Fri Jan 14 17:31:31 PST 2005 Kevin Karplus In this directory, I'll redo all the previous work, but starting from a better set of initial alignments, using all available templates, including the new one I got today from Titia Sixma. I will take advantage of the work in ../redo-casp-method, but I'm hoping to be able to get as good or better models quickly, since I'll be using the two templates that aren't yet in PDB more thoroughly (I'm adding them to the template library today). I'll have to make decent 5mer alignment files for 1i9bA, bulcA, and aplyA, so that 5mers can be created from the monomers. Sat Jan 15 11:08:17 PST 2005 Kevin Karplus On try1-opt2, the unconstrained cost is already as good as the try7-opt2 run for ../redo-casp-method. I tried running gromacs, using George Shackelford's instructions. That is, copying the casp6/gromacs directory, connecting to it and running dogromacs.sh ACH7_SHORT.try1, using the OPLS-AA/L all-atom force field and outputting group 1 (protein) atoms. I then fixed the last two atom names to be O and OXT instead of O1 and O2, and moved the ACH7_SHORT.try1-opt2.gromacs.pdb.gz file to decoys/ As usual, gromacs improved the clashes and breaks at the cost of completely messing up the sidechains. Repacking it with rosetta did not quite restore the disulfides, but still kept the clashes and breaks low. For the try2 run, I'll start from the existing models. I tweaked the cost function, increasing the weights of components that the repacked gromacs try1-opt2 did well on until it just barely got the best scores. Sat Jan 15 11:59:50 PST 2005 Kevin Karplus I just noticed that the aplysia protein (aplyA) may not have been added to the template library completely, and so may not have been properly searched for---it *seems* to have come out ok though, so I don't think any harm came of it. It seems to score better than the bulcA (Bulinus sequence), but not as well as the 1uw6A and 1i9bA sequences. Unfortunately, the alignments to aplyA and bulcA were NOT included in all-align.a2m (because the names did not begin with a digit), so were not included in the try1 and try2 alignment sets. For try3, I should probably try again from the alignments with the alignments for bulcA and aplyA included. I'll also add a new RandomConform command to undertaker, and do two different random starts before TryAllAlign commands. Sat Jan 15 13:29:34 PST 2005 Kevin Karplus I ran try2-opt2 through gromacs, as with try1-opt2, and repacked it with Rosetta. This is the best Rosetta energy for the three-template runs, but not yet as good as ../redo-casp-method/decoys/ACH7_SHORT.try13-opt2.gromacs.repack-nonPC.pdb (or even try12-opt2.gromacs.repack-nonPC, try11-opt2.gromacs.repack-nonPC, try13-opt2.repack-nonPC). Once again gromacs messed up the sidechains (especially one of the disulfides), but it actually increased the breaks and left the clashes almost unchanged. Repacking with rosetta kept the clashes fairly low, but did not fix the proben disulfide. SCWRL runs (on the try*-opt1 models) destroy both disulfides, as well as introducing a lot of clashes. Sat Jan 15 14:03:37 PST 2005 Kevin Karplus try3-opt2 doesn't score quite as well as try1-opt2 with the try3 costfcn, probably because the high clash and break costs made it too difficult to change the backbone much. I ran gromacs with both the 0: GROMOS96 43a1 Forcefield (official distribution) and 3: OPLS-AA/L all-atom force field (2001 aminoacid dihedrals) forcefields (as ACH7_SHORT.try3-opt2.gromacs0.pdb.gz and ACH7_SHORT.try3-opt2.gromacs.pdb.gz, respectively). Both completely mess up the sidechains, but the gromacs0 does slightly better on other measures, and slightly better after repacking by Rosetta, though Rosetta likes the repacked gromacs better than the repacked gromacs0 version of try3-opt2. Perhaps the most striking thing in try2-opt2, relative to the designs in ../redo-casp-method, is that R99 is pointing into the monomer, instead of into the center of the channel. try3-opt2 also has R99 pointing into the monomer, but has moved the N-terminal helix into a position that is not compatible with a pentamer. try1 and try2 have both left it close to where it was in the 1uw6A and aplyA structures. I like the way the aplyA alignment ACH7_SHORT-aplyA-t04-local-str2+CB_burial_14_7-0.4+0.4-adpstyle5.a2m places the R99, but to orient it that way it had to omit D97. Still, it might be worthwhile to try an optimization run with the sheet constraints from that alignment. Sat Jan 15 16:48:13 