Mon Jun 28 09:40:44 PDT 2004 T0218 Due 13 aug maybe fold-recognition? maybe new fold. Mon Jun 28 13:50:32 PDT 2004 Kevin Karplus The alignment to 1ixvA is causing scwrl to run a LONG time (over 40 cpu minutes) causing the show-align.under script to stall. Mon Jun 28 14:39:11 PDT 2004 Kevin Karplus scwrl up to 53 minutes, I'm going to kill it. Thu Jul 29 18:03:38 PDT 2004 Martina Koeva Based on the t2k and t04 predictions from str2 and dssp-ehl2, I have decided to define 8 strands and 7 helices for now: s1: Q44-V48 s2: K51-T59 s3: G74-W78 s4: L118-V121 s5: I152-V155 s6: I173-N178 s7: K214-L221 s8: G229-E234 and h1: A8-E17 h2: G66-V73 h3: I132-M148 h4: K186-F194 h5: P244-T256 h6: I276-Y286 h7: K290-Y300 and have specified them in try2 as strand and helix constraints. I have also included the 280.rr.constraints, have rescaled the hbond parameters and have increased the constraints weight. Last night I ran VAST on this very scattered structure and it came up with nothing. There was only one hit, which was more or less trash. VAST ID: VS60238 Password: T0218try1 There seems to be more or less consensus in the top hits between c.62.1.1 and c.55.3.4. One thing that I like about those folds is that they both indicate 3 layers of secondary structures: a/b/a. Looking at the burial predictions from both CB_burial_14-7 and near, it appears that the strands (given that they form a sheet) want to be buried on both sides, which fits in with the template structure. I looked at all 10 Robetta models and there are quite a few contradictions, where Robetta folds helices into what we predict as strands and vice versa. One, I thought, common pattern between the models is that Robetta tries to pull the strands and the helices into more or less separate domains. So if one were to look at the molecule, half of it would be all strands in 1 sheet, and the other half would be only helices. Given our burial predictions, I do not believe the Robetta models very much. On the other hand, some of the strands it has gotten right and it has paired them into a sheet. If I do not find a successful way of building a sheet, I might try and get a hint from the way Robetta has built its sheet. Thu Aug 5 15:25:51 PDT 2004 Martina Koeva Started try3 with the constraints from dssp-ehl2 from t04. In an attempt to get a sheet forming, I used the try1-opt2 sheet constraint, as well as an additional sheet constraint that I liked in the Robetta model. Sat Aug 7 13:39:32 PDT 2004 Kevin Karplus I've not looked at the predictions here yet, but I'm creating a t04.many.frag fragment library, since there isn't one already. This should be useful in subsequent runs, to give a larger set of fragments to use in modeling. After I've downloaded the files and looked at the models, I'll add more comments to the README file. Sat Aug 7 14:13:04 PDT 2004 Kevin Karplus None of try1, try2, or try3 look like a protein---we are not forming the necessary beta sheets. Robetta model1 has a halfway decent structure for E26-L87. We may want to copy sheet constraints from that. I created the robetta*.sheets and robetta*.helices files in decoys, using the superimpose-best.under script. If there are any constraints worth picking up, they can be copied from the appropriate files. It's too bad that the rr constraints are so weak on this protein---we really could use some help. Since I don't have any great ideas what to do next on this, I'll leave it for Martina to play with. Sun Aug 8 17:17:47 PDT 2004 Martina Koeva I added the sheet constraints from the Robetta model 1 for try4. My plan right now is to attempt to make a sheet of strands and sandwich the sheet between 2 layers of helices (a layer on each side). Mon Aug 9 16:27:25 PDT 2004 Martina Koeva Try4 and try5 are basically the same, except that i forgot to put weights on the sheet constraints in try 4. However, there is no sheet formation in try5-opt2 at all. Even