Wed Jun 23 08:48:58 PDT 2004 T0215 Due 19 Aug Wed Jun 23 11:20:26 PDT 2004 Kevin Karplus Fold recognition? new fold? No strong hits. Wed Jun 23 16:24:29 PDT 2004 Kevin Karplus This seems to be new-fold on a 3-helix protein. The helices are easy to create, but packing them will be difficult. We can make the N-terminal HIS tag into an extended structure, but other than that, we'll probably have to rely on random packing. Since helices are so easy to form, maybe we can use just fragments and not alignments in try2 (though using the alignments to generate the specific fragment library). Wed Jul 21 19:31:32 PDT 2004 Kevin Karplus I'll try running try3 with a new (possibly buggy) version of undertaker that has "bonus" keywords allowed on constraints. Thu Jul 22 11:09:29 PDT 2004 Kevin Karplus The try3 run has finished, and scores well with try3 cost function. It also scores best with the unconstrained.costfcn. Try1, try2, and try3 are pretty much in agreement about the final two helices, but disagree substantially about what comes before that. I think I like try3 best, but I wonder if the HIS tag should be more exposed. I'd also like to pack it a bit tighter. Perhaps a polishing run would be useful. Sat Jul 31 05:57:27 PDT 2004 Kevin Karplus try4 looks fairly good, but should probably be packed tighter. Thu Aug 5 14:43:21 PDT 2004 Martina Koeva Have started try5 in an attempt to pack the structures a little tighter. constraints => 2 1 wet6.5 => 10 20 dry5 => 10 20 dry6.5 => 20 40 dry8 => 10 20 dry12 => 5 10 soft_clashes => 40 30 Thu Aug 12 16:19:09 PDT 2004 Martina Koeva I have not looked at this target for a week now. Try5-opt2 is scoring right now the best with the try5 cost function, as well as the unconstrained cost function. However, I am concerned that there is still a big hole in the middle (if one were to look at the protein in spacefill). Maybe I should try and introduce an additional constraint in an attempt to pull the helices a little closer. ---------------------------------------------------------------------------- Date: Thu, 12 Aug 2004 16:52:46 -0700 From: Kevin Karplus To: martina@soe.ucsc.edu Cc: karplus@soe.ucsc.edu Subject: Re: Question related to T0215 On a well-polished model, it is often quite difficult to get anything to move, even to close a big hole, since moving things will almost certainly create a gap or a clash, at least for a while. This creates an activation barrier that makes it difficult to move to a lower-energy conformation. Sometimes it is easier to get things to move by adding constraints and rerunning from alignments. Having high-energy (bad) conformations makes it easier to pass the activation barrier and arrive at a better minimum. ----------------------------------------------------------------------------- Sat Aug 14 01:35:56 PDT 2004 Martina Koeva So I went back to alignments, which we only had in try1 and added a few hand-picked constraints in an attempt to bring the helices closer together. I have select leucines on the inside of the helices, since those should be buried within the core of the protein. In particular I have chose Leu26,Leu54 and Leu66. The constraints have been added to a try6 cost function. I have raised back soft_clashes to 40. Try6 is running on meow. For try7 I will go back to optimizing from existing models with the hope that try6 models will have an effect and would allow undertaker to successfully close the gap in the core of the protein. Sat Aug 14 15:15:15 PDT 2004 Martina Koeva As expected try6 is scoring quite a bit worse than try5 even with the try6 cost function. It has formed a nice-looking hairpin, except that we do not predict any strands in this structure. For try7 I am going back to starting from the existing structure, hoping that the try6 models will have an effect and help with getting to a better energy minimum. Sat Aug 14 21:43:06 PDT 2004 Martina Koeva Try7-opt2 is back to pretty much back to where try5-opt2 was. The hole in between the three helices is still there. Moreover, try5-opt2 scores a little bit better than try7-opt2 with the try7 cost function. I will have to go back to alignments, introduce some more/different constraints, make try8 run from alignments again and go back in try9 to starting from models with the hope that try8 would help undertaker take the structure out of this local energy minimum. Sun Aug 15 15:08:25 PDT 2004 Martina Koeva Before I go back to alignments I will attempt to start over from existing models, fixing the weights and distances on the constraints. These changes are reflected in try8.costfcn. Try8 running on hoot. Sun Aug 15 21:14:16 PDT 2004 Kevin Karplus try8 finished and scores a tiny bit better than try5 on the unconstrained costfcn. The differences are a bit bigger on the try8 costfcn, because the weak constraints make some difference. I think we have gone past the point of diminishing returns with this optimization, since the sidechain repacking by Rosetta makes a much larger difference than any backbone changes we're making. Unless Martina has some clever ideas for how to get a different structure, I think it is time to choose the best models and submit. Mon Aug 16 15:55:07 PDT 2004 Martina Koeva Here is a list of proposed submission candidates: try8-opt2 (best score with unconstrained function) try5-opt2.repack-nonPC (lowest Rosetta energy) Note: negative scores? try1-opt2 (fully automated) try6-opt2 (scores quite bad; but forms a hairpin, with an additional strand at the N-terminus; it's an interesting structure, even if not very believable) T0215-1lqlA-t2k-global-adpstyle5 (model 1 from undertaker-align.pdb) Mon Aug 16 17:11:51 PDT 2004 Kevin Karplus I accept Martina's list and will submit them, though changing the order to put the full auto in model 4, since that is more in keeping with what I've done before (even though try6 is really ugly). Thu Nov 18 23:05:28 PST 2004 Martina Koeva Based on the smooth gdt scores: best sam-t04 45.7959 (the Rosetta repacked version of our top model try8-opt2) best submit 45.7920 (model1) model1 45.7920 auto 38.3845 align 28.4919 robetta best 45.4691 (robetta model9) robetta1 42.2667 Fri Nov 26 07:49:07 PST 2004 Kevin Karplus The secondary structure is right, and the helices are packed more or less right, though the C-terminal helix is not packed properly into the bundle. We beat robetta on smooth GDT, but robetta 9,4,8,1 beat us on all-atom rmsd. Robetta-9 matches us on straight GDT. There were too few homologs to do contact prediction and the hand-added packing constraints did not help---the added constraints were wrong. Unconstrained costfcn does decent job of ordering decoys.