PST 2005 Kevin Karplus The try4-opt2 model scores almost as well as try2-opt2, using any of the cost functions, despite starting from alignments rather than from polished models. Rosetta does not like that much though. Sat Jan 15 17:15:12 PST 2005 Kevin Karplus I tweaked the try5 costfcn so that try4-opt2 slightly outperforms try2-opt2, and will use all models as the starting point for try5. Sun Jan 16 04:43:23 PST 2005 Kevin Karplus I ran gromacs on try5-opt2 using all 5 energy functions that gromacs provides as options to see if any of them are useful (not the 6th one, that gromacs says not to use for new work). All of them produce absolutely terrible sidechains, probably as a result of some error in the description that is shared by all the energy functions. The gromacs4 file seems to be the best after repacking, and gromacs2 the worst, but the differences are fairly small. Gromacs4 is the one that does the least damage to the disulfide, gromacs3 does the most damage to the disulfide. Even after repacking with rosetta to fix the bad sidechains, gromacs reduces clashes and breaks, but increases costs of bystroff code (phi-psi class), predicted alpha angles, and sidechain scores. Rosetta universally refers the gromacs optimized backbones for repacking, almost entirely because of the reduced clashes. Because gromacs4 worked best for try5-opt2, and because try2-opt2.gromacs3 was the backbone that rosetta liked best for repacking, I tried making try2-opt2.gromacs4 and repacking it. Rosetta did not like this one as well, again the Erep (Van der Waals repulsion) term is the big difference. The unconstrained costfcn, which has a rather small weight for clashes, likes try5-opt2 best. To move Rosetta's favorite to the top is not really possible, since undertaker sees try2-opt2 as having fewer clashes and breaks than try2-opt2.gromacs3.repack-nonPC---indeed try2-opt2 dominates in almost all components of the cost function. Sun Jan 16 06:13:24 PST 2005 Kevin Karplus For try6, I added a gap constraint to make sure that the C190-C191 peptide is not broken. I also added some salt bridges taken from try5-opt2, and a couple of speculative sheet constraints taken from ../redo-casp6-method/try8.costfcn I'll try a run from the alignments with this cost function, then another run starting from all models. Sun Jan 16 08:06:50 PST 2005 Kevin Karplus The try6 run *still* has a break between C190 and C191, despite the constraints. The try3-opt2 structure did not, so maybe I should cut and paste to get the disulfide from there. I made two chimeras: try5-try3-chimera.pdb and try6-try3-chimera.pdb. I copied S184-D197 from try3-opt2 into try5-opt2.gromacs4.repack-nonPC and try6-opt2, respectively. Interestingly, they both have breaks reported before C190 (cost 0.196487). Oops---so does try3-opt2. How did I miss that before? Let me try again, using try3-opt2.gromacs0.repack-nonPC as the source of the disulfide. Sun Jan 16 10:59:02 PST 2005 Kevin Karplus I was trying to debug undertaker, when I found a bug in gromacs output. In addition to misnaming the final oxygen atoms O1 and O2 (instead of the standard O and OXT), it also misnames ILE CD1 and ILE CD. After writing a script in casp6/scripts to fix gromacs files by changing these atom names, I reran the scoring for all the pdb files. Now the gromacs files do fairly well, with sidechains only slightly worse than undertaker or rosetta. This observation caused me to rebuild the try5-try3-chimera and try6-try3-chimera, using try5-opt2.gromacs0 and try6-opt2 and the bases and try3-opt2.gromacs0 as the source for S184-D197. Superposition was done using undertaker (with the superimpose-best.under script), focusing initially on the residues near the crossover cut: printAllConformPDB best-models.pdb superpose \ residue R183 residue S184 residue D197 residue V198 Now the try5-try3-chimera scores best with the try6 cost function. Sun Jan 16 15:12:50 PST 2005 Kevin Karplus After optimizing existing models with try7, the best models according to various scoring functions are Rosetta score: try2-opt2.gromacs3.repack-nonPC try7-opt2.gromacs4.repack-nonPC try7 & try6: try7-opt2 try7-opt1 try7-opt2.repack-nonPC try7-opt2.gromacs0 try7-opt2.gromacs4 try5: try7-opt2 try5-opt2 try7-opt1 try7-opt2.repack-nonPC try4: try7-opt2 try7-opt1 try5-opt2 try7-opt2.repack-nonPC try2-opt2 try3 & try2: try7-opt2 try7-opt1 try2-opt2 try7-opt2.gromacs4 try7-opt2.gromacs0 try7-opt2.repack-nonPC try1: try7-opt2 try7-opt1 try2-opt2 try5-opt2 