Mon Jun 24 09:16:05 PDT 2002 25 June 2002 Kevin Karplus The T0139.try1-opt.pdb seems to be a rather random assortment of helices, fairly compactly placed, but not matching the predicted secondary structure all that well. The best-scoring (1iuaA) alignments are just for 261-286. They have a very short helix-coil-helix-coil-helix region, but are far from a complete fold. Also, some of the coil region here is predicted to be helix and some is predicted to be strand---not a good 2ry match! The next best (1jx6A) is also quite fragmentary, and has some predicted helices stretched out into straight lines. We'll probably have to improve the cost function for undertaker and hope for an ab initio result---perhaps without seeding the alignments. 25 June 2002 Kevin Karplus The result of try2, which started from a random conformation, is an all-helical protein that does not particularly respect the 2ry-structure predictions. We may have to use a scoring function with predicted alpha values, which I haven't implemented yet. 27 June 2002 Kevin Karplus CAFASP consensus is very weak. It seems that no one has a clue on this protein. 12 July 2002 Kevin Karplus Using the new fragments, try3-opt still gets an all-helical protein. This actually may not be as bad as I thought, since the str network and the combined dssp_ehl2 prediction does seem to favor helices throughout. There is still a region (approx K248-Q258) that is winding into a helix and probably shouldn't be. 24 July 2002 Kevin Karplus I edited define-scores.script and undertaker.script to use the new pred_alpha2 cost function, to scwrl conformations after each iteration, and to save Template.atoms to speed up subsequent undertaker runs. It looks like 1gdoA and 1gdoB are the popular templates for this target, though the alignments only go from S257 to R291. Try4-opt-scwrl is the new best scorer and does have some of the same problems with secondary-structure prediction as some of the earlier bests. Perhaps I should do another run with no alignment insertion---just fragments and specific fragments. This might not make any difference, since the decoys on try4 were almost all free of breaks. I'll also run the optimization longer, to sample more of space. In try5-opt (the new best score), we can see the helices bending where they should, but the long helices not staying straight. Perhaps I should add some constraints to keep the helices straight. The strongly-predicted helices are S8-R22 (or A20), E59-Q77 (oops---that's in the sequence numbering, not the PDB numbering, which would be S232-A244, E283-Q301) Let's add constraints for CA S232-A244 and E283-Q301. With the new constraints try5-opt is still the best scorer, so let's start optimizing from there. 24 July 2002 Kevin Karplus Now try6-opt is the best, but it is still folding up the helices. Let's try turning off radius and the larger-diameter burial terms. Redoing the score function still left try6-opt on top. Let's do another run, not seeding the conformation with anything, using the new score function. 25 July 2002 Kevin Karplus The try7 conformations have the helices straight, but the turnaround is not so good. Perhaps the "contact_order" term is encouraging the two to run antiparallel too much---hmm, seems unlikely since contact order is not particularly positively correlated with cost. The pred_alpha2 cost seems to be a strong correlation with overall cost, and the clash functions are even stronger. Dry6.5 has a strong correlation with overall score also, as expectd from the weighting used. The new "JiggleSubtree" operator is useless on this target, as we are not introducing any breaks. Let's try again, removing contact_order from the score function. The best now is try8-opt, which has NO distant contacts. Instead it has the two main helices almost anti-parallel a long way apart (separated by the middle two helices). NOT a likely conformation. HOW ARE WE GOING TO GUESS HOW THIS PACKS? Target t0139 has only a very close homolog (mouse from human) in the t2k alignment, so all the searches are rather limited in their ability to find distant homologs, and the 2ry prediction is likely to be weak. According to Swissprot and PDB, the N-terminal domain of DNA Fragmentation Factor 40 (DFF) is in 1ibx---not much help since t0139 is the C-terminal domain of DFF (225 on, after the second CASPASE cleavage site). Swissprot does report some alternate splicing and sequencing variants: 262-331 LVTKEDP.... shrunk to VGGNQGM (this amounts to deletion of the C-terminal helix) 291 R->W The mouse DFF shows similar alternate splicing to delete the final helix. 