Tue Aug 6 14:31:43 PDT 2002 t0169 Clear homology modeling to d.108.1.1---the same family as T0152. I suspect that we may have to do the same sort of constraints to hold the sheets together as we used in 152, but T0169 is a LOT closer to the templates, so should be much easier. 7 August 2002 Kevin Karplus try1-opt-scwrl looks OK, but not great. The region from 118 to 141 looks misaligned, so that the two strands in the region don't sit at the edge of the sheet. We'll probably have to add CB constraints, and maybe do some hand alignment, to get the sheet to form properly. 8 August 2002 Yael Added constraints to try and get the last part of the sequence contencted to the edge of the sheet. Next, we may have to make helix constraints on 113-121 since for some reason undertaker thinks it is a stard (though also the alignment and thess preditions assume it is a helix). 9 August 2002 Yael T0169.Try2-opt scores better then T0169.try1-opt. The end part of the sequence is now conected to the rest of the sequence but still far from being aligned right. There are still quite a few gaps and breaks. I added hbonds constrains to try and form the hairpin in the end of the sequence and length constraints to try andform the helix (K113-K121) this hopefully should help getting this part of the sequence aligned better. Added also CB constrains to try correct the register of 2nd and 3rd strad of the sheet. 9 August 2002 Yael Though try3-opt scores the best now, the structure dosen't look good. Seems that the constraints didn't help. The last part of the sequence looks much better in try2. The helix didn't form either. We should think of another way to align this part of the sequence to the sheet.... 22 Aug 2002 Kevin Karplus Added try-heavy target and tried try4 run without constraints. Without constraints, current best is try3.19.20. I suspect that this structure MAY be close enough to try giving to VAST to get a better alignment back. Let's see how try4 goes, then take the best structure and give it to VAST in the morning. Fri Aug 23 10:27:37 PDT 2002 Kevin Karplus New best is try4-opt. There is one bad break mid-helix (before S16) and a couple of others in loops (before S51 and K113). Although the orientation of the pieces seems pretty good, the protein is quite foamy. Let's try running VAST to se if we can get a better alignment: run VS30766 password casp5t0169. VAST does not see the badly broken helix. Best VAST hit is to 1b87A (1.1Ang over 141 residues). I cleaned up the loop alignment slightly, and am rerunning with only this alignment as an initial alignment. Fri Aug 23 16:31:46 PDT 2002 Best is now try5-opt. It looks like some constraints are needed: Extend parallel strand R103-V107 R103.CB T67.CB V104.CB A68.CB K105.CB Y69.CB C106.CB I70.CB V107.CB H71.CB Add parallel strand D125.CB R103.CB I126.CB V104.CB E127.CB K105.CB K128.CB C106.CB G129.CA V107.CA Add antiparallel strand K128.CB N141.CB G129.CA A140.CA T130.CB V138.CB K131.CB S137.CB I put these constraints into try6-opt. Try5-opt still scores best with these constraints. Let's try re-optimizing from try5-opt. 24 Aug 2002 Kevin Karplus 9:52 try6-opt is new best score, but I don't see any real improvement over try5-opt. The run spent allmost all its time in CloseGap, which is not very productive. Increasing the constraint weight makes try5.1.60 score best, which has a big beta sheet (but part of the sheet is predicted to be a helix)--this is very similar to try4-opt. Let's try another optimization run, with the constraints set high, from try4-opt, try6-opt, try5-opt, and try2-opt. 17:03 the try7 run seems to have died in the last SCWRL call before the final run on the super pool. The best score is for try7.9.100. I don't really like where strand D125-K131 has ended up (at right angles to the sheet) and I don't like how S137-T142 has coiled into a helix. Perhaps I need some straightness constraints? 25 Aug 2002 00:05 No, try7 didn't die, it was just stuck in SCWRL for a while. New best is try7-opt, but it has the same problems with S137-T142 and D125-K131 as try7.9.100. In try8.constraints, I tightened the constraints positioning D125-K131 by using Hbond constraints, added a straightness constraint for S137-N141. Also, the try6 constraints were inconsistent. When I fix all these constraints, the best score is try3-al10.pdb, and try7-opt is also pretty good. Given that initial alignments are scoring best, perhaps the try8 run should try adding alignments to try7-opt. 25 Aug 2002 10:15 Kevin Karplus try8-opt is now best score. S115-A140 almost forms a hairpin with S135-N141, but does not attach to R103-V017. Sheet Hbonds in try8-opt: D2.N I45.O D2.O I45.N R4.N S43.O R4.O S43.N T42.N L56.O T42.O L56.N F44.N G54.O F44.O G54.N T46.N S51.O Note:continuing would pair 52, not 51 T46.O S51.N F55.N G74.O F55.O G74.N Note continuing would pair 73 not 74 I57.N H71.O I57.O H71.N F59.N Y69.O F59.O Y69.N S61.N T67.O S61.O T67.N A68.N R103.O A68.O K105.N I70.N K105.O Cleaning up these constraints and adding them to the try8 constraints, but removing the H-bond constraints for K128+N141, T130+F139, since I'm not sure of the register for that hairpin, leaves try8-opt as best score (naturally, since H-bonds are derived from it). Let's do an optimization run with the try9.constraints and see what we get. 15:28 Top score is now try9-opt. I'll extend the H-bonds for the parallel connection of Y69 and K105 and try making up some hbonds for the last hairpin, putting them in try10.constraints. With these constraints try9.0.100 scores best. Let's reoptimize from best few, with high InsertFragment to try to flex things into place. 20:13 Top score is now try10-opt. There is a bad clash between G129 and A140. Do I have conflicting constraints? Oops, Yes I do. I have E127 hbonding on both sides of the strand! Some possible arrangements 103> RVKCV 125> DIEKGTK 141< NAFVS 103> RVKCV I favor this one 125> DIEKGTK 141< dyNAFVS 103> RVKCV 141< NAFVS 125> DIEKGTK 103> RVKCV 141< dyNAFVS 125> DIEKGTK I want R103 and and D143 to line up next to the same residue in the strand, probably D125. Hbonding is available on T102.O, V104, C106, and T108. With the pairing I've chosen, this leaves I126, K128, and T130 for H-bonding on the other side of of the strand. With these corrected constraints in try11.constraints, the best-scoring models are try10.2.100, try10-opt, try10.8.100. Let's try again from them (and toss in some older models that score OK, incase we've managed to get into a position that is haard to unwedge from). 27 Aug 2002 11:59 Best score now try11-opt, but the last couple of strands are STILL messed up. Maybe I should swap the last two strands bonding R103-T108 to S137-Y142. 103> RVKCVTs 143< dyNAFVSig 122< gfDIEKGTKtv 29 Aug 2002 13:50 Best score is now try12-opt (with try12.constraints) D125-K131 is a little ugly but not too bad. Packing is not quite solid, but not bad. Without constraints, best score is try5-try6.10.100 (that seems like a strange naming convention---our usual naming puts the optimization run name first). That run does NOT look good to me, as everything after about V107 looks wrong. The first part looks good though---It has a different arrangement of beta strands, hbonds: S43.N G58.O S43.O G58.N I45.O L56.N I45.N L56.O S47.N G54.O S47.O G54.N F55.N G74.0 F55.O G74.N bulge F59.N Y69.0 F59.O Y69.N A68.N R103.O A68.O K105.N I70.N K105.O Note: different arrangement of S42-S47 means that DIRT strand at D2 needs different Hbond specs > mDIRTit > tsFITSEhn or < tiTRIDm > tsFIFSEhn I put the first (parallel) arrangment into try13.constraints, copying compatible constraints from try12.constraints. The best score with the try13.constraints is try12-opt. 29 Aug 2002 17:42 try13-opt now best. We need to pack it a bit tighter---maybe constrain W22.CZ2 near H119.CE1? There is also the problem of the strand S137-D143 being upside down. Let's add a few more constraints for that part of the sheet, s well as the weak constraints for packing the helices in tighter. Now try13-try11-7.1.60 scores best, try13-opt second best, and try13.10.60 third. Let's reoptimize from them. 22:37 try14-opt seems to have made a mess of G123-G135. Perhaps I put in wrong constraints. I'm too tired to debug tonight. Fri Aug 30 10:00:23 PDT 2002 Kevin Karplus Without constraints, the best score is for model try6-opt, which is missing the hairpin from D125 to N141. Next best is try9.0.100, which only sort of has the hairpin and doesn't attach it to the sheet. Of more recent runs, the best-scoring is try10.5.100, which almost has the hairpin and Hbonds C106.O-E127.N and V104.O-D125.N, but the strand is almost at right angles to the rest of the sheet, because the preceding helix isn't long enough to let the strand flatten out. Perhaps I should try two pairings: V104.O D125.N C106.N D125.O C106.O E127.N T108.N E127.O or C106.N F124.O C106.O I126.N T108.N I126.O T108.O K128.N Note: try14 is trying to put the last two strands in the other order, which is not working. I've put the first set of constraints into try15.constraints With these constraints, the