try7-opt2.repack-nonPC try7-opt2.gromacs0 try7-opt2.gromacs4 unconstrained: try7-opt2 try7-opt1 try5-opt2 try2-opt2 try7-opt2.repack-nonPC try5-opt1 try4-opt2 Sun Jan 16 17:58:11 PST 2005 Kevin Karplus For try8, I turned up the clash and break weights, and removed some of the more speculative constraints. The clashes seem to have gone down substantially, but the breaks are not closing much. I may need to find some more powerful gap-closing operators, for closing these tiny gaps. Gromacs does a fairly good job of closing them, but often at a rather high cost in other parts of the cost function that gromacs doesn't evaluate. Sun Jan 16 18:44:38 PST 2005 Kevin Karplus Rosetta now likes best try8-opt2.gromacs3.repack-nonPC, which is ranked 17th by the try8 costfcn, and 32th by the unconstrained2 costfcn. Rosetta probably likes it because it broke the vicinal disulfide, which Rosetta doesn't believe in. The best model according to unconstrained2.cosfcn is try8-opt1. The try8-opt2 is slightly worse, because the disulfides are not as good. There are a few better models with this cost function in ../redo-casp-method (Notably try11-opt2, try12-opt2, and try13-opt2). Perhaps it is time to optimize for the unconstrained2 cost function, using the best models from both directories. Mon Jan 17 09:41:18 PST 2005 Kevin Karplus It may have been premature to include the ../redo-casp-method models for optimizing with unconstrained2.costfcn. The try9 run seems to have just optimized the old ../redo-casp-method/decoys/ACH7_short.try12-opt2 model, rather than using a lot of crossover with the try8-opt2 models. Perhaps it would have been better to use the try8 cost function, to keep the constraints in place for another round, as try9-opt2 does slightly worse than try8-opt2 with try8.costfcn, mainly on the basis of the constraints. Rosetta still likes best try8-opt2.gromacs3.repack-nonPC, but try9-opt2.gromacs4.repack-nonPC is almost as good. Question: why did the pred_alpha values change so much in gromacs optimization, when I can't really see any backbone changes in rasmol? I used superimpose-best.under to print out phi-psi tables for try9-opt2 and try9-opt2.gromacs4 Hmm, there seem to be gross errors in the tables for try9-opt2, probably where there are breaks in the backbone---I'll have to look at how alpha is computed at breaks. Mon Jan 17 10:38:36 PST 2005 Kevin Karplus I'll do a remake on the entire directory, because George Shackelford tells me that the error in the traincontactnn was indicating a missing or empty file---one that was *not* empty. Perhaps it was a transient file-server error? I'll try again to get contact predictions. The bug in the alpha computations is probably encoraging undertaker to make tiny breaks, since they make the alpha (and bys?) values unnaturally good by supressing the computation of the alpha and bys scores at those points. Mon Jan 17 11:31:41 PST 2005 Kevin Karplus traincontactnn is definitely buggy---it still fails with the same error message though ACH7_SHORT.a2m is ok. Mon Jan 17 12:44:18 PST 2005 Kevin Karplus I fixed the bug in undertaker that caused mis-scoring of alpha and bys when there were tiny breaks. Now the try8 costfcn prefers try9-opt2.gromacs0. I should do another optimization run, using the try8 function (or a similar one that still has constraints) to reoptimize from all models except the try9 ones, which have been "contaminated" with the ../redo-casp6-method models. Mon Jan 17 13:02:28 PST 2005 Kevin Karplus OK, starting a run with very high clash and break costs as try10. It includes try2, try7, and try8 models as starting points. Mon Jan 17 14:17:46 PST 2005 Kevin Karplus The try10 models have only 2 breaks (undertaker is no longer introducing lots of tiny breaks, hooray!). Running try10-opt2 through gromacs introduces a break in the vicinal disulfide, but removes one before P28. The other break (before P170) is not touched by gromacs. try10-opt2 and try9-opt2 are scoring very similarly with most cost functions, though cost functions that put a high premium on breaks and clashes prefer gromacs-optimized versions of try9-opt2 over the undertaker backbone (not surprising, since try9 was done when undertaker was introducing tiny breaks). Rosetta's favorites are now try8-opt2.gromacs3.repack-nonPC try9-opt2.gromacs4.repack-nonPC try10-opt2.gromacs4.repack-nonPC Mon Jan 17 14:35:08 PST 2005 Kevin Karplus At P170, try9 has quite a different solution than try8 and try10, and it might be worth copying. The loop