30 July 2002 Kevin Karplus I put contact_order back in, but with a lower weight than before (0.1 instead of 1), aadded phobic_fit with a small weight, and added dry5, to make packing more favored for hydrophobic residues. Try8-opt-scwrl scores best. There are plenty of hydrophobic residues, so doing another run from the beginning might try to pack them more. The new versions of JiggleSubtree and OptSubtree (which will break between arbitrary atoms) may be of some use. If nothing reasonable comes out of try9, then I'll need to add some constraints, either for guesses about how the helices pack, or arbitrarily between several of the hydrophobics, to try to get some collapse. Tue Jul 30 16:17:00 PDT 2002 Kevin Karplus try9-opt is the new best scoring. It has not respected the 2ry predictions much, nor the constraints that tried to keep the helices straight. Wed Jul 31 09:33:00 PDT 2002 try10-opt is the new best. This is the closest to something reasonable so far. The helices are bit bent, but the packing is starting to make some sense. Let't restrart from there (rather than completely from scratch), with heavy weight on fragment insertion. 1 Aug 2002 Kevin Karplus hmm, for some reason, try10-opt scores BETTER than try11-opt. oops---it looks like I started from a random conformation, doing alignment and fragment inserstion, without using try10-opt as a starting point. I should run again using try10-opt and try11-opt as starting points. I think I like the looks of try11-opt better. Hmm---playing with the weights of the scoring function makes try8, or try10 look best. I should probably do a run starting from try8, try10 and try11, with a high probability of crossover, to see what comes out. 1 Aug 2002 Kevin Karplus The score function seems to favor the open conformation of try8 too much, since the iterations of try12 so far look a lot like try8. I'll have to play with parameters until try 10 (or better, try11) are favored, then reoptimize once more. Fri Aug 2 04:19:23 PDT 2002 kevin Karplus try12 does indeed favor a fairly open conformation, though the two long helices have come together at one end. Increasing the weight for the more global burial terms (dry8, dry12, and phobic_fit) still leaves try12 on top, probably because there are no breaks and the small radius terms are fairly well satisfied. I'll do an optimization run from try10 and try11 (NOT try8 or try12) to see if cleaning them up will result in a low-scoring structure. (Yes, on try13, even the first iteration does better than try12-opt.) We can be pretty sure that the open conformation is wrong, since the NMR spectroscopists claim that "it is most likely to be monomeric". Fri Aug 2 09:42:43 PDT 2002 Kevin Karplus try13 is getting some halfway decent structures, but is not exploring enough of the space---it uses crossover far too much, whigh tends to homogenize the pool rather quickly. OptSubtree and InsertFragment ARE getting used a fair amount, so things are not totally hopeless. When try13 finishes, I should probably do another run with crossover turned way down. Fri Aug 2 11:25:58 PDT 2002 Kevin Karplus try13 finished, and try13-opt is the new best scorer. Let's try adding weak constraints to bring V285 near L245 and L295 near V235. These might help drive the folding. We can strengthen the helix-straightening constraints a bit also. The spacing is such that we would expect them to be on the same side of the helices. Constraint 75 538 1 3.5 12 // V235.CB L295.CB Constraint 144 460 1 3.5 12 // L245.CB V285.CB Increasing the weight of the constraints doesn't change the lowest-scorer, as try13-opt already does better than most at these constraints. T0139.try7.6.40.pdb gets the constraints much better, but at the expense of not packing the small helices at all well. Let's do another run with NO intitial conformations but the extra constraints. 2 Aug 2002 Kevin Karplus I rather like the way T0139.try14+1bl8A-T0139-fssp-g...+T0139-1h2rS-2track-local-adpstyle5.pw.a2m.gz:1h2rS.19.80.pdb looks, though it has covered the hydrophobic patches in a different way. If this doesn't for the basis for the final superiteration, I should probably go through and look at each intermediate, selecting out interesting ones and guessing some constraints based on them. 3 Aug 2002 Kevin Karplus Try14 does not score nearly as well as try13, probably because it is looking for a different packing than the constraints are. It is very frustrating to use Rasmac to look at the molecules, because the redraw is so slow that it doesn't rotate molecules smoothly. I'm having a hard time figuring out whether I like the try13 or try14 run better and what constraints to try next. 8 aug 2002 Kevin Karplus Let's remove the packing constraints and rescore. try13-opt still scores best, mainly because of pred_alpha2. I'm not sure what to do next---try more examples from random start to see what comes out? guess the packing and constrain? (Incidentally, rasmol 2.7.1 is plenty fast on the Mac---it is only the UCB version of rasmol that is painfully slow.) Let's guess that L242 and L293 pack against each other (a guess from T0139.try14+1bl8A-T0139-fssp-g...