best models are try14.4.80, try13-try11-7.6.60, try14-opt, ... Oops, I didn't check the hairpin constraints---there are inconsistent Hbond requests---let's remove the Hbond constraints on the hairpin for try15, and rescore (same few come out on top). I put the second set of constraints into try16.constraints. Scoring with them puts try14-opt first but has roughly the same ones on top. Fri Aug 30 13:28:34 PDT 2002 The try15-opt file has trouble with the hairpin with G129 and A140 clashing as the hairpin twists around F124 and D125 make a good initial part of the strand, though a bit too curved. try16-opt has not managed to make sense out of the hairpin or the strand, but scores best with try15.constraints and try16.constraints. For try17, let's try adjusting the pairing for hairpin K128-N141, adding another short parallel strand on the end, and changing the first short strand from parallel to antiparallel. With try17.constraints and clashes and constraints turned up, try15-opt scores best. 31 Aug 2002 Kevin Karplus Best score is currently try17-opt, with the try17.constraints. Adding nonalpha_hydrogen_bonds to the scoring function, still leaves try17-opt as the best. Leaving in the Hbond scoring, but taking out the try17.constraints makes try4.15.60 score best. It has a lot more Hbonds, because it has made the predicted helix at R149-F152 into a strand. The prediction there is very weak, though, so a strand is certainly feasible. Strand S137-N141 is not hbonded to anything though it looks like it belongs parallel to F72-H76. The breaks in try4.15.60 are pretty bad though, so let's up the break penalty and rescore. Now try16.1.80 scores best, but it looks terrible, with breaks in bad places and lots of clashes. Reducing the break penalty to 20 and increasing the clash penalty to 2.5 makes try4.15.60 score best again. Let's start from there, adding constraints for S137-N141. 135> giSVFANydg 65> peTAYIHFSGVhpd (The other strong possibility is that S137-N141 folds back antiparallel to D125-K131 135> giSVFANydg 123< gfDIEKGTKtv but all the charged residues on D123-K131 make it look a lot like an edge strand. ) Let's add JUST the CB constraints for the new strand pairing as try18.constraints, and reoptimize with the nonalpha_hydrogen_bonds turned on. The best-scoring current decoys with this scoring function (omitting very similar ones) are try4.15.60, try7-opt, try12-try11-5.7.120, 14:11 The try18 run died. The current best score is try18.2.100, but it has mangled the helices and looks terrible. It seems to be trying to put K128 near V135 and strand S137-141 near H71-G74 simulatenously, but the two requests are not really consistent (only the strand placement was requested--the rest came the existing structures. If we set the score function to use hydrogen_bonds instead of nonalpha_hydrogen_bonds, the best score is still try18.2.100 If we use the try17 constraints and no Hbond term, instead of the try18 constraints, the best score is for try18.1.100, then try17-opt. With no constraints, but hydrogren_bonds on, the order is try18.2.100, try18.1.100, try18.0,100, try18.3.100, try4.15.60, try4.12.60, ... Let's try re-optimizing from try18.1.100, try18.2.100, try17-opt, with hydrogen bonds on and no constraints. 31 Aug 2002 20:51 Hmm---it looks like try19 was optimizing WITH the try18.constraints as well as with hydrogen bonds. The best with those constraints is now try19-opt. Without try18.constraints, but with hydrogen bonds the best score is also try19-opt. The structure looks terrible though, with all the strands having lost their hbonds and being scrambled up. The problem is with the pred_alpha2, which seems to be messing things up with VERY strong scores. If I take out the hydrogen_bonds scoring and use my version of undertaker, the problem goes away,so there appears to be a bug in Jonathan's code that is stepping on the pred_alpha data---probably a memory management problem. With no constraints, the best are currently try6-opt try9.0.100 With try18 constraints, the best are currently try7-opt try10.5.100 Looking at try7-opt, gives me an idea for another strand packing of F124-I126 and K128-V133 and I136-N141. Maybe K128-V133 and I136-N141 fit in the split between H71-H76 and K105-S109 71> HFSGVh 135> gISVFANy 101> ctrvKCVTsp 135< gnvtKTGKE Let's do another run, optimizing with the OLD version of undertaker and no Hbonds, and try20.constraints. Currently try20.constraints favors try10.5.100 and try7-opt. 1 Sept 2002 08:22 Kevin Karplus Best with try20.constraints now