Wed Aug 11 10:40:12 PDT 2004 T0273 DUE 28 Aug 2004 Wed Aug 11 12:48:19 PDT 2004 Kevin Karplus The t2k and t04 predictions don't agree. We'll probably have to treat this as new-fold predictions. Wed Aug 11 17:22:17 PDT 2004 Kevin Karplus try1-opt2 came up with a fold prediction which is in moderate agreement with the secondary structure predictions. If we like the fold, we can probably slide the alignment around a bit to get better helix and strand matches. For example, A113-V117 could be made to go where Y134-N137 are now, cutting out an extra helix and getting a better fit of sheet and helix. This would require careful editing of the strand constraints. Wed Aug 18 17:20:54 2004 George Shackelford In theory both I and Bret are working on this prediction. But theory often fails in this line of work. This seems to be trying to be a barrel but that may mean its a sandwich. I think sandwich. There are a few decent RR constraints as well. I'm going to do a second run and exclude t2k constraints and include some of the RR constraints and see what happens. Either the 'barrel' gets better or not. Again I needed to replace the basic cost values from the starter-directory. try2 running on cluck(?) Thu Aug 19 11:04:26 2004 George Shackelford This baby has barrel on the brain. Now I did notes at home but my connection is down. I'll have to start over again from school. I don't believe the barrel and I'm going to define a few sheets to break it up and do something more natural. Strands s1 27 -30 s2 35 -38 s3a 67 -70 (?) s3b 74 -77 (?) s4 86 -92 s5 112-118 s6 140-146 s7 151-156 s8a 162-165 (?) s8b 169-172 s9 180-183 (?) Weak stuff in several places. I'm going for sheets s1^vs2 s4||s5 (or anti?) s5||s6 s6^vs7 Try3 running somewhere. Fri Aug 20 01:06:31 2004 George Shackelford So try3 didn't do too well. It failed to form the requested sheets. I didn't specify hbonds and it didn't offer any. For try4 I'll specify some hbonds and hope they'll work. SheetConstraint Q27 S31 V38 G34 hbond S31 10.0 # s1^vs2 SheetConstraint D86 V90 A113 V117 hbond F89 10.0 # s4||s5 SheetConstraint A113 V117 L141 V145 hbond L144 10.0 # s5||s6 SheetConstraint L141 V145 F155 A151 hbond V145 10.0 # s6^vs7 Try4 running on peep. Fri Aug 20 14:34:11 2004 George Shackelford Damm barrel! I'm cranking up the first and fourth sheets to get some antiparallels. Where are those helices coming from?? I've looked around in the constraints files and the sections that show up as strands should be strands, not helices! There must be an alignment that is screwing (or fixing) up the works. I'll do a try5 with those tighter constraints. Now running on whinny. Fri Aug 20 22:50:48 2004 George Shackelford Well that broke up the barrel, but it left me with little else. I'm going to copy some sheets from a couple of Robetta models I like: model1 and model10 (because it scores well). Sat Aug 21 16:21:53 2004 George Shackelford Working with robetta model1, I adjusted a couple of sheets and added another. SheetConstraint Q27 S31 V38 G34 hbond S31 30.0 # s1^vs2 SheetConstraint D86 V90 A113 V117 hbond F89 10.0 # s4||s5 SheetConstraint V114 V117 L141 L144 hbond V143 10.0 # s5||s6 SheetConstraint L141 D146 R156 A151 hbond L144 30.0 # s6^vs7 SheetConstraint Q27 S31 V38 G34 hbond S31 10.0 # SheetConstraint G66 F69 F75 ?? Now to start try6 on crow(?) Sat Aug 21 22:55:01 2004 George Shackelford Try6 is mangled again by the 'barrel.' I have researched the best hits list and I have found that 1qwg is a barrel! I asked Kevin how I could leave it out - Sure, you exclude templates by not including them. That is, instead of using all-align.a2m, use the alternative syntax in the .under files of using "include read-alignment-scwrl.under" lines. (See any of the .under files, it is included in the try1.under files in a commented-out form.) You may need to run "make read_alignments" to make sure you have all the read-alignment-scwrl.under files you need. Also you have to have the list of templates in the MANUAL_TOP_HITS list. Then do the make read_alignments. The includes follow directly. I also commented out the 'ReadFragmentAlignment' line that contains the all.align.a2m. Now to see what I can get - Running try7 on crow. Sun Aug 22 09:48:59 2004 George Shackelford Try7 still is going for the barrel. I'm going to take out 1ezw and 1m5w which I have noted may cause problems. I also boosted the anti-parallel sheets and boosted the beta hbond cost factor. Try8 running on crow. Sun Aug 22 12:42:12 2004 George Shackelford Ouch! that sure 'fixed' things. Now we need to get some helices back. Time to back off and retry. At least we aren't getting any barrels. Turning down the hbond_geom settings to normal. Readjusting cost of SheetConstraints. Upping cost of break. Try9 running on caw Sun Aug 22 15:32:27 2004 George Shackelford Better but there are still some sections that aren't making sense. 47-60 is a consensus helix but shows up as part of a sheet. Need to stop that. I did a unconstrained scoring and I'm doing worse and worse. Only try6, try3, and try5 score better than Robetta model 6. I am going to take another shot at getting better results out of what we have on try9. I've got to make it work because I don't like the barrel. I've got new sheets in place and try10 is ready to go. I have included t04.ehl2 and I've boosted the costs for its helices and strands. try10 running on crow. Mon Aug 23 22:32:15 2004 George Shackelford Kevin suggested at the meeting that I should try using the sheets provided by robetta model 6 (which look nice) and work with try3 as a start (because it scores well) and see what I can get. Adjust the cost function to control what works. The robetta sheets don't cover everything; I'm going to supplement them with some of my others. The hbonds I have must be wrong, though. Try12 running on crow. Tue Aug 24 10:07:08 2004 George Shackelford Try12 scores the best using unconstrained. I am going to 'adjust' the costfcn again, using try3 and try12 as inputs. Up goes the constraints cost and up goes the hbond_geom costs. Try13 running on crow. Tue Aug 24 17:07:47 2004 George Shackelford Try13 hasn't changed. I'm going to REALLY crank up the constraints cost and try 14. Try14 running on crow. Tue Aug 24 23:52:28 2004 George Shackelford Ok, I THINK try14 is getting somewhere. I've adjusted the constraints to match some sheets and I've added another call to get that parallel s6||s7 going. It looks like a natural to me. Try15 going on crow. Wed Aug 25 13:31:35 2004 George Shackelford Try15 doesn't yet cut it. I then started a tighter try16 this morning and that seemed a bit better. Then as I was analyzing it, I noticed that I hadