Mon Jun 7 09:39:39 PDT 2004 T0196 DUE 2 July 2004 Mon Jun 7 12:03:13 PDT 2004 Kevin Karplus The first target seems to be a fairly simple comparative model. There are a lot of "Elongation Factor Tu" proteins in the t2k alignment (interestingly, the t04 alignment has very few proteins). The homology to 1exmA, 1efcA, ... is quite strong. The alignment is to the "Reductase/isomerase/elongation factor common domain" a "(6,10) barrel of circularly permuted topology". (SCOP domain b.43.3.1). The hard part will be packing the initial helix against the beta structure, as it is part of a different domain in the found homologs. The try1.under script keeps causing undertaker to crash. I don't know if this is a problem with the script or with undertaker. I fixed some minor bugs in undertaker so that it could handle an all-zero PDB file, but there are still crashes happening. Found the bug! The is_NO_Hbond() test was sometimes being called for out-of-range residues. I added range checks in is_NO_Hbond, and it seems to work ok now. Mon Jun 7 14:11:32 PDT 2004 Kevin Karplus try1-opt2 is done, and doesn't look too bad---the initial helix has been joined up and started to pack agains the barrel, but the N-terminal HIS tag has gotten in the way. The barrel is very similar to the barrels of the initial alignment, but with tiny tweaks for better packing. For try2, I've added some constraints: HelixConstraint S9 K16 0.9 StrandConstraint K48 E56 0.5 StrandConstraint K65 G68 1.1 StrandConstraint G74 E79 0.7 StrandConstraint I94 I98 1.0 // Try packing the initial helix against the beta unit: Constraint L15.CD2 G95.CA 0 3.6 8 0.5 // Maybe pick up some constraints from the mutual-info predictions: // These are not consistent, so should only be used with bonus_constraints // cost function. // From t2k-thin30 # Two of these three are reasonably consistent---I84 and I89 are on opposite # sides of the groove that I think the initial helix packs into. Constraint R17.CB E112.CB -10. 7.0 12.0 0.116958 Constraint R17.CB K84.CB -10. 7.0 12.0 0.110032 Constraint E19.CB I89.CB -10. 7.0 12.0 0.107884 // From t2k-thin62 # These form three distinct patches on the surface Constraint G8.CA E40.CB -10. 7.0 12.0 0.11769 Constraint G8.CA I52.CB -10. 7.0 12.0 0.11634 Constraint G8.CA S70.CB -10. 7.0 12.0 0.120141 Constraint G8.CA S97.CB -10. 7.0 12.0 0.106387 # These form a surface patch. Constraint H7.CB E40.CB -10. 7.0 12.0 0.114363 Constraint H7.CB E56.CB -10. 7.0 12.0 0.123808 Constraint H7.CB R93.CB -10. 7.0 12.0 0.121084 # but this is separate Constraint H7.CB E79.CB -10. 7.0 12.0 0.111146 When try2 is done, I'll clean up the constraint file, just keeping those constraints that look like they can be satisfied consistently. Mon Jun 7 15:24:13 PDT 2004 Kevin Karplus Try2-opt2 tried a different way of packing the N-terminal helix, but I didn't like it much---there seem to be two main choices, with L15 near G95 in both. I'll pick the one I think more likely and try building it for try3 (it is the one that try1 was approaching). I'm very pleased that undertaker is not blowing up beta sheets any more, and that the packing is looking quite reasonable. When I get a backbone I like, I may want to try letting Rosetta repack the residues (except for prolines). If try3 does an adequate job of placing the helix, I'll try picking some more specific interactions to constrain the helix, and reoptimize starting from try1, try2, and try3 (rather than starting over from scratch with alignments). Mon Jun 7 20:07:35 PDT 2004 Kevin Karplus The try4 optimization from tries 1-3 may have been a bit premature---it hardly moved try3's backbone at all, not packing the helix any tighter. The HIS tag moved out a bit, but did not straighten out. I should probably rerun from just the alignments, since the initial helix is much more movable when it is not yet attached. For try5, I'll go back to using the alignments (as in try3), but with the try5 constraints and slightly tweaked cost function (more weight on clashes and breaks). Mon Jun 7 23:18:25 PDT 2004 Kevin Karplus The try5 model has begun to lose Hbonds, so I should increase the weight of the beta and beta pair cost functions. I'll also try tweaking the constraints to see if I can get tighter packing of the initial helix. Tue Jun 8 00:07:12 PDT 2004 Try6 did not do even as well as try5. Let's try again with a smaller break and soft_clashes penalty, to allow freer movement initially. Tue Jun 8 01:12:42 PDT 2004 It looks like try7 will be a new best. If it looks good, I should try reoptimizing from tries 1-7, perhaps after tweaking the constraints. Tue Jun 8 07:18:00 PDT 2004 Oops, on try8 I tried to change 6 superiterations to 10, and because I was too tired, ended up with 106. At least this will be a good indication whether lots of superiterations on the second round is worth anything (I suspect it is just a waste of computer time). The range of costs is from 38.39939 to 39.45157, which is not a very wide range. The question is whether the extra diversity helps in the final pool--it may be another half an hour before the job finishes running. The try8 model looks pretty well packed, but there are some clefts that result from the helix not settling in tightly. Perhaps we could replace the constraints with different ones? L11.CD1 V43.CG2 perhaps I think it is time I released this target for others to play with. What I want to see in a good model: 1) The beta barrel is not disrupted. 2) The HIS tag at the beginning is extended, not helical. 3) The N-terminal helix is nicely packed against the barrel. 4) The predicted burial (seen with the "burial" script) shows that the right residues are buried. 5) There are no bad breaks or clashes. I think that one could play around with different constraints to pack the helix differently, trying to get it to settle in more tightly. Wed Jun 9 07:29:38 PDT 2004 Kevin Karplus try9 is just a short run to check out the rearrangement of the files to put the main parameters of the cost function and the constraints together in try9.costfn Wed Jun 9 10:35:15 PDT 2004 Kevin Karplus try9 is not much different from try8. We may need some new constraints and rerunning from the initial alignments to get out of the local minimum. I think that the helix needs to settle deeper into the groove on the beta barrel. The sidechain of R93 needs to move out of the way a bit and F14 needs to move closer to I52. Mon Jun 14 15:39:18 PDT 2004 Katzman, Draper, Koeva Included the t2k.undertaker-align alignments and use TryAllAlign to generate new conformations. Try changing the constraint of I26 to a point deeper in the groove at V37 hoping this will help rotate the helix. // Try packing the initial helix against the beta unit: constraint I26.CD1 V37.CA 0 5 10 1.0 Increased the weight of constraints to 30 in SetCost. Tue Jun 15 11:32:17 PDT 2004 Katzman, Draper, Koeva There was virtually no change in the try10 result compared to try9. There was an additional break introduced at G47. -------------------------------------------------- From karplus@soe.ucsc.edu Tue Jun 15 09:55:47 2004 Date: Tue, 15 Jun 2004 09:55:44 -0700 From: Kevin Karplus To: katzman@ieee.org Cc: karplus@soe.ucsc.edu, sol@soe.ucsc.edu, ggshack@soe.ucsc.edu, learithe@soe.ucsc.edu, martina@soe.ucsc.edu, bbarnes@ucsc.edu In-reply-to: <5.1.1.6.1.20040615091651.00b13150@pop.sbcglobal.yahoo.com> (message from Sol Katzman on Tue, 15 Jun 2004 09:23:20 -0700) Subject: Re: T0196 pointers Sol asked > We could use some pointers on how to make progress with T0196. > > We spent time yesterday understanding the process and we know WHAT > we want to do, but were not very successful in figuring out HOW > to do it. You can look at try10.under to see what we did. Looking > at the results, it looks like there was virtually no change from the > try9 pdb output. When you start with existing models using ReadConformPDB, the result is usually to do minor polishing to the best of the previous models. They are already highly optimized, so are trapped in a local minimum in conformation space. You can get changes if the cost function is different enough (totally different constraints, for example) so that the existing models are far from optimal. What I usually do to get major movement is to start over from the alignments (the "TryAllAlign" commands) without any ReadConformPDB. This usually gets new starting points from which better models can be built. > For example, we wondered how to uncoil the last few residues of > the helix to give us more flexibility to place the hydrophobic > portion of it next to the hydrophobic portion of the sheet. It is generally not a good idea to unwind helices that are predicted to be helix---those predictions are usually right. (Exception, histidine tags are usually disordered in crystal structure or extended, but our networks predict them to be helical.) Since the initial alignments have breaks at the end of the helix in T0196, there is lots of flexibility if you start from the alignments. Increasing "super_iter" on the first round of optimization also increases the diversity of starting points. > And we did not have a good idea on how to really accomplish the > constraint that would bring I26 next to V37. Given that the helix is initially not attached in the alignments, putting in constraints should get it to move more or less where you say. You can decrease the break weight weight if the program seems to be more interested in attaching it than in placing it in a good place. Or decrease soft_clashes if it seems to find the groove too small. You might want to try increasing the initial probability of applying OptSubtree and OptSegment operators, which try to use the constraints (and other pairwise terms) to move pieces around. -------------------------------------------------- Tue Jun 15 11:35:36 PDT 2004 Katzman, Draper, Koeva For try11, we went back to the original t2k alignments, without reading in the confomations from the first 10 tries. We kept the restricted G8-R17 helix constraint and the strand constraints, and focused on the constraints from the helix (L11,F12,L15,I26) to the strands (G95,I94,G53,V37) trying to pack the helix into the cleft of the strands. We also decreased the break weight and increased the super-iter as suggested in Kevin's email above. The try11 results got the shortened helix closer to the strands, but failed to make most of the strands. And in fact it made the strand at the C-terminus into a helix. For try12, use the StrandConstraints from t04.dssp-ehl2.constraints. rather than the smaller set from t2k, while leaving the single short HelixConstraint. Not much different from try11. ------------------------------------------------------------------ Wed June 16 4:30pm Jenny Draper I used DeepView's backbone-rotation capabilities to twist the L29-P33 loop into a conformation that brings the helix packed in approximately where we want it; I did this by rotating the Phi-Psi angles of K31 & P33 using DeepView's ramachandran-plot "drag the spot to a new angle space" method. Obviously, this sucker is full of clashes; but it might help us as a starting point for getting the helix where we want without killing the barrel? It's called: decoys/T0196.try9-opt2-jen-rot1.pdb Wed Jun 16 16:41:06 PDT 2004 Sol Katzman Used Jenny's pdb as a starting conformation (without any other conformations) for try13 Try13 looks better than try9 did. The portion of the helix between L11-F14 is packed into the sheet pretty well. ------------------------------ From karplus@soe.ucsc.edu Fri Jun 18 22:14:38 2004 Date: Fri, 18 Jun 2004 22:14:35 -0700 From: Kevin Karplus To: karplus@soe.ucsc.edu, sol@soe.ucsc.edu, ggshack@soe.ucsc.edu, learithe@soe.ucsc.edu, martina@soe.ucsc.edu, bbarnes@ucsc.edu Subject: T0196 Looking at the README file and the most recent models for T0196---I think that we may be trying too hard to force the helix into the groove. We might want to try starting from try13-opt2 and running an optimization with higher weights for dry5, dry6.5, and dry8 (to try to get somewhat tighter packing), and reduce the weight of constraints that are trying to get the helix to pack. The loop from I28 to P33 is too far from the rest of the model. You can also make a different attempt at sidechain packing by making decoys/T0196.try13-opt2.repack-nonPC.pdb, which will use a version of Rosetta to repack the sidechains. Sometimes Rosetta does a better job, sometimes not (Rosetta and undertaker have somewhat different views on what constitutes a clash---I spent quite a bit of time this spring comparing the two clash checkers and finding bugs in both). You may need to turn up the weights of soft_clashes and break to start polishing the model. When we submit this, we'll probably want a model where we've forced the helix where we think it belongs, and a different one where undertaker did the optimization without so much forcing. We may want one with undertaker's sidechains and one with Rosetta sidechains. We can also include a model straight from the alignments without much tweaking. I'll add these comments to the README file, but leave the implementation to Sol (or someone Sol agrees to have do it). Kevin ------------------------------------------------------------ Mon Jun 21 16:48:51 PDT 2004 Sol Katzman For try14, as Kevin suggested, decreased weight of break (45 -> 30) and increased dry weights: dry5 (5 -> 10) dry6.5 (6 -> 12) dry8 (5 -> 10) I also created T0196.try13-opt2.repack-nonPC.pdb but cannot see much difference between it and try13-opt2.pdb. Tue Jun 22 15:03:00 PDT 2004 Sol Katzman I made a mistake in try14. I was supposed to decrease the weight of the constraints, not the weight of break. Looking at the loop I28-K32 shows that in try9-opt2 there was a kink at S30 that brought it closer in to the rest of the model. In Jenny's hand modified version try9-opt2-jen-rot1 this kink was properly maintained, but a conflict developed between S30 in the loop and S57 in the interior. The undertaker output after that blew those two serines apart, even though the neighboring K31,K32 still pointed outward, as did I28,L29. For try 15, I will decrease the weight on the constraints (general constraint weight 30 -> 15, as well as reducing the individual atoms constraints that we used to pack the helix into the groove.) I will also read in the try9-opt2, try9-opt2-jen-rot1 models, since they are a better starting point than the opt13 model. Wed Jun 23 08:34:13 PDT 2004 Sol Katzman In try15 there is