+T0139-1h2rS-2track-local-adpstyle5.pw.a2m.gz:1h2rS.19.80.pdb). try13-opt still scores best, but let's try again from random start. 10 Aug 2002 Kevin Karplus Even with a random start (try15) try13-opt scores best. None of the packings seems particularly tight yet. Should we do another random start, or try optimizing the existing ones? At the moment, our best bet seems to be to submit multiple models, starting with try13, try15, try14. Tweaking the parameters and removing the guessed constraint for hydorphobic contact makes the order different: try13, try12, try10, try15, try9, try11, try14. Let's try again optimizing from nothing with the new score function. 11 Aug 2002 Kevin Karplus Best scores from different runs are now try13-opt, try12-opt, try16-opt, try10-opt, try15-opt Tweaking the parameters to favor packing changes the order to try12-opt, try13-opt, try16-opt, try15-opt, try10-opt-scwrl Hmm, this must NOT be favoring packing, since try12 just brings the N and C terminus near each other, and does not do anything else to bury hydrophobics. I tweaked the parameters until try12-opt dropped behind try16-opt. Order is now try13-opt, try16-opt, try12-opt, try15-opt, try14-opt Tweaking more, to try to make try16 look better, I get the order try13-opt, try16-opt, try14-opt, try15-opt, try12-opt Let's do a run starting from try16, 14, 15, 12, with CrossAndInsert set high, with this new scoring function. The result should be a better packed version of try16. 11 Aug 2002 Kevin Karplus Reran "make" with the new template library---no change among the top hits. When try17 is done, may want to do another run from scratch with the new fragment library. 12 Aug 2002 Kevin Karplus Try17-opt is new best scorer, but is still pretty foamy and has exposed hydrophobics. 12 Aug 2002 Kevin Karplus try18-opt is still rather foamy and has bad breaks, but buries hydrophobics better than try17-opt (though try17-opt score better by a little). Let's start from try17-opt and try18-opt with a bigger break penalty. 13 aug 2002 Kevin Karplus try19-opt is new best. It seems to have been based on try17-opt, but it is starting to look pretty good. I'll reoptimize once more, with slightly higher break penalty, then submit the best few models from different runs (deadline is Thursday 15 Aug). After tweaking the cost function try19-try17.24.80.pdb scores best, though it doesn't look quite as good to me. Hmm--I probably gave too much weight to gen6.5 and wet6.5, which tend to unpack things. Tweaking weights again moves try19-opt back to being the best, though try19-try17.24.80 is second best. 14 Aug 2002 Kevin Karplus Best-scoring is now try20-opt. I 'stole' the robetta models from the CAFASP site, but they do not score at all well. I looked at robetta1.pdb, and it is VERY loose. try20-opt looks ok except for some small breaks. Increasing the break penalty makes T0139.try20-try19-25.3.90.pdb score best. Let's try reoptimizing. From: Rachel Karchin Date: Wed, 14 Aug 2002 23:32:14 -0700 To: karplus@soe.ucsc.edu CC: rachelk@soe.ucsc.edu, yael@biology.ucsc.edu, learithe@soe.ucsc.edu Subject: T0139 I found a paper about this protein. JMB paper dated Aug. 9, 2002. (new!) http://www.ncbi.nlm.nih.gov:80/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=12144788&dopt=Abstract Title: Solution Structure of the DFF-C Domain of DFF45/ICAD. A Structural Basis for the Regulation of Apoptotic DNA Fragmentation. Abstract: DFF45/ICAD has dual functions in the final stage of apoptosis, by acting as both a folding chaperone and a DNase inhibitor of DFF40/CAD. Here, we present the solution structure of the C-terminal domain of DFF45, which is essential for its chaperone-like activity. The structure of this domain (DFF-C) consists of four alpha helices, which are folded in a novel helix-packing arrangement. The 3D structure reveals a large cluster of negatively charged residues on the molecular surface of DFF-C. This observation suggests that charge complementation plays an important role in the interaction of DFF-C with the positively charged catalytic domain of DFF40, and thus for the chaperone activity of DFF45. The structure of DFF-C also