Wed Jun 16 16:08:34 PDT 2004 T0207 DUE 30 Jun 2004 Wed Jun 16 16:42:43 PDT 2004 Kevin Karplus We have a very tight deadline on this prediction. Luckily, it is a short protein and it looks like we have very strong hits to domain c.47.1.1, so we may not have to work too hard on this one. The target searches all got strong scores, but the template searches had much more modest scores. Luckily, the best scores with the template library were skill to domain c.47.1.1, so we still have confirmation of the superfamily. Wed Jun 16 17:05:06 PDT 2004 Kevin Karplus The alignments to the various templates are all full-length and very similar in the aligned regions. There are gaps in the turns that undertaker will try to patch, but I suspect that using enough templates will allow us to mix and match fairly freely, so that almost no ab-initio prediction is needed. Wed Jun 16 17:45:20 PDT 2004 Kevin Karplus There are some small breaks and a little foaminess to try1, but it is basically ok. Let me try doing another run, with breaks and soft_clashes turned up and more alignments, to try to jiggle things a little tighter. Wed Jun 16 22:04:25 PDT 2004 Kevin Karplus Things haven't moved much in try2, but they look a little better to me. There is still a bit of foaminess as undertaker is trying a bit too hard to avoid clashes. I'll do one more run from the alignments with break weights turned up a bit, then try a polishing run from all the models so far. After that, I might try seeing what Rosetta thinks of the model, both as is and after repacking the side chains. For try3, I'll also turn off SCWRLing the initial alignments, so that we don't get stuck in SCWRL's idea of where the sidechains should go. Thu Jun 17 06:01:53 PDT 2004 Kevin Karplus try3 did not score as well as try2---the initial scwrling does seem to help. (Either that or the strand and helix constraints help.) I just noticed that I had not included the C11-C14 disulfide in the scoring. I'll add that for try4 and rerun (with strand and helix constraints, but still not scwrling the start) Thu Jun 17 11:35:21 PDT 2004 Kevin Karplus try4 pulled apart the hbonds in the middle, in an attempt to reduce clashes, resulting in a roughly right but foamy model. Thu Jun 17 15:29:14 PDT 2004 Kevin Karplus try5 is much nicer---looking almost like a real protein. There are still a few clashes and the protein is not packed quite as tight as the real ones---a little too much space between the helices and a little too much between the two hairpins. Now, the question is how I squish the protein just a bit tighter. Should I up the dry5 and dry6.5 terms a lot? Should I identify some knobs and holes that fit together and add constraints? Should I repack sidechains with Rosetta then reoptimize? Thu Jun 17 16:32:15 PDT 2004 Kevin Karplus I fixed up a script and some Make.main targets to make T0207.try5-opt2.repack-nonPC.pdb Rosetta was unable to remove some clashes, ending up with bad repulsive energies for several residues: 72 LEU -4.5 111.6 1.7 0.0 -0.2 3.2 1.9 -0.8 0.0 -0.1 0.0 0.0 113.1 76 PRO -3.6 114.0 1.1 0.0 0.1 0.2 0.0 0.0 0.0 0.0 0.0 0.0 111.8 11 CYS -3.0 65.9 1.9 0.0 0.0 1.0 0.0 -0.8 0.0 -1.7 0.0 0.0 66.7 14 CYS -4.4 65.6 2.7 0.0 0.3 0.5 0.0 -0.8 0.0 -1.7 0.0 0.0 65.6 6 TYR -7.0 60.6 2.6 0.0 0.0 2.2 0.5 0.0 0.0 -0.8 0.0 0.0 59.6 49 PRO -4.3 32.8 1.1 0.0 -0.8 0.1 0.0 -0.7 0.0 0.0 0.0 0.0 28.1 40 LEU -4.3 21.5 1.5 0.0 -0.2 3.1 1.7 -0.8 0.0 -0.1 0.0 0.0 22.7 28 ALA -1.4 14.9 0.8 0.0 0.9 -0.0 0.0 0.0 0.0 0.1 0.0 0.0 15.1 61 TRP -4.1 11.5 2.0 0.0 1.7 1.9 0.2 -0.7 0.0 -1.8 0.0 0.0 14.2 4 THR -3.4 12.2 2.0 0.0 0.1 0.9 0.0 -0.2 0.0 0.4 0.0 0.0 11.0 51 LEU -4.0 10.9 1.3 0.0 -0.2 2.4 0.3 -1.5 0.0 -0.1 0.0 0.0 9.2 60 ASP -2.7 9.0 1.9 0.0 0.7 0.7 0.0 0.0 0.0 0.8 0.0 0.0 8.9 26 ALA -1.4 9.2 0.6 0.0 -0.1 -0.0 0.0 0.0 0.0 0.1 0.0 0.0 8.3 Undertaker doesn't like this repacked version very much, but let me try doing an optimization run from it, to see what undertaker comes