the strands are not forming into strands correctly. Some of them are still curling into helices. For try6 I am changing the cost function quite a bit in terms of weights. I have raised all of the dry5,6.5 and dry12, as well as wet6.5 weights. I have turned down the weight on sidechain, but raised phobic_fit. I have also increased the break weight. I have left the hbond_geom_beta_pair at 15 (the weight that it was changed to for try5). I have also increased by a factor of 10 the strand and helix constraints. Finally in the .under file I have included the T0218.t04.many.frag file (NOTE: Do not forget to comment out again the Template.atoms file after this run.) Tue Aug 10 02:45:19 PDT 2004 Kevin Karplus try6 is still pretty feeble. Of course, the very thin alignments mean that the neural net predictions are not worth much and the rr predictions are weak. I noticed that MANUAL_TOP_HITS was not set, so I set it and ran make extra_alignments and make all-align.* to update the alignment set. I also set PREFERRED_AL_METHOD to t04, since that has marginally more sequences in the alignment, and ran make to update the links to rasmol scripts. Wed Aug 11 12:06:59 PDT 2004 Martina Koeva I have started try7 and have increased sheet constraints weights even more. At this point the only thing that I really want to achieve is to have those few strands that Robetta pulls into a sheet in a sheet. So far though we are not even successful at that. For try7 I have also increased the rr constraints' weight by a factor of 10 and have turned up even more the break weight. I don't really want to increase the hbond_geom_beta_pair anymore, since it is already at 15, but I did increase the weight on hbond_geom_beta. Wed Aug 11 15:12:43 PDT 2004 Martina Koeva Try7-opt1 looks terrible. I am not sure where to go next. I can try submitting try6 to VAST again, but I don't think anything useful wil come from the structural alignment. The last structure that I submitted to VAST was not much worse than try6, so I am not clear that we will get much improvement from it. VAST ID: VS60678 Password: T0218try6 Thu Aug 12 00:22:54 PDT 2004 Martina Koeva So VAST came up with nothing, except for two hits for a little subdomain and they look pretty gappy too. At this point I am about to give up. I am not clear as to why we cannot form any sheets, or even strands, where strands are predicted. I will go back and look at the score functions, but none of the models we have really look like proteins. Thu Aug 12 01:07:40 PDT 2004 Martina Koeva Ok, so after I looked at the sheet and strand constraints, is it possible that we are not really forming a sheet, because we do not even have the strand constraints for that segment of residues at all in the .costfcn right now. I will set up those strand constraints for try8 and leave the sheet constraints as they are. Thu Aug 12 01:48:59 PDT 2004 Martina Koeva I picked up a couple of strand constraints from bys.t04.constraints that cover the segments of residues for which I am trying to form sheets: # from t04 bys constraints StrandConstraint T23 E26 7.48 StrandConstraint L31 E32 6.28 StrandConstraint V76 D79 6.89 StrandConstraint T84 S89 6.06 StrandConstraint K92 P94 6.51 I have scaled the weights by a factor of 10. Try8 is running on coo. Thu Aug 12 05:37:37 PDT 2004 Kevin Karplus With the unconstrained cost function, try8-opt2.repack-nonPC now scores second best, after robetta-model2.pdb. Both the robetta model and try8-opt2 look like junk, but we probably won't get much better. It might be worth upping the weights of rr constraints, (that is, take out the "bonus" keyword). Thu Aug 12 09:09:55 PDT 2004 Kevin Karplus try9-opt2 looks pretty junky also. Thu Aug 12 11:51:12 PDT 2004 Martina Koeva try8-opt2 and try8-opt2.repack-nonPC also seem to be scoring better than try9 with the try9 cost function. Thu Aug 12 11:57:28 PDT 2004 Martina Koeva All