FSSP rep for 2minB is 1qguB, which has a Z-score of 5.9 with 1dbwA, so there is some structural similarity. I'll try including a number of the VAST alignments in the initial pool for try11 (and no initial conformations) to see what comes up. Try11 certainly scores much better than anything I've seen so far, though try11-opt scores better than try11-opt-scwrl. There is a very bad break mid helix between V67 and Q68. The sheet looks pretty good though, so I'm tempted to try reoptimizing from here. 23 July 2002 Kevin Karplus try12-opt (not try12-opt-scwrl) is currently the best scoring. The beta sheet looks pretty good, but there are some bad breaks. The C20-C53 SSbond does not really look possible in this conformation, and the C8-C34 is misaligned by 1. Let's rescore taking out the SSbonds and adding in the new pred_alpha cost function. With this change, the best decoys are try12.9.30 and try12-opt-scwrl, with very little difference in score between them. After re-optimizing try12-opt-scwrl without the SSBonds, try13-opt scores best, but it looks like C8 C34 could definitely bond. Putting SSbond 8 34 back into the score function still leaves try13-opt as the best. Let's do one more run, increasing the break penalty and trying to close some of the bad gaps. 23 July 2002 Kevin Karplus New best is T0138.try14.7.30.pdb (slightly edging out T0138.try14-opt-scwrl.pdb). There are still 5 breaks, one of which is slightly smaller than in try13-opt-scwrl. In the try14 run, CloseGap was used a lot, but ReduceBreak was not---I think I need to look for bugs in ReduceBreak. T0138 is looking more or less done, assuming we believe in the helix-strand-helix for T14-T46. We predict "helix" for the whole region (except D27), but the prediction is weak for the region that we've made be a strand. 29 July 2002 Kevin Karplus Started over with no conformations read, but with lots of alignments (including the VAST alignments with previous predictions). The new low-cost structure is try15-opt-scwrl. I tried tweaking the score function a bit---adding the dry5 cost, decreasing break weight, and decreasing constraint weight. try15-opt-scwrl still scores best, but one strand has separated a little from the sheet and there are some nasty clashes between S101 and A79. Let's reoptimize starting from try14-opt-scwrl and try15-opt-scwrl. 30 July 2002 Kevin Karplus For some reason S101 and A79 are clashing very badly even in try16-opt. Maybe I have the constraints for the Hbonds wrong? Standard parallel Hbonds: O N H105 G83 V81 H105 E103 V81 A79 E103 S101 A79 V77 S101 Y99 V77 match what was echoed in undertaker-try16.log Let's try increasing the penalty for clashes and backbone clashes, decreasing the penalty for breaks, and increasing the weight of phobic_fit, to see if we can get one of the "good" models to score better. Hmm, try16-opt-scwrl still best, though it scores worst on most parameters---it seem to be picking up all its value from known_ssbond and pred_alpha2---let's try reducing both. I've tweaked the parameters until one of the try14 intermediates scores best, though only slightly. Let's do another run, starting from scratch, and see if it gets wedged also, or does well. 30 July 2002 Kevin Karplus After try17, try14.9.30 still scores best and MANY decoys beat try17-opt. Let's try another run, with the best structure or two from each of several runs. Tue Jul 30 17:25:19 PDT 2002 Ater try18, try18-opt scores best, but strand Y99-H105 is a bit messed up at the beginning, and there are still some bad breaks. The C8-C34 disulphide is not formed, though a change of rotamers should get it. Wed Jul 31 09:11:18 PDT 2002 Kevin Karplus Strand 99-105 is still too bent. Maybe I need to add a constraint to keep it straighter. Looking at the straight strand C53-V57 CB CYS A 53 -9.424 -66.541 10.507 CB VAL A 57 -18.380 -67.985 21.918 14.578 Angstrom It looks like CB-CB on the same side of a strand is about 3.64 Ang*chain separation. So L98-L106 should be about 29.15 Angstroms. I'm not sure the strand should be quite that straight. Perhaps more important is that S101 to G103 should be about 7.29 Angstroms Adding constraints Constraint 785 796 7 7.3 8 // CBs S101 G103 Constraint 755 823 22 29.15 35 // CBs L98 L106 and tweaking the parameters of the score function makes the lowest-cost decoy be try19-opt-scwrl. Let's try reoptimizing from there, Wed Jul 31 14:13:26 PDT 2002 Kevin Karplus Although try20-opt scores a little better, it really doesn't look much different. Tweaking the score function to be less afraid of breaks still leaves try20-opt-scwrl as the best. Let's re-optimize with several starting points and the tweaked scoring function. 1 Aug 2002 Kevin Karplus try21-opt is new lowest cost. The helices have been a bit messed up to get closed gaps, but the packing is pretty dense. Increasing the break penalty still leaves try21-opt on top. I played around with the weights of the scoring function a bit (turning off the SSbond scoring), then started a new job to reoptimize from try21-opt. This one is looking pretty close to done. I'd like the helices to be straighter, but I'm not sure how to get that and still close gaps. Fri Aug 2 03:34:38 PDT 2002 Kevin Karplus tr22-opt-scwrl now best. Packing is a little better, but SSbond has been discarded. I'll add constraints to straighten helix for F63-F72 and G37-T46 and comment out the constraints that kept strand3 straight. Doing so leaves try22-opt-scwrl as best. Putting a little weight back on the c8-c34 disulphide makes try22-opt better than try22-opt-scwrl, but there are many decoys with a much better known_ssbond score (so increasing the weight too much would really change the ordering). Increasing both the constraint weight and the break weight still leaves try22-opt on top. Let's try reoptimizing once more with the new weights and constraints. Fri Aug 2 13:41:22 PDT 2002 Kevin Karplus try23-opt is still a bit "foamy", not densly packed, and has some bad mid-helix breaks. A lot of fragment insertion was tried, and a fair amount of subtree jiggling, but CloseGap was rarely successful. The positioning of Y99 to A102 looks particularly bad. I'll try adding some CB constraints to move this strand parallel to its neighbor, as well as putting back in some old constraints to straighten around the kink. Even after making the changes, try23-opt still scores best. 3 Aug 2002 Kevin Karplus try24-opt looks much denser and cleaner, though there are still a few bad breaks. (particularly 67-68). Let's do another optimization run with the break and clash penalties turned up a bit. 4 Aug 2002 Kevin Karplus try25-opt scores a little better, though the big breaks at 67 and 100 are still there. Packing still looks pretty good. It would be nice to fix up the helix around 63-68. I changed the weighting parameters for the score (somewhat arbitrarily) so that the position of the local minimum would change, to give undertaker a chance to explore a little more around the current best. Doing so changes the current best, making try25.17.80.pdb, try25.2.80.pdb, try25.10.80.pdb, try25.14.80.pdb, ... score better than try25-opt. The gains are mainly in dry12 and pred_alpha2. Tweaking again to raise the penalty for clashes changes the order again---now the best iterations are 2, 17, opt, 10, 14, ... . Let's try reoptimizing from the first 5 of those. 5 Aug 2002 Kevin Karplus Best is now try26-opt. I don't expect to be able to improve it much, so let's submit it. 11 Aug 2002 Kevin Karplus Remaking with new template library moves 1d5wC to the top, but it was already in the "try-alignments" target, so I see no reason to redo the undertaker runs, even if the fragments are now slightly better.