up with. I'll also try increasing the "dry" terms and clash weights slightly. Thu Jun 17 20:37:57 PDT 2004 Kevin Karplus Even with the try6 cost function, try5 scores a bit better. We seem to have lost some of the Hbonds in the center. I'll strengthen the constraints for try7, and start from all the decent models. As usual, undertaker does not care much for Rosetta's repacking, but the try5 repacking is still preferred to try6's. Fri Jun 18 07:05:32 PDT 2004 Kevin Karplus try7-opt2 seems to be better packed than earlier attempts. When Rosetta repacks it, it seems to me do a better job, though undertaker doesn't like it as much. Perhaps I need to increase soft_clashes a bit, though it would take an increase in weight of about 160 for undertaker to like it as much. One problem is that undertaker likes a rather crumpled arginine sidechain, while Rosetta prefers a straight one. Another is that the crumpled arginines put more atoms in the "dry" spheres that undertaker is using to encourage tighter packing, so undertaker sees Rosetta's packing as less dense. There are still bad backbone clashes (M55.O-A75.O and L46.O-R47.O) and L72.O clashes with P76.CD, which Rosetta can't fix, since we're not letting it touch the prolines. There are also two breaks, before G27 and E58. I'll let undertaker try reoptimizing once more, using a modified cost function that increases the clash and break costs and decrease the "dry" and "sidechain" costs. I'll also fix the Hbond constraints, since some of the Hbonds I specified in earlier runs seem to be bogus. With any luck try8 will be the final run, and we'll submit try8-opt2 and try8-opt2.repack-nonPC as our best models, and 3 unpolished alignment-based models as our runners-up. Fri Jun 18 09:26:48 PDT 2004 Kevin Karplus Interestingly, try7 was built mainly from fragments, not from previous models as I had intended---I'd mistyped all the file names for the ReadConformPDB. I caught the problem after try8 had run for a while, so I killed and restarted try8 with the file names fixed. Now try8 is very quickly fixing breaks and reducing clashes, since it no longer has to start from scratch! It would be misleading to claim parents for try7, since they were built mainly from fragments, not alignments. I suppose I could include all the top hits, which provide the longer specific fragments, as the parents. Fri Jun 18 13:14:42 PDT 2004 Kevin Karplus try8-opt2 has smaller breaks than try7-opt2, but still has two breaks: before R8 and E58. (Interestingly, try8-opt1 only has the break before E58.) The only bad clash that Rosetta sees after repacking is W61-P62. This looks like the W61.CB atom with the P62.CD atom at 2.826 Angstrom, which has a limit of 2.70993 for a neighbor bump, so undertaker doesn't see it as a problem. I'm going to increase the limit to 3.05 (the smallest other CB-CD distance). try8 seems to have almost exactly the same packing density as try7, and the Rosetta repacking doesn't really seem to be making improvements. I don't like the buried R57, but if I make a salt bridge D53.OD2-R57.NH2 D53.OD1-R75.HN1, the buried charge is counteracted. I can find corresponding residues in 1nw2A, but they DON'T form a salt bridge there---instead the lysine (1nw2A:K85) hbonds to the final helix (1nw2A:Q99.O and Q99.OE1), and the sheet twists to expose 1nw2A:E82. The sheet twisting is harder here, since we have a 4 residue separation, not a 3 residue one. Fri Jun 18 20:46:43 PDT 2004 Kevin Karplus try9-opt2 is the new best-scoring model, but try9-opt2.repack-nonPC does not score as well as try8-opt2.repack-nonPC. The try9 models have essentially the same breaks as the try8 models. The try9 models do manage to form the salt bridge I was trying for. But Rosetta removes the saltbridge when repacking, preferring to move R57 out to the solvent and leave D53 buried. Perhaps as a second option, I should try to form a hydrogen bond to the terminal helix---perhaps to A74.O, which