from V22-Q27 (before the P28 residue) is also quite different in try9, but I find the triple Hbond off of V22.O in try9-opt2.gromacs0 a bit dubious. I'll do one more optimization run, including all models, but with no constraints. First, I'll may a try10-chimera by taking P21-L29 from try10-opt2.gromacs4 and the rest from try10-opt2. Mon Jan 17 16:44:24 PST 2005 Kevin Karplus New optimization run startes as try11. The try10-chimera seems to be the best initial model on this run. Mon Jan 17 22:05:59 PST 2005 Kevin Karplus Rosetta's order for the models is now decoys/ACH7_SHORT.try8-opt2.gromacs3.repack-nonPC decoys/ACH7_SHORT.try9-opt2.gromacs4.repack-nonPC decoys/ACH7_SHORT.try10-opt2.gromacs4.repack-nonPC decoys/ACH7_SHORT.try11-opt2.gromacs4.repack-nonPC The try8-opt2.gromacs3 model messed up the vicinal disulfide. The try11 costfcn puts try11-opt2 first with gromacs not helping---even the clashes and breaks get worse with gromacs. try9-opt2 is also ok. It is interesting that rosetta and gromacs seem to agree more with each other than with undertaker. The big differences in rosetta scores are in the VdW repulsive terms, so there is clearly still a difference in the clash detectors. I've probably done about as much as I can with the monomer, and it is time to switch back to the pentamer, and see what I can do with that with the improved undertaker. Once I get a decent pentamer again, I'll *finally* get back to doing design work. Tue Jan 18 22:28:11 PST 2005 Kevin Karplus I set up the subdirectory 5mer/ today, and ran the undertaker script make-5mer-try11.under to create starting points for optimization. The initial 5-mer is not as good as the highly optimized 5mers I had in ../redo-casp-method/5mer, but seem to be a decent starting point. Interestingly, the aplyA alignment seems to have produced the best initial pentamer. Wed Jan 19 07:30:44 PST 2005 Kevin Karplus Foo! undertaker crashed after about 6 hours, but before it output any results for 5mer/try1. Wed Jan 19 10:47:20 PST 2005 Kevin Karplus The assertion failure was in XYZpoint::unitize(), where a zero-length vector needed to be made unit length. It is not clear where this was called from, so it is not clear which test is missing. Since I just added all the has_* checks in the Conformation class, based on the older ones in the ChainsResiduesAndAtoms class, I suspect one of them is incorrect. Fri Jan 21 17:30:30 PST 2005 Kevin Karplus The fix I made to has_phi seems to have removed the bug, but there are other problems now. Although 5mer/try1 did not use TweakMultimer on the first round, 5mer/decoys/5mer.try1-opt1.pdb has blown up the pentamer into widely spaced monomers. This appears to have happened during a "ShiftSubtree" or "OptSubtree" operation, neither of which should have had that effect. Perhaps I need to have the compactness scores weighted a bit higher relative to the clash and break, at least until the worst clashes are removed. Fri Jan 21 22:35:33 PST 2005 Kevin Karplus For try2 (started before try1 finished), I added in the phobic_fit parameter and scaled down n_ca_c, soft_clashes and break, though (by mistake) I left backbone_clashes high. The 5mer runs are very slow, probably because of the cost of re-symmetrizing after every conformation change. I probably want to look for a more efficient way of handling homomultimers. Sat Jan 22 09:06:14 PST 2005 Kevin Karplus The try2-opt1 run keeps the pentamer together and has started reducing the clashes. I may be able to reduce the breaks by crossing with try1-opt2, but both have breaks before P170, A102, and L29. This is more breaks than I expected to see. No wonder! The make-5mer-try11.under script read try1-opt2, not try11-opt2, so I've been starting with poor models! Sat Jan 22 11:52:20 PST 2005 Kevin Karplus Foo, the 5mer/try2 run crashed after running for 12 hours. Maybe I need to allow core dumps again, so that I can pick through the remains of a crash like this. For 5mer/try3, I'll run from the proper 5mer-try11-*.pdb files, with a score function intermediate between try1 and try2. I'll also turn off the maintenance of the 5mer, and homogenize the models when I'm done. Sat Jan 22 14:51:16 PST 2005 Kevin Karplus The 5mer/try3 run was pretty quick, and in 5mer/ I ran "make homogenize-try3-opt2" which created lots of symmetrized versions based on taking different monomers superimposed on the 3 templates. Although the try3 runs score best, the symmetrized 5mer-aplyA-try3-opt2D and 5mer-aplyA-try3-opt2C do fairly