still large gap between S30 and S57. (10A) For try16, add a constraint to keep those Serines together. Also, try a Rosetta repack of try9-opt2-jen-rot1 as an input conformation. Wed Jun 23 16:55:58 PDT 2004 Kevin Karplus I wanted to look at the progress on this model, so I made an "unconstrained.costfcn" which has no constraints except the extended HIS tag. Using this in score-all.under, I see that there is a non-readable pdb file in the decoys, which causes undertaker to stop reading conformations. I renamed it to have "pbd-pc" as its extension, so that read-pdb.under wouldn't include it. I'm not sure why undertaker choked on it---perhaps I need to add more error checking, but the PDB input is read by libpdb, which is not my code. >> Sol: This was the file created by Jenny using DeepView on a PC >> The problem is PC vs. Unix newlines. The file was converted >> and the PC version has now been deleted. The best current model without constraints is try14-opt2, followed by the try16 and try15 ones. After that are try9,8,10,13,1. It might be worth doing a polishing try without constraints (copy unconstrained.costfcn to a tryXX.costfn file) starting from all decoys. We should probably submit models with different helix positioning---for example try14, try9, and try1, plus one or two models from the alignments. In try14, I'm a bit worried about the buried D49---it should either be exposed or a salt-brdige formed, perhaps with K101 or K104. Similarly I think E56 should be exposed or salt-bridged to K36. Increassing the weight of wet6.5 may help expose these---salt bridges can be added as Hbond commands. There are sitll some bad clashes in try14-opt2, so it might be worth doing Rosetta sidechain repacking for the top few models before doing a polishing run. Wed Jun 23 19:28:14 PDT 2004 Sol Katzman The result of try16 did pull S30 and S57 together. For try17 do a polishing run: a) repack-nonPC of opt2: try14,try15,try16,try9 b) include opt1 and opt2 and opt2-repack of the four above c) include opt2 of try8,try10,try13,try2 d) use unconstrained.costfcn as try17.costfcn Fri Jun 25 09:56:33 PDT 2004 Sol Katzman In try17 the serines S30 and S57 got blown apart again. The best scoring models using unconstrained.costfcn are now: try17-opt2 (26 clashes) try17-opt1 (31 clashes) try14-opt2 (42 clashes) try16-opt2 (37 clashes) try16-opt1 (40 clashes) try15-opt2 (49 clashes) Sat Jun 26 08:43:48 PDT 2004 Kevin Karplus I think that we should hedge our bets on this one. Take our best-scoring unconstrained model (try17-opt2) as our first guess, take a couple with other guesses of helices packed against the barrel in different places, and at least one with the helix extended (perhaps one of the original alignments?) I've set up a possible set in superimpose-best.under, selected mainly on the basis of the unconstrained.costfcn (except for the models from alignments). InFilePrefix decoys/ ReadConformPDB T0196.try17-opt2.pdb ReadConformPDB T0196.try9-opt2.pdb ReadConformPDB T0196.try1-opt2.pdb InFilePrefix ReadConformPDB T0196.t2k.undertaker-align.pdb model 1 ReadConformPDB T0196.t2k.undertaker-align.pdb model 2 Thu Jul 1 21:29:50 PDT 2004 Kevin Karplus I did one more polishing run (try18) with soft_clashes and breaks turned up a bit. This managed to get the cost down a little bit. The changes are very minor ones to the rotamers---things pack just a tiny bit better. The loop that makes a hole (L29-P33) was not moved. I'm submitting try18-opt2 as our first model and declaring this target completed. Models submitted 1 T0196.try18-opt2.pdb 2 T0196.try9-opt2.pdb 3 T0196.try1-opt2.pdb 4 T0196.t2k.undertaker-align.pdb model 1 =T0196-1exmA-local-str2+CB_burial_14_7-0.4+0.4-adpstyle5.a2m 5 T0196.t2k.undertaker-align.pdb model 2 =T0196-1efcA-local-str2+CB_burial_14_7-0.4+0.4-adpstyle5.a2m Fri Sep 24 12:19:37 PDT 2004 Kevin Karplus I added 1xe1A as the correct answer and produced GDT plots. We did quite well (beating robetta), but did not submit our best model. name length missing_atoms rmsd rmsd_ca GDT smooth_GDT T0196.try7-opt2.pdb.gz 116 0.0000 5.8876 5.9142 -71.1538 -66.8193 model3.ts-submitted 116 0.0000 6.2945 6.4412 -70.8791 -66.6645 full-auto model5.ts-submitted 116 47.0000 4.4155 4.2927 -68.9560 -65.7084 model4.ts-submitted 116 42.0000 4.3970 4.2788 -70.8791 -65.6431 model1.ts-submitted 116 0.0000 6.3086 6.3700 -68.9560 -64.6214 robetta-model1.pdb.gz 116 0.0000 6.3613 6.4689 -69.5055 -64.5145 model2.ts-submitted 116 0.0000 6.5861 6.6872 -67.5824 -63.9845 robetta-model3.pdb.gz 116 0.0000 6.3203 6.4890 -68.1319 -63.8931 robetta-model5.pdb.gz 116 0.0000 6.1616 6.2212 -66.2088 -62.6192 robetta-model4.pdb.gz 116 0.0000 6.5595 6.7042 -64.2857 -61.3960 robetta-model2.pdb.gz 116 0.0000 5.9223 5.6426 -64.2857 -61.1893 Indeed our full-auto and alignments were better than our model1 or model2.