T0169.try20.4.80==try20-opt, but D125-K131 and S137-N141 don't form Hbonds with anything. With try18.constraints, best is try20-opt. With try17.constrains, best is try17-opt. With try16.constraints, best is try16-opt. With no constraints, best is try6-opt. None of these look great to me---I'm having trouble figuring out where to put those C-terminal strands. Maybe a VAST run would help me recognize a structural homolog? Run VS30951 password casp5t0169. With VAST, the best alignment is to 1qsmD, which ends before this C-terminal part. I hand edited to allow a little more room for reorganization where the loops were different lengths. I also changed the sequence from 1qmsD to 1qmsA, since we have that one in our template library already (it is 100% identical and within 0.5 Angstroms). Let's try inserting that alignment into several of the the best models, then reoptimizing without constraints. From karplus@bray.cse.ucsc.edu Sun Sep 1 08:44:52 2002 Date: Sun, 1 Sep 2002 08:43:53 -0700 From: Kevin Karplus To: karplus@soe.ucsc.edu, rachelk@soe.ucsc.edu, weber@soe.ucsc.edu, learithe@cats.ucsc.edu, yael@biology.ucsc.edu, baertsch@soe.ucsc.edu, rph@soe.ucsc.edu, afyfe@soe.ucsc.edu, jcasper@soe.ucsc.edu, oscarhur@soe.ucsc.edu Subject: T0169 The next unsubmitted target that is due is T0169 (due 4 sep). I'm having trouble findng a place for the C-terminal strands. Can anyone help? ------------------------------------------------------------ 1 Sep 2002 16:52 Best score with no constraints is now try21-try20.6.100 Both the C-terminal strands V138-Y142 and V153-K154 are wound up into helices. No hints here that I can see about where the strands are best placed. 2 Sept 2002 15:51 Jonathan says he's fixed the bugs in the hbond counting. With either hydrogen_bonds or nonhelix_hydrogen_bonds and no constraints, the best score is for try21-try20.6.100. The nonhelix hydrogen bonds have a a mean of 0.1666 bonds/residue, but a standard deviation of only 0.048, so we need to up their weight if they are to affect the score (the cost has a standard deviation of 5.5 over the same set of decoys). If we up the weight to 10, try19-opt scores best---something is wrong here, as it has -0.0641026 for the nonhelix_hydrogen_bonds, and try21-try20.6.100 has -0.121795, so raising the weight of nonhelix_hydrogen_bonds should NOT have moved try19-opt up. Let's see if the other scores have changed relative to what they are computed with undertaker---it is still the pred_alpha2 value that is being changed. Tue Sep 3 10:52:02 PDT 2002 Kevin Karplus I superimposed the top few *-opt models scored with no constraints: try21, try6, try20, try9, try7,try5, try8 using superimpose.under to put them in T0169.nocons.super These agree for a large part of the model, and try8 seems best where they disagree. I also superimposed the top few *-opt models scored with each of the constraint sets: try18.super: try21, try6, try7, try9, try12, try8 Again, try8 looks best. try17.super: try17, try15, try16, try13, try14, try7, try4, try9 at the N-terminus try4 looks best (try9 is similar). try4 also looks interesting at the c terminus. try14 has a bad clash but tries to get an extra strand on the sheet. try16.super: try16, try17, try15, try14, try13, try4, try7, try6, try5 try4 gets more sheet than the rest, but puts some residues predicted to be helix into the sheet. try15.super: try17, try15, try16, try14, try13, try7, try4, try11, try12 Tries 7, 4, 11, and 12 all get similar N-termini, with one more strand than most models. tries 11 and 12 get longer strands than most, but try4 still has the best c-terminus. Try15 does get a bit of the penultimate strand, but has bad clashes. try14.super: try14, try13, try15, try16, try17, try12, try10, try11, try8 None of tries 14,13,15,16 look particularly promising. try17 gets a longer strand2, at the cost of a bad helix break. Tries 10,11, and 8 all have similar N-termini, but differ on the last two strands. Try11 gets the final strand parallel to the sheet but has messed up the penultimate strand. try13.super: try14, try13, try15, try16, try17, try12, try10, try11, try8 (same set as for try14.super) try12.super: try12, try10, try11, try8, try9, try13, try14 Tries13 and 14 have the N-terminus parallel rather than anti-parallel, and try13 manages to length the second strand as a result, though it doesn't have the Hbonds quite lined up. The final strand is almost in position---just a little too high, but the penultimate strand is not. Perhaps try13 deserves some more polishing with appropriate