right, another major mess up. I just copied try8.costfcn over the try9 costfcn, instead of copying try9 over to a try10 costfcn. As a result, the try9 costfcn is completely lost and I don't see any way of recovering it right now, except for the log files. Thu Aug 12 12:56:42 PDT 2004 Martina Koeva Try10 has been set up. I used the sheet constraints from try8-opt2.sheets. I added some more strand constraints from the t04.bys.constraints file. # from try8-opt2.sheets SheetConstraint Q36 E38 D43 I41 hbond E38 14 SheetConstraint I46 V48 R53 K51 hbond V48 14 SheetConstraint L153 V155 L175 K177 hbond I154 14 # from align1.sheets SheetConstraint Q22 E26 I46 L42 hbond T23 7 SheetConstraint E27 E32 I41 Q36 hbond S28 7 SheetConstraint K51 S54 D79 V76 hbond K51 7 SheetConstraint F55 V56 G74 V73 hbond V56 7 SheetConstraint G74 R80 F102 K96 hbond G74 7 # from align5.sheets SheetConstraint P172 N178 V205 K199 hbond G174 7 SheetConstraint K202 V205 I222 V219 hbond F204 7 All weights on the strand constraints have been rescaled by 10. Thu Aug 12 14:22 2004 Bret Barnes Ok, so I spent some of last night and this morning reviewing what has been done so far. This one looks pretty damn sloppy. I came up with two different topologies. I'm going to steal try11 and try12, but it will take me a little bit to set up the costraint files. I think one thing I might added to my constraint files are some constraints to keep helicies that I think should be on the same side of the sheet next to eachother. Kind of like as if they were RR constraints. They will definetly be bonus constraints, but it might help pull the structure together. Thu Aug 12 15:34:54 PDT 2004 Martina Koeva Try10 crashed after producing an opt1 and opt1-scrwl structures. Since there are already structures formed, I am renaming the try to try13 and restarting it with no changes made to costfcn. Thu Aug 12 15:54:50 PDT 2004 Martina Koeva I have been examining try10-opt1, since it is the only clue as to what I can expect from try13. Interestingly enough, in my enthusiasm to form sheets I have introduced some sheet constraints that are a quite dubious: For example, SheetConstraint G74 R80 F102 K96 hbond G74 7 We have a pretty good pairing of two strands there, but we predict the majority of those segments to be helical.Similarly, SheetConstraint E27 E32 I41 Q36 hbond S28 7 forms a good sheet of two strands, but we do not predict those segments to be strands.Otherwise, some of the other strand pairings show potential to form sheets. Right now, I expect try13 to show an improvment over try8, which is about the best that we have so far and it looks like junk. Thu Aug 12 16:28 2004 Bret Barnes running try11 on meow. Thu Aug 12 17:00:51 PDT 2004 Martina Koeva I am not having any luck today. I just happened to run try13 on croak and had to kill it myself a couple of minutes ago: ------------------------------------------------------ Date: Thu, 12 Aug 2004 16:42:45 -0700 (PDT) From: Krishna M. Roskin To: Martina Koeva Subject: running on croak Matrina, Can you please no run CPU intensive jobs on my machine (croak)? I need it's CPU. -krish ---------------------------------------------------------- I am restarting it as a try14 now on bark, since we have already generated model files for try13. From karplus@soe.ucsc.edu Thu Aug 12 17:46:18 2004 Date: Thu, 12 Aug 2004 17:46:17 -0700 From: Kevin Karplus To: martina@soe.ucsc.edu, bbarnes@soe.ucsc.edu CC: karplus@soe.ucsc.edu Subject: T0218 needs to be packaged for shipping tonight I'm going home for supper now. When I finish supper, I'll want to send off T0218. I can't do another all-nighter tonight. Please list in the README file the models to be sent and why. I would like to include the best scoring model with an unconstrained costfcn the model with the lowest Rosetta energy try1-opt2 model 1 from the undertaker-align.pdb : T0218-1muhA-t04-local-str2+CB_burial_14_7-0.4+0.4-adpstyle5 