well. The 5mer/try4 run will use the same costfcn as try3, but will start from the various symmetrized models. If I get little improvement, I'll see if gromacs can fix up any of the breaks, then do a run with the symmetry enforced by undertaker. Sat Jan 22 19:22:05 PST 2005 Kevin Karplus The 5mer/try4 run did not get quite as good as the 5mer/try3 run (though very close), but the symmetrized versions have fewer clashes than the symmetrized try3 ones. Sun Jan 23 05:16:26 PST 2005 Kevin Karplus The 5mer/try5 run has done a little better than the try4 run on the symmetrized models, but not by much. There do seem to be some gapless models though, which would be nice. Sun Jan 23 09:28:12 PST 2005 Kevin Karplus The 5mer/try6 run is being done with few generations, but with maintenance of the symmetry. As of try6-opt1, it seems to have reduced the conflicts without introducing breaks, and without introducing large gaps between the monomers. Sun Jan 23 15:08:29 PST 2005 Kevin Karplus 5mer/try6-opt2 had no breaks and only minor clashes, but that does seem to include some backbone clashes. Compared to try6-opt1, it has more clashes and fewer Hbonds, but compensates by having better dry5, dry6.5, and sidechain scores. gromacs is no longer useful, as it increases the breaks while making only small changes to the clashes. The backbone clashes in try6-opt2 are N171.C to G172.O, and are only about 0.1 Angstrom too close. I should now let Rosetta play with the dimer interface, to see if I can fix up any of the remaining clashes, then take the chimeric monomer and polish it once more with undertaker. Tue Jan 25 21:32:44 PST 2005 Kevin Karplus I did a little playing around with the Makefiles, so that I could make new 5mers from a chimera made from the Rosetta repacked dimer. In doing so, I accidentally stepped on the the old 5mer/decoys/5mer-try11-aplyA.pdb ... files, so I re-created them with a new name 5mer/decoys/5mer-try11-opt2-aplyA.pdb ... I created 5mers from the dimer dimer/decoys/dimer-5mer.try6-opt2.repack-nonPC.pdb.gz 5mer-5mer.try6-opt2.from-dimer-aplyA.pdb Fri Feb 4 12:56:49 PST 2005 Kevin Karplus I thought I needed new commands in undertaker to create a 5mer from the monomers copying an existing 5-mer symmetry transformation, but I found that I had already implemented that functionality, so I wrote a new script make-5mer-from-try6.under that creates 5mer/decoys/5mer-5mer.try6-opt2.repack-nonPC.from-dimer.pdb from decoys/ACH7_SHORT.5mer.try6-opt2.repack-nonPC.from-dimer.pdb (which was the chimera created from Rosetta's repacking of the dimer from 5mer/decoys/5mer.try6-opt2.pdb) using the 5mer/decoys/5mer.try6-opt2 symmetries. Sat Feb 5 17:16:35 PST 2005 Kevin Karplus The 5mer-5mer.try6-opt2.repack-nonPC.from-dimer.pdb file does not score as well with the 5mer/try6.costfcn as try6-opt2 itself, but does better than any of the attempts to use the same monomer with any of the symmetries from the 3 crytal structures. The dry costs have gone up, as have the soft clashes, and the hbonds have gone down. Oops, I accidentally deleted it while running it through gromacs (forgot to gzip it before running make). I re-created it and ran it through gromacs again. Sun Feb 6 06:19:55 PST 2005 Kevin Karplus gromacs destroyed the 5mer anyway, because I forgot to unpack the multimer into separate chains, so gromacs tried to close the large breaks between chains. To fix the problem, I did make decoys/5mer-5mer.try6-opt2.repack-nonPC.from-dimer.unpack.gromacs4.pdb.gz in 5mer/ This version scores almost as well as 5mer-5mer.try6-opt2.repack-nonPC.from-dimer, but has introduced 17 new breaks in order to reduce clashes. Sun Feb 6 16:15:19 PST 2005 Kevin Karplus I reoptimized from the 5mer-5mer.try6... to get 5mer.try7-opt2, then from that and 5mer.try6-opt2 to get 5mer.try8-opt2. I took the first two momoners from each and repacked to get dimer/decoys/dimer-5mer.try7-opt2.repack-nonPC.pdb and dimer/decoys/dimer-5mer.try8-opt2.repack-nonPC.pdb The dimer/decoys/dimer-5mer.try7-opt2.repack-nonPC.pdb is Rosetta's favorite so far---even beating the previous runs in ../redo-casp-method. I made chimeric monomers from both the try7 and try8 repacked dimers, and 5mers from the monomers. Mon Feb 7 09:35:08 PST 2005 Kevin Karplus The 5mer/try9 run failed because of an undertaker bug, but after creating try9-opt1, so I'll try adding some bug fixes and extra assertions to undertaker and run try10 including the try9-opt1 pdb file as a starting point. Mon Feb 7 12:59:18 PST 2005 