provides a rationale for the loss of the chaperone activity in DFF35, a short isoform of DFF45. Indeed, in DFF35, the amino acid sequence is truncated in the middle of the second alpha helix constituting the structure of DFF-C, and thus both the hydrophobic core and the cluster of negative charges are disrupted. try18, try19, try20 and try21 all look pretty good to me. try18 looks a bit better than try21 which has kind of a hole in the middle. The structures definitely have the four packed helices and several patches of negatively charged residues on the surface. But none of them have a single large cluster of negative charge as described in the abstract . . . -- rk 15 Aug 2002 Kevin Karplus Wow! The pictures in JMB 311:317-327 (2002) should really help us get the solution. i wonder if they'll withdraw this from CASP. Let me try to get a bunch of constraints from the pictures. True helices: S239-E247, S257-K265, P268-L275 or N276, K280-T300. (The third helix is presented differently in the Figures 1 and 5). Based on Figures 4 and 6c, the second and 4th helices pack against each other. It looks like the closest contact is approx T264 to C289. From Figure 6A: The negative patch is centered on E290 and consists of D278, E283, E287, E290, E261, E267, E266. E283-E290 is 2 turns of a helix, so approx 10.5 Ang. E290-E267 looks like the same distance, as does E266-E283 and E261-E267, and E283 is closer to E267 (maybe 6Ang) than to E266. E290-E261 also looks like about 6Ang. K281 seems to be about equidistant from E247, D278, and K248 (so about 4 Ang?). D259 seems to be 4-5 Ang from H240. E292 seems very close to E252 (3Ang?) and R296 seems to be about 10Ang from E252. From Figure 6B I279 is about 4 Ang from P268. L242 touches A270 and is about 4Ang from V273. I241 touches L262. L260 touches L293 and comes close to C289. L262 is about 4 Ang fro A244 and L255. A250 is about 5Ang from V288 and L253 is about 4 Ang from L295. W277 is about 6 Ang from A244. From 6A and 6B R246 is near V273 (4Ang), L262 is between D 259 and H240, E247 is near W277, K248 is near L255, P268 is sandwiched between I267 and E283, I 241 contacts K265 and E266, R296 is near L253. There are a lot of problems with trying to guess constraints from the pictures. 1) atoms aren't labeled, and only a few residues are. 2) some residues are mislabeled (there are no residues E267 and V288, shown in 6a and 6b) 3) distances are hard to guess. I've thrown in a lot of constraints, in the hope that the mistakes will tend to average out. After optimizing with the constraints, we should probably reoptimize with the constraints turned down, to pack without conflicting constraints. With the constraints in place, the best scoring decoys are try21.0.90 and try21-opt-scwrl, but neither one is great. Let's reoptimize from them with a short protocol (try22), then turn down the constraints and reoptimize. Simultaneously, let's try an optimization run from scratch, and see if we can get to a reasonable solution faster that way (try23---nope, it seems to take a long time and not get good scores). Hmm, try22 doesn't seem to be doing much, probably because the OptSubtree and JiggleSubtree operators are called too rarely. Let's try another short run with them turned up (try24---this seems to be getting better faster than try22). Date: Thu, 15 Aug 2002 08:28:52 -0700 (PDT) From: yael@biology.ucsc.edu To: Kevin Karplus cc: rachelk@soe.ucsc.edu, , , , , , , Subject: t0139 In-Reply-To: <200208141118.g7EBIdm23618@bray.cse.ucsc.edu> Thanks Rachel for finding the paper, I could not down load it from home but based on the anstract it seems that most of our decoys agree with the 4 helix arrangment though non of them have a one large negative cluster. I would suggest to add to one of (try18,19,20,21) try14 and try3. Yael 15 Aug 2002 Kevin Karplus With the jmb.constraints and reoptimization, the best score is T0139.try24.7.40.pdb (NOT try24-opt-scwrl). Rats I don't like try24-opt, try24.7.40, nor try23-opt. The try24 runs suffer from a handedness problem, with helix4 on the wrong side of the helix1-helix2-helix3 cluster. We could try hand-editing try24.7.40 to add (-12, 12, 3) to all residues starting at L275, then reoptimize with OptSubtree turned way up. Another idea---try all of try3,14,18-24 as starting points, turn constraints way up and reoptimize with high OptSubtree and CrossAndInsert probability. Let me start that, since it requires little thought. OK, I started a new run (try25) in which I tightened the constraints on the final helix, increased the weight of the constraints, and started from many initial conformations. I'm only doing