That leaves on other model to choose. Which do you want to include? Thu Aug 12 19:47:18 PDT 2004 Martina Koeva Ok, so I am voting for the following: 1. try8-opt2.repack-nonPC (the best scoring model with the unconstrained costfcn after the robetta-model2) 2. not sure ... I have never scored myself files with the Rosetta energy function. I have been trying to go through Make.main to find how to do that... 3. try1-opt2 (fully automated run) 4. T0218-1muhA-t04-local-str2+CB_burial_14_7-0.4+0.4-adpstyle5 5. try10-opt1 (it does not score very well, but it does score well for an opt1. The cost function is the same as the one in try14, which is still running. If it finished, I imagine that try14-opt2 will be scoring higher. It is one of tries in which we are finally forming strands and sheets even though they are still spread out). From karplus@soe.ucsc.edu Thu Aug 12 19:56:23 2004 Date: Thu, 12 Aug 2004 19:56:22 -0700 From: Kevin Karplus To: martina@soe.ucsc.edu CC: karplus@soe.ucsc.edu Subject: best T0218 I got try8-opt2 as best unconstrained, and by doing grep total decoys/*repack* | sort -g +13 I got that try2-opt2.repack-nonPC is the lowest Rosetta energy. What I need from you is an opinion on which other model is worth trying. Also, I noticed that you had never tried doing a run that combined the existing models, so I'm starting one now (try15) on cheep. ------------------------------------------------------------ Thu Aug 12 20:01:08 PDT 2004 Kevin Karplus OK, I'll use try10-opt2 as the other model. I'll submit try8-opt2 best scoring try10-opt1 promising, but incomplete try2-opt2.repack-nonPC lowest Rosetta energy try1-opt2 full auto T0218-1muhA-t04-local-str2+CB_burial_14_7-0.4+0.4-adpstyle5 If we want to add try14-opt2 or try15-opt2 when they finish, I can knock out either try10-opt1 or the final model from alignment. Thu Aug 12 20:22:35 PDT 2004 Kevin Karplus I've done the submission. I've also restarted try15, with many short runs of high crossover for the first part, in the hope of getting something other than just a polishing of robetta-model2. Thu Aug 12 21:00 2004 Bret Barnes At this point there doesn't seem to be much of anything forming. Perhaps its best to start looking at other targets since it seems we are not making much progress. I'm going to review the status file and see what are next targets due soon and see if there isn't anything I can do to help out on those. Thu Aug 12 22:28:08 PDT 2004 Martina Koeva Try14 is still running, but try15 is already done. I have rescored models with the unconstrained cost function and try15-opt2 scores better now with it. As far as the Rosetta energy function, try2-opt2.repack-nonPC is still a little ahead of try15-opt2.repack-nonPC. I am still waiting for the try14-opt2 and if it scores better than try10-opt1, which it should, I would say we should change the submitted models to: try15-opt2 (instead of try8-opt2) try2-opt2.repack-nonPC (unless try14-opt2.repack-nonPC happens to score better) try1-opt2 T0218-1muhA-t04-local-str2+CB_burial_14_7-0.4+0.4-adpstyle5 try14-opt2 (instead of try10-opt1) Oh, try14 has just finished and I rescored all models again. Try14 scores way down with the Rosetta energy function after try2,try15,try1,try3,and try4. With the unconstrained function try14-opt2 still does not score as well as try8-opt2, but I do think it is a better model than try10-opt1. Try15-opt2 still scores the best. With the try14 cost function, try14-opt2 scores the best with try14-opt2.repack-nonPC (better than try15-opt2 and try8-opt2). In this case I would say we should go with the list I just put above. Fri Aug 13 07:43:31 PDT 2004 Kevin Karplus The try15-opt2 output turned out to be only a minor polishing of robetta-model2, so I won't submit it. If it had succeeded in actually mixing robetta-model2 with something else, I would have used it. I will replace try10-opt1 with try14-opt2.