Kevin Karplus The 5mer/try10 run gets a good score with try10.costfcn (beating try8), but the symmetry does not seem to have been preserved perfectly (probably an undertaker bug). I resymmetrized each of the monomers (and the 5 chimeras from the interfaces), and got only small differences, but 5mer.try10-opt2B seems to be the best, saving a little bit on the breaks. (The extra break seems to be in chain A, in the part that interfaces with chain E---before W55). I could either try polishing up try10-opt2B in undertaker, or do another Rosetta interface optimization to see if I can get Rosetta and undertaker to agree on model. I think I'll do the Rosetta run first. Mon Feb 7 13:11:06 PST 2005 Kevin Karplus dimer/decoys/dimer-5mer.try10-opt2B.repack-nonPC.pdb gets the best Rosetta energy so far (finally getting into the negative numbers). Let's see how undertaker likes a symmetrized chimera. Mon Feb 7 13:22:40 PST 2005 Kevin Karplus 5mer/decoys/5mer-5mer.try10-opt2B.repack-nonPC.from-dimer.pdb is the best of the Rosetta-reoptimized pentamers. Undertaker sees it as worse in all components of the cost function than try10-opt2B, but not by much, so we may finally be getting to the point where the cost functions disagree in detail, and not just on the weighting of different terms. There is a break before W679 and before W887 in 5mer-5mer.try10-opt2B.repack-nonPC.from-dimer.pdb.gz that is not present in the other monomers, but I don't know where it comes from. There is only one break reported in the chimeric monomer: Conformation[1] ACH7_SHORT.5mer.try10-opt2B.repack-nonPC.from-dimer.pdb.gz has 1 breaks ACH7_SHORT.5mer.try10-opt2B.repack-nonPC.from-dimer.pdb.gz breaks before (ACH7_SHORT)G172 with cost 0.160981 Could it be from accumulated round-off error in repeated trasformation? The breaks seems a bit big for that. Mon Feb 7 13:47:20 PST 2005 Kevin Karplus I re-created the 5mer-5mer.try10-opt2B.repack-nonPC.from-dimer.pdb.gz file and saw no bugs when creating it. I looked at the file in rasmol, and the bond before W887 looks just fine---the length (1.327) is within roundoff error of the ideal leangth (1.328). I'll try redoing the scoring, but I suspect that there is subtle bug in the scoring by undertaker or in reading in the files, and it may be difficult for me to track down. Tue Feb 8 09:04:35 PST 2005 Kevin Karplus I added a PrintConformBreakCost command to undertaker, to see why W887 was unusual. It seems that the omega torsion angle is very clost to the limit undertaker imposes for non-planarity of peptide planes (160 degrees), and so roundoff error can move the different monomers slightly above or below the threshold. I'll need to look at the possibility of including a peptide-quality cost function, to penalize the non-planarity in a more continuous way. Tue Feb 8 17:32:56 PST 2005 Kevin Karplus I added a bad_peptide cost function to undertaker, and have added that cost function to 5mer/try11.costfcn for optimization. The models I've constructed so far do seem to have worse than normal peptide planes (mean value is about 0.337 for the bad_peptide cost, while x-ray models average about 0.15. I think that the focus on clashes and breaks has allowed the models to torque the peptide planes more than is healthy. I'm wondering if I should reoptimize from a long way back in the process, when the peptide planes were cleaner. Unfortunately, that looks like backing up to try1-opt2 or try2-opt1 (try1-opt2 is the best scoring of the models with bad_peptide<0.3). Wed Feb 9 09:59:01 PST 2005 Kevin Karplus I added a HealPeptide operator to undertaker, using the same PatchBackbone technique as for HealGap, but selecting the peptide to apply to based on the bad_peptide cost, rather than choosing a break. I am applying this to the monomeric model, starting with early models that had relatively low bad_peptide costs (try3-opt2, try3-opt2.repacck-nonPC, 5mer-try1-opt1, try1-opt2, try1-opt2.repack-nonPC). Wed Feb 9 11:29:13 PST 2005 Kevin Karplus The HealPeptide operator gets called a lot, and the try12-opt2 monomer has a bad_peptide score in the normal range for a chain. Running gromacs raises the bad_peptide score a lot, so it seems that gromacs does not have much penalty for tweaking the omega torsion angles. I'll try doing another polishing run on the monomer for try13, to see if I can get the remaining breaks and clashes reduced enough, then I'll try this monomer in the 5mer. I suspect that it will be quite different, since the loop in the center that