Tue Jul 11 09:13:01 PDT 2006 T0369 Make started Tue Jul 11 09:15:23 PDT 2006 Running on whidbey.cse.ucsc.edu Tue Jul 11 09:58:38 PDT 2006 Kevin Karplus BLAST gets no good hits, but the HMMs see 2f22A as a good possibility. Fri Jul 21 15:22:49 PDT 2006 Firas Khatib well, there is quite a bit of disagreement between the alignments! I like the burial of try1 according to "near", but the "burial" script is VERY different, with S112-T134 being all red! the strands from 104-111 & 63-65 do not seem to be predicted using dssp-ehl2 or str2 (either t2k, t04, or t06) I am going to rerun try1 using just str2-t06's constraints try2 is running on shaw with T0369.t06.str2.constraints instead of dssp-ehl2 Sat Jul 29 18:47:49 PDT 2006 Martin Madera According to str2, four long helices... this looks like a 4-helix bundle. But the predictions may be wrong, or not the whole story, because of the strongly-conserved histidines bang in the middle of helix four. The HMMs really like 2f22A (unknown superfamily) and 1rxqA (which is in a.213.1.1). The two structures are clearly from the same superfamily and look very similar. Both have many missing atoms, though 1rxq is better than 2f22. 2f22 is a dimer with a large, well-packed interface. 1rxq has four molecules in the unit cell, but it's clear that it's also a dimer. However, the interfaces are different, *sigh*. (You can see that clearly by focusing on where the small beta sheets are.) Try1 screwed up the helices. In particular, the third helix (77-96) is in the wrong place, it needs to be moved down (looking at the protein with the beta sheet at the top) and packed against the other three. Try2 suffers from the same problem. So we clearly need to add constraints to keep the helices in place. Looking at T0369.undertaker-align.pdb.gz, models 1 & 2 have the full 4-helix bundle, so may be a good source of constraints. I have decided to focus on: select 12,53,85,120 From model 1: Distance LEU85A.CA-VAL12A.CA: 8.329 Distance LEU85A.CA-ALA120A.CA: 10.593 Distance LEU85A.CA-LEU53A.CA: 6.925 Distance VAL12A.CA-ALA120A.CA: 7.153 Distance VAL12A.CA-LEU53A.CA: 10.661 Distance ALA120A.CA-LEU53A.CA: 7.643 From model 2: Distance LEU85A.CA-VAL12A.CA: 6.720 Distance LEU85A.CA-ALA120A.CA: 10.833 Distance LEU85A.CA-LEU53A.CA: 8.288 Distance VAL12A.CA-ALA120A.CA: 7.795 Distance VAL12A.CA-LEU53A.CA: 10.745 Distance ALA120A.CA-LEU53A.CA: 7.909 Which gives me: # constraints to keep the 4-helix bundle together Constraint L85.CA V12.CA 4.0 7.0 10.0 1 Constraint L85.CA A120.CA 7.7 10.7 13.7 1 Constraint L85.CA L53.CA 4.5 7.5 10.5 1 Constraint V12.CA A120.CA 4.5 7.5 10.5 1 Constraint V12.CA L53.CA 7.7 10.7 13.7 1 Constraint A120.CA L53.CA 4.8 7.8 10.8 1 Try3 and try4 started on orcas. They're the same as try1, but adding the above constraints. Sun Jul 30 18:34:44 PDT 2006 Martin Madera Try3 failed. I don't understand why, the errors in try3.log are cryptic: Error: Couldn't open file /var/tmp/from_scwrl_1220454422.pdb or /var/tmp/from_scwrl_1220454422.pdb.gz for input Hmm. Try4 worked and the .under files are identical (up to s/try3/try4/), and so are the .costfcn files... strange. Try4 blew up *again*. Let me check the constraints: Constraint L85.CA V12.CA 4.0 7.0 10.0 1 ... 18.765 Constraint L85.CA A120.CA 7.7 10.7 13.7 1 ... 19.400 Constraint L85.CA L53.CA 4.5 7.5 10.5 1 ... 16.943 Constraint V12.CA A120.CA 4.5 7.5 10.5 1 ... 6.728 Constraint V12.CA L53.CA 7.7 10.7 13.7 1 ... 6.550 Constraint A120.CA L53.CA 4.8 7.8 10.8 1 ... 