seems to be about the right distance away and doesn't have an Hbond. I'll try that for try10. If I turn off the constraints in the cost function, try9-opt2 currently scores the best, so I'll probably go with it as our number-1 choice. (Better yet, I'll run try11 optimizing existing models without the constraints and see what happens.) Sat Jun 19 07:33:31 PDT 2004 Kevin Karplus With no constraints in the scoring function, try10-opt2 and try11-opt2 cost almost the same, and only very slightly better than try9-opt2. Neither try10 nor try11 makes a salt-bridge. Try10 buries just D53, while try 11 buries both D53 and D57. Try10 and try11 have essentially the same breaks as try9. Failure to make the salt bridge between D53 and R57 did not fix up the break between R57 and E58. If I use a costfunction that include the SSbond and the salt bridge, but no other constraints, the best score is for try9-opt2. Interestingly, the salt bridge that formed was not the OD1-NH1, OD2-NH2 bridge I requested, but OD1-NH2, OD2-NH1. Perhaps I should reoptimize with that salt bridge requested, to get the bridge without bumping into the R57.CZ atom. I'll also do one more attempt at optimization, with the salt-bridge changed to D53.OD1-R57.NH2 D53.OD2-R57.NE, which may be a little less strained. Sat Jun 19 18:32:16 PDT 2004 Kevin Karplus Try12 produced the salt-bridge it was trying for, but try13 did not. Using the cost function for the salt-bridge that try12 was trying for, try12-opt2 is the best scorer. Rosetta's repacking seems to make clashes a bit worse (of course, Rosetta disagrees). Of the repacked models, Rosetta likes try9-opt2 best. Try12 does not look significantly better to me than try9---the backbones are almost identical and there are only minor changes in the rotamers. I am tempted to make try12-opt2 be my #1 model, try9-opt2.repack-nonPC be my #2 model, and use the remaining 3 models for the top 3 models from alignment (without fixing up their sidechains---just the initial sidechain replacement). I tried formatting these and sending them to CASP. We can replace them if anyone gets any great ideas before the deadline. I think I'm going to stop work on them now. Sun Jun 20 19:21:01 PDT 2004 Kevin Karplus Reformatted today with the improved scripts, and successfully submitted 5 models. Thu Jun 24 20:59:03 PDT 2004 Kevin Karplus Got news today that T0207 was leaked, so the structure is being removed from CASP6. Too bad---I was curious about how I did. Sun Sep 19 12:49:30 PDT 2004 Kevin Karplus try13-opt2 was 2.7821 Ang, but some of the earlier ones were better: try7-opt1-scwrl was 2.6683. For that matter try13-opt1-scwrl was 2.6834 and had the best (1.8197) CA_rmsd. But most of our models beat the robetta ones. SCWRL did surprisingly well on these models. Wed Sep 22 11:20:22 PDT 2004 Kevin Karplus On the GDT score, our best model was still try13-opt2 at 82.6667% The best robetta model was robetta3 at 68.6667% Fri Sep 24 17:12:08 PDT 2004 Kevin Karplus Looking at the GDT curves and the (revised) smooth_GDT scores, it is clear that our best two submitted models (model1 and model2) easily beat the robetta models, and our best model (try7-opt1.scwrl) was better still: name length missing_atoms rmsd rmsd_ca GDT smooth_GDT T0207.try7-opt1-scwrl 78 0.0000 2.6683 1.8235 -81.6667 -75.5420 model1.ts-submitted 78 0.0000 2.7580 1.8198 -78.3333 -72.8550 model2.ts-submitted 78 0.0000 2.7926 1.8201 -78.3333 -72.8499 model5.ts-submitted 78 23.0000 4.0689 3.4354 -67.0000 -63.4664 robetta-model3.pdb.gz 78 0.0000 3.7499 2.7034 -67.0000 -62.2137 robetta-model2.pdb.gz 78 0.0000 4.3215 3.1369 -61.6667 -57.8474 robetta-model1.pdb.gz 78 0.0000 4.7477 3.1844 -62.0000 -57.6046 robetta-model5.pdb.gz 78 0.0000 4.8915 4.1172 -59.0000 -53.7890 robetta-model4.pdb.gz 78 0.0000 5.2357 3.8212 -56.3333 -52.7316 model3.ts-submitted 78 67.0000 4.2224 3.4924 -57.6667 -52.5412 model4.ts-submitted 78 89.0000 5.3708 4.3864 -48.0000 -43.9989