constraints. try11.super: try11, try10, try12, try8, try9, try17, try15 try10.super: try10, try11, try9, try8, try12, try17, try15 try9.super: try9, try10, try8, try21, try20, try11, try12, try17 try8.super: try9, try10, try8, try11, try16, try17, try15 try6.super: try9, try21, try8, try20, try11, try12, try7 try3.super: try11, try10, try17, try16, try9, try17, try15 try2.super: try9, try11, try8, try20, try11, try12, try10 If any of these constraints sets is right, the model it lists first should proably be submitted. The candidates are try21, try17, try16, try14, try12, try11, try10, try9 Essentially al of our models agree on strands 2-5. Decisions to make (or alternative models to propose): N-terminus: strand 1 D2-T5 strand 2 T41-E48 parallel (try13) or antiparallel (try8, 10, 11) The parallel connection allows a slightly longer strand 2, but does not explain why there are 2 helices, instead of 1 between the strands. C-terminus: strand 5 R103-V107 strand 6 D125-K131 strand 7 S137-N141 6^7v not part of sheet (try8) 5^6^7v (try10, 11) 5^7v6^ (try13) 5^6^ and 7^4^ (try20) 5^7v no 6 (try4) Based on the number of charges in strand 6, I favor 5^7v6^ to put 6 on the edge. We seem to have linked the combinations of the N and C termini options. Let's try polishing up try13 for both the C and N termini, with better constraints at each end. I put the new constraints in try22.constraints. Scoring with them makes try14-opt look best 3 Sept 2002 Kevin Karplus The try22 run finished, but pointed out to me one problem with try13 and try14 (the parallel positioning of strand1)---they both lifted the helices away from the sheet, exposing a lot of hydrophobics. try8 and try11 are foamy, but not nearly as bad. Let's try the antiparallel connection at the N terminus, with the try13,try22 style strands at the C-terminus. 19:00 try23 is having trouble cleaning up the strands---too many collisions and not enough change of shape. I should probably turn JiggleSubtree way down, as it is not doing much good. May be I need to drag things into better places---pulling T130-Q147 out away from the K105 strand and M1-5 out away from L56. Foo, the try23 constraints were inconsistent, with S43 having hydrogen bonds on BOTH sides. 22:09 The try24.constraints don't seem to be helping. I may have to break the offending pieces off and move them away to let the whole thing repack better. I may have to turn break down when doing that then, turn it up to heal the gaps. Try breaking off 1-7 and 122-end move D2 to be in line with L56 and I45 move D125 to be in line with K105, Y69, and G58 (that is, N about -0.1 -3.5 -19.3) Wed Sep 4 09:24:10 PDT 2002 Kevin Karplus Neither try24 nor try25 produced particularly convincing C termini. If I score without constraints, the best are try21-try20.6 and other try21 variants. It doesn't respect the secondary structure prediction for the last two strands, but not of my attempts to pack them have worked well. Perhaps I should do a polishing run for a try8-style arrangement, where the last two strands for a hairpin, but not worry about where the hairpin goes. 122> lgfDIEKGTKtv 144< gdyNAFVSIgn With these (try26a) constraints, the best score is for try12-try11-5.1.120, which actually has a different alignment: 122> lgfDIEKGTKtv 144< gdynAFVSign try12-try11-5.1.120 comes pretty close to making the alignment I want for the C-terminus, and has a decent N-terminus. With these (try26b) constraints, the best score is for try9.0.100, which I don't like nearly as much. Adding and modifying constraints to get more of the properties I liked in try12-try11-5.1.120 gives try26c.constraints. OK, now it comes out as teh best scoring with thos constraints---let's optimize with them Wed Sep 4 13:36:25 PDT 2002 try26-opt is looking a bit better. Let's do more polishing with the try26c constraints, but turn break up and clashes down to get a tighter packing. Wed Sep 4 15:05:42 PDT 2002 The try27.0.120 is not as much improved as I thought it would be---it is doing a lot of CloseGap and very little OptSubtree. As soon as try27 ends, I'll do another run with OptSubtree set higher. Wed Sep 4 17:54:44 PDT 2002 Well, try18 did a lot of OptSubtree and improved the score, but still did not close the critical gaps. I'm turing CloseGap back up for try29, and turning the break penalty up also. I'm not entirely happy with this prediction for t0169, but I'll submit it anyway when the run ends, since time is running out. model 1: try29-opt model 2: try14-opt model 3: try11-opt model 4: try8-opt model 5: try20-opt