7.528 Right. I'll bump up the constraint weights to 10, and constraints from 10 to 30. Two more runs from alignments, try5 and try6. Running on camano and squawk. Sun Jul 30 22:25:56 PDT 2006 Martin Madera Try5 and try6 didn't work. Hmmmmmmm. AAAAH! It seems I copied the .under file from try2 and didn't notice that try3-try6 were all trying to polish try1! DOH! Try7, try8, try9 (with try3 / try4 costfcn) running on camano and squawk. Mon Jul 31 10:55:48 PDT 2006 Martin Madera Success at last! Try7 managed to hold the bundle together but introduced a chain break. Try8 and try9 still got torn apart. This indicates that I still need to increase the weight of the constraints. I'll go with the try5/try6 costfcn, but otherwise still from alignments, I want to see some alternative structures. These are try10 (camano), try11 and try12 (abyss). I have also started try13, a polishing run on try7 that I hope will close the break. I kept the constraints from try7 but tripled soft clashes and break. Running on shaw. Mon Jul 31 15:29:12 PDT 2006 Firas Khatib try13 is very similar to try7, still has the break, but scores better with the try13 costfcn, the try7 one and even unconstrained. I will attempt try14 based on try13 to fix the chainbreak by adding constraints to keep the helices in place. Mon Jul 31 17:44:30 PDT 2006 Firas Khatib as of right now, try14-opt1 is doing the best. here is a recap: try14 -> try13 -> try7 -> the follwing alignments: 1rxq, 2f22, & 5eau try12 -> the follwing alignments: 1rxq, 2f22, 5eau, and 5eat try10 -> the follwing alignments: 1rxq, 2f22, and 5eau try2 -> try1 -> the follwing alignments: 1rxq, 2f22, 1r2r, and 5eau we don't really trust try2 and try4, but they score best using the unconstrained cost function so we should include one of them as model 5. Mon Jul 31 18:12:33 PDT 2006 Firas Khatib so far, try14/try13 are the best, we will try to fix that chain break I will run try15 with try12 as input since it has the chain break in a different location that might be easier to fix. I don't really like try10 and feel that it will probably not be submitted try15 uses try14's costfcn but with try12 as input, it's running on shaw Mon Jul 31 18:23:50 PDT 2006 Firas Khatib try16 uses try14-opt1 as input with try14's costfcn but a high break cost try16 is running on lopez as soon as try14-opt2 finishes (and if the break is not healed) I will do the same for it. Mon Jul 31 18:37:44 PDT 2006 Firas Khatib try17 is running on shaw using try14-opt2 as input and try14's costfcn, but with breaks increased to an insane 500! constraints are at 200. try16 is basically the same with breaks at 400 and constraints at 100. this is all due to the fact that since try7 we cannot close that break at F71 even with a high break penalty of 300! recapping: try17 -> try14-opt2 -> try13 -> try7 -> alignments: 1rxq, 2f22, & 5eau try16 -> try14-opt1 -> try13 -> try7 -> alignments: 1rxq, 2f22, & 5eau try15 -> try12 -> alignments: 1rxq, 2f22, 5eau, and 5eat try10 -> alignments: 1rxq, 2f22, and 5eau try2 -> try1 -> alignments: 1rxq, 2f22, 1r2r, and 5eau Mon Jul 31 18:49:34 PDT 2006 Firas Khatib try11 doesn't have the chain break, but is foamy. I will run try18 with try11 as input and see if this can be fixed. try18 is running on camano with try14's costfcn and try11 as input try19 is running on camano with try11's constraints and increased dry terms with try11 as input. recapping: try17 -> try14-opt2 -> try13 -> try7 -> alignments: 1rxq, 2f22, & 5eau try16 -> try14-opt1 -> try13 -> try7 -> alignments: 1rxq, 2f22, & 5eau try15 -> try12 -> alignments: 1rxq, 2f22, 5eau, and 5eat try19 & try18 -> try11 -> alignments: 1rxq, 2f22, and 5eau try2 -> try1 -> alignments: 1rxq, 2f22, 1r2r, and 5eau Mon Jul 31 19:15:25 PDT 2006 Firas Khatib we are currently waiting on try15-try19 to finish up. they are expected to be worth the wait. once they are done, Martin will write up a few sentences for each. Mon Jul 31 20:36:23 PDT 2006 Kevin Karplus The constraints are probably way too strong---you can have huge breaks and still score well if your constraints