has R99 is positioned differently. I may have to mash it down to avoid conflicts in the center of the ion channel. Wed Feb 9 18:22:03 PST 2005 Kevin Karplus Foo, I just got a chance to look at try12-opt2 and try13-opt2 and I realized that they are moving the helix. That shouldn't have come as a surprise, since that was the problem with try3-opt2 also, and they were mainly based on try3-opt2. Perhaps I need to try again with the helix pinned, or start from the try5-try3-chimera. Wed Feb 9 18:49:42 PST 2005 Kevin Karplus For try14 I tried pinning the helix with Hbond Q4.NE2 G74.O and I put in sheet constraints from try11-opt2 and try13-opt2, removing the conflicting constraints for the strand around F100. I started from a number of models that scored close to try13-opt2, in the hopes of getting productive crossover: ReadConformPDB ACH7_SHORT.try9-opt2.repack-nonPC.pdb ReadConformPDB ACH7_SHORT.try8-opt2.pdb ReadConformPDB ACH7_SHORT.try13-opt2.pdb ReadConformPDB ACH7_SHORT.try11-opt2.repack-nonPC.pdb ReadConformPDB ACH7_SHORT.try10-opt2.repack-nonPC.pdb ReadConformPDB ACH7_SHORT.try12-opt2.pdb Thu Feb 10 13:17:55 PST 2005 Kevin Karplus try14-opt2 looked ok, so I did a polishing run starting from all models, which mainly polished try14-opt2. I then made 5mers of the try14-opt2 monomer, which (naturally) had a lot of clashes, scoring about as badly as the 5mers made from the try11-opt2 monomer with the 5mer/try11.costfcn. I'll start a 5mer/try12 run from just the 5mer-try14-opt2 starting points, using the same cost function as for 5mer/try11. I'm a little worried that undertaker will simply push the monomers apart to remove the clashes---I may have to do the optimization gradually raising the cost of clashes, so that the model can reshape itself to fit. Thu Feb 10 16:38:52 PST 2005 Kevin Karplus 5mer/decoys/5mer.try12-opt1.pdb is showing the results of doing TweakMultimer too soon---the pentamer has spread apart enough to avoid clashes, but is not compact enough to really provide a good basis for residing the multimeric interface. I have a few routes I can try: 1) reoptimizing from the same starting pentamers, but with the TweakMulitmer operator turned off for the first round, to try to keep the pentamer more tightly together. Also make the soft-clashes penalty ramp up gradually. 2) create some new pentamers, using the same monomer, but with different symmetry operations (perhaps copying the symmetries from 5mer.try11-opt2.pdb) 3) create new pentamers using the monomer from 5mer.try12-opt2, but with symmetries from other runs. Perhaps the cleanest thing to do would be to create a few new starting pentamersk that are fairly densely packed, and do the optimization from them with a ramped up soft-clashes penalty. Fri Feb 11 15:47:08 PST 2005 Kevin Karplus I created some new pentamers: 5mer.monomer-try15-opt2-on-try11-opt2.pdb 5mer.monomer-try14-opt2-on-try11-opt2.pdb 5mer.try12-opt2A-on-try11-opt2.pdb 5mer.try12-opt2B-on-try11-opt2.pdb 5mer.try12-opt2C-on-try11-opt2.pdb 5mer.try12-opt2D-on-try11-opt2.pdb 5mer.try12-opt2E-on-try11-opt2.pdb Then I did an optimization starting with them that gradually ramped up the weights for clashes and breaks (5mer/try13.under) In 5mer/ I made homogenize10-try13-r4 to get 10 different monomers symmetrized. The 5mer.try13-r4A multimer scores almost identically with try13-r4. These score a bit bettter than the old 5mer.try10-opt2B, but most of the improvement comes in the bad_peptide score and the sidechain score. Soft-clashes, break, andphobic_fits are all worse. The dry scores are comparable. I wonder if phobic_fit is misleading, since in the pentameric structure the hydrophobic residues form a torus, rather than a sphere. Maybe I should switch to using just radius of gyration (or radius_fit). There may be poor fit there also, because of the torus, but probably not as bad. I should probably make a dimer and repack with Rosetta, to see if I have something that Rosetta likes. Sat Feb 12 20:06:28 PST 2005 Kevin Karplus I added a new command to undertaker "TweakPeptide" to undertaker, which tries rotating the peptide planes without moving the C_alpha atoms. It seemed to be working fairly well, but I did get a crash, because I forgot to include "recompute_edge_lengths". Sun Feb 13 09:17:55 PST 2005 Kevin Karplus 5mer/try14 managed to reduce the n_ca_c and soft_clashes costs of try13-r4, without changing break or bad_peptide much. Hbonds got a tiny bit better. Unfortunately, both try13 and try14 suffer from having