a few generations and superiterations, so that we can look at the results soon---the target is due by this evening. Date: Thu, 15 Aug 2002 10:47:49 -0700 (PDT) From: Jenny Draper To: yael@biology.ucsc.edu cc: Kevin Karplus , rachelk@soe.ucsc.edu, jcasper@soe.ucsc.edu Subject: Re: t0139 Sounds good. I don't think I could add much. Try 22 seems to close the hole in 20/21 a little better. I don't like trys 23 & 24. -Jenny Thu Aug 15 10:58:34 PDT 2002 Kevin Karplus try25-opt made piddling improvements, not moving the helix the way I wanted it to. I think that the reason is that the helix is on the WRONG side, and OptSubtree does an interpolation to try to figure out where to resolve constraints---if something is on the wrong side the interpolation will tend to cause a lot of clashes and so not produce a decent score. Perhaps I should sometimes extrapolate AWAY from the current position in OptSubtree? 11:43 Kevin Karplus No, try26-opt is not succeeding in reducing the constraints much either. I've made a different change to OptSubtree that sometimes makes a large random transform of the subtree before doing the optimization---this should help unstick jammed-in-the-wrong-place problems. 12:33 Kevin Karplus Nope---try27 scores better, but still hasn't put the helix on the right side. Either my constraints are all screwy, or OptSubtree isn't able to move the helix. Let's do one more try with the constraints cranked WAY up, while I try to make a script to move the helix to the right side manually. try28 tries again from try21-opt. I wrote a crude script "move-residues.perl" to move the helix in try24.7.40 (nnot because I liked try24.7.40 much, but because I had looked at it to determine a possible initial offset). The result of the move is in T0139-moved.pdb. try29 tries again from T0139-moved.pdb. Hmm---the new OptSubtree operator seems to be fracturing the chain into tiny pieces that it moves wherever---satisfying some of the constraints but shredding the protein into nonsense. This may be because I turned the constraints up way too high. Let's try again (from T0139-moved.pdb) with break penalties up again and constraints down. try30 started from T0139-moved but without the huge weight on constraints seems to be getting into the ballpark. When that run finishes I'll start again from there with more normal weights on the constraints and breaks, to see if we can fix up the packing. 2:29 Kevin Karplus try30-opt is roughly right. Let's turn down constraints and turn up breaks and reoptimize it. I've created a "true" rasmol script that uses the 2ry labeling from the paper, coloring the disordered residues as "turn". 16:00 Kevin Karplus try31-opt is looking pretty good--still a bit too much space, but I'll try another run with some extra constraints (now jmb3.constraints) to try to close things up. There may not be time for another run after that---I'll probably submit the results of that run and replace them if I can get another run to finish. 16:32 Kevin Karplus With the new scoring function, try32 intermediates are finally overtaking the try24 ones. 17:36 Kevin Karplus Try32 is new best. Let's now turn off all constraints except helix straightening. With helix-only constraints, best are try32.1.40, try32.9.40, ... . Since some of the constraints may be incorrectly read from the picture, let's try jiggling some of these to close the remaining hole. It may be difficult because the C terminus has folded back and may prevent some desirable movement. Perhaps I should add a constraint to move R307 out 14 Ang from T300. With the JMB constraints removed, the try24 models score better than the try32 ones, but I'll start run try33 from just try31 and try32 models. It'll probably be 20:00 before this run finished, so I'll submit try32.9.40, just in case. 19:55 Kevin Karplus try33-opt-scwrl is new best score in the "right topology" series, but it has gotten V273 stuck in the gap holding E267 and E283 apart. Perhaps we should try another quick run with the jmb3 constraints. With the jmb3 constraints, the top score with the right topology is try33-try32-9.9.40.pdb, with try33-opt as second best. 21:08 Kevin Karplus Getting down to the wire here, and I don't see much change in try34-opt. Let's submit the best scoring one with jmb constraints: try34-opt and 2nd the best-scoring one (in the try30+ series) with helix-only:try34-opt-scwrl They're almost identical, but we have to go a long way down the scores to get a different one, and almost all the try30+ series are quite similar. They're not the best we could with the data in the article, but the best we could do in 24 hours with the data.