are just a tiny bit better satisfied. try16 looks like it came the closest to closing the breaks (close enough that gromacs thinks it can close them). If it looks ok, we should probably do just one more polishing run from the try16.gromacs0.repack-nonPC model. Mon Jul 31 20:52:22 PDT 2006 Kevin Karplus starting a try20 polishing run from all the *.gromacs0.repack-nonPC models, with no constraints but with high dry weights and high break weight. I expect that try16 and try17 will contribute most to the optimization. Mon Jul 31 21:19:02 PDT 2006 Martin Madera Notes on the various runs. Try1-try6, try8-try9 are failures: the third helix is clearly in the wrong place. However, try2 is our best-scoring model! Try7 is the first run from alignments where undertaker kept the helix in right place, but it has a large break. However the helix constraints were pretty weak, so try8 and try9 (which had the same constraints) still blew up. Try10-try12 was another series of runs from alignments, but with stronger constraints to keep the third helix in place. Firas doesn't like try10 (and I agree, two big breaks etc.). Try11 has no major chain breaks in rasmol, the first model to achieve that! Firas thinks it's foamy, I don't think it's that bad. Try12 has created two small helices in the loop leading up to the helix 3 (which is probably wrong) and has a chain break. Try13 was an unsuccessful attempt to close the breaks in try7, but it improved the score. Try14 was an attempt to close the break in try13, but again unsuccessful. Then try15 -> try12 -> alignments: 1rxq, 2f22, 5eau, and 5eat No chain break in rasmol. An alternative structure for the loop leading up to the third helix. try16 -> try14-opt1 -> try13 -> try7 -> alignments: 1rxq, 2f22, & 5eau ... but still a chain break in rasmol! However, try16 is our second best model according to the unconstrained cost function. try17 -> try14-opt2 -> try13 -> try7 -> alignments: 1rxq, 2f22, & 5eau ... still a chain break in rasmol! Scores slightly worse than try16 according to unconstrained costfcn. try19 & try18 -> try11 -> alignments: 1rxq, 2f22, and 5eau ... try19 is our 3rd best scoring model, the first one with a sensible helix (unlike try2) and no visible chain breaks in rasmol (unlike try16). I would say this is our best model. And, after I wrote this, I noticed that Rosetta thinks that try19-opt2.gromacs0.repack-nonPC is our best model! try20 is a polishing run of all gromacs models. Rosetta doesn't like it and neither does the unconstrained costfcn. So, what to submit? - try19 is the best model (according to what I think and what Rosetta thinks), third best according to undertaker - try16 is our second-best scoring model, but it still has a chain break in rasmol; it's the best from the try7 -> try13 -> try14 lineage - try15 scores quite badly, but it has no visible chain breaks in rasmol and has a different structure for the linker before the third helix - try20 has fixed all the clashes and breaks, but for some reason it scores poorly - try2 ... we probably think it's wrong (the helix is in the wrong place), but it scores well a more detailed method section on these five: ---------------------------------------------------------------------- Our models are based on 1rxq, 2f22 & 5eau. We've had trouble keeping the four-helix bundle together in undertaker, and when we managed to do that, we had trouble fixing breaks in the chain. Model 1 is try19; it isn't our best-scoring model (it comes third) but it is the first model with the third helix in the right position and without any major breaks in the chain. We like it, and Rosetta agrees. Model 2 is try16. This is our second-best model score-wise. It has the helix in the right place, but it also has a large break in the chain after