the pentamer spread out just a little bit, so the interface is loose. I'd need to increase radius_norm weight (and maybe dry8) by quite a bit to get tighter packing. Perhaps I should take the try14-opt2 monomer and place it according to the try11-opt2 symmetries or event the tighter try8-opt or try6-opt2 symmetries, then reoptimize without letting TweakMultimer do much. Tue Feb 15 15:49:42 PST 2005 Kevin Karplus 5mer/decoys/5mer.try15-r4.pdb does not pack the multimer as tight as any of the models it started from (which, admittedly, had bad clashes). All of them are more loosely packed than even the loosest of the models built the symmetries in the crystal structures. I've got to pack this tighter if I'm going to do design! Thu Feb 17 13:51:27 PST 2005 Kevin Karplus The 5mer/try16 run seems to have been moderately successful in packing the multimer tighter. I'll do a 5mer/try17 run with soft-clashes turned up a bit (but still having the 5mer/try16-opt2 score best) to see if we can get a better tight fit. Thu Feb 17 21:31:50 PST 2005 Kevin Karplus The 5mer/try17 run is looking pretty good, but there are probably too many breaks and clashes---when the dimer is fed to Rosetta for repacking, it really hates the result, seeing many bad clashes. Fri Feb 18 11:42:01 PST 2005 Kevin Karplus The really bad clash for Rosetta is W149:A with L119:B, which had an W149.NE1-L119.O Hbond in the try17-opt2 model, that Rosetta really mangled. Perhaps there is some over-packing in this region, and Rosetta sees all tryptophan rotamers as bad. F104 (which Rosetta also sees as having bad clashes) is in this neighborhood also. Another bad clash is E98:A with I123:B---the E98 residue looks a bit cramped in the try17-opt2 model. Perhaps there should be a hydrogen bond between E98.OE1:A and I123.N:B? T103 (the next worse according to Rosetta) is between E98 and F104, so is part of the same bad region. Thu Mar 3 15:00:03 PST 2005 Kevin Karplus 5mer/decoys/5mer.try18-opt2.pdb reduces the clashes somewhat from try17, while keeping the packing tight. Gromacs will reduce the clashes a lot (roughly 3-fold), but at the cost of messing up the peptide planes and the n-ca-c bond angles (not to mention the vicinal disulfide). The try18-opt2 model scores best with costfcns try18, try17, and try16. It is beaten by try15-r4 in try15, and by try15-r4, try15-r3, try15-r2, and try14-opt2 in try14. Earlier costfcns, which put less emphasis on compactness and more on clashes score it poorly. Perhaps I should try optimizing with undertaker from the gromacs-optimized models, trying to fix the bad peptides while retaining the low-clash features. Thu Mar 3 21:11:55 PST 2005 Kevin Karplus 5mer/decoys/5mer.try19-opt2.pdb was obtained by optimizing the gromacs-optimized files (mainly try18-opt2.unpack.gromacs4) using undertaker. The cost using try19.costfcn was reduced to substantially less than what try18-opt2 cost. Thu Mar 3 22:22:55 PST 2005 Kevin Karplus Running gromacs on try19-opt2 does reduce clashes further, but at considerable additional cost in breaks, bad N_CA_C bond angles, and bad peptide planes. Giving the first two chains of try19-opt2 to rosetta for repacking results in a better rosetta score than try17-opt2, but not as good as the loosely packed try7-opt2 and try10-opt2B. Other than the vicinal disulfides, which rosetta will never like (unless there have been some code changes since the version I'm running), the biggest conflicts are for Y168:B, W55:B, E98:A, F104:B, T103:B, I123:B, R79:B, F104:A, P121:B, which are almost all copied into the monomeric chimera. The 5mer created from the chimera made from the rosetta repacking is 5mer/decoys/5mer-5mer.try19-opt2.repack-nonPC.from-dimer.pdb It scores better with the try19.costfcn than other rosetta repackings, but when compared with try19-opt2 has substantially increased costs for dry5, dry6.5, dry8, and soft_clashes, but somewhat reduced sidechain costs. It is a little worse on non-backbone hbonds. Sun May 1 09:47:28 PDT 2005 Kevin Karplus I've not worked on this project for a while, mainly because I've been too busy, but also because I was getting frustrated at not being able to get a homology model that made a tight enough pentamer to be worth trying to design from. I put together a new cost function today using new clash detection and alpha histograms from the dunbrack-2191 set, a larger training set than I've been using. try19-opt2 is still the best-scoring model, but it may be worth doing a new optimization round.