A69. Model 3 is try15. Undertaker created two short helices in the loop connecting the second and third long helices, which is probably wrong. But it is an interesting alternative structure. Model 4 is try20. Here we have managed to close most chain breaks and eliminated most clashes, but the result scores poorly. Model 5 is try2. The third helix is clearly in the wrong position, but according to our cost function this is the best-scoring model! ---------------------------------------------------------------------- Updated MANUAL_TOP_HITS and best-models. Done with this target. Mon Jul 31 22:23:25 PDT 2006 Kevin Karplus If you are going to include try19-opt2 because Rosetta likes it so much, you may want to use try19-op2.gromacs0.repack-nonPC as one of the models. One you start getting gaps to be small, the standard unconstrained costfcn is no longer the best way to judge models---you need to start increasing the weights of the finer detail cost function components. For example, the try20 costfcn prefers try20, try16, try17, try14, try19. I'm going to move try20 to the top, and replace try19 by try19-opt2.gromacs0.repack-nonPC. try2 scores worse (as it deserves) with the try20 costfcn, mainly because of bad breaks. Mon Jul 31 22:48:37 PDT 2006 Kevin Karplus submitted with comment Our models are based mainly on 1rxq, 2f22, & 5eau. We've had trouble keeping the four-helix bundle together in undertaker, and when we managed to do that, we had trouble fixing breaks in the chain particularly for the linker T61-V76, which caused problems for most of the servers also. Model 1 is try20-opt2. Here we have managed to close most chain breaks and eliminated most clashes, but the result scores poorly with the standard unconstrained costfcn, mainly because of a few poor matches to the predicted alpha angles. Try20 was optimized by undertaker from try17-opt2.gromacs0.repack-nonPC. try17-opt2 < try14-opt2 < try13-opt2 < try7-opt2 < alignments (last alignment added was to 5eau) Model 2 is try19-opt2.gromacs0.repack-nonPC, the backbone that Rosetta liked best of all the ones it repacked sidechains for. It was made from try19-opt2 by reoptimizing with gromacs and repacking the sidechains (except PRO and CYS) with rosetta. Try19-opt2 was created by undertaker using two rounds of optimization. try19-opt2 < try11-opt2 < alignments (last alignment to 5eau) Tyr19 was the first model with the third helix in the right position and without any major breaks in the chain. We like it, and Rosetta agrees. Model 3 is try16-opt2. This is our second-best model with a couple of different undertaker cost functions (both unconstrained.costfcn and try20.costfcn). It has the helix in the right place, but it also has a large break in the chain after A69. try16-opt2 < try14-opt1 < try13-opt2 < try7-opt2 < alignments (last alignment added was to 5eau) Model 4 is try15-opt2. Undertaker created two short helices in the loop connecting the second and third long helices, which is probably wrong. But it is an interesting alternative structure. try15-opt2 < try12-opt < alignments (last alignment added was to 5eat). Model 5 is try2-opt2. The third helix is clearly in the wrong position, but according to the unconstrained cost function this is the best-scoring model! ------------------------------------------------------------ Mon Jul 31 22:55:03 PDT 2006 Kevin Karplus Submission was rejected: Martin had incorrectly put 5eauA and 5eatA in the Makefile---there is no chain A in those PDB files. I removed the extraneous chain ID and resubmitted. Sun Aug 20 20:09:15 PDT 2006 Kevin Karplus Our best model was try12-opt2 (after running through gromacs) Our best submitted was model4 (try15-opt2). These beat all but a few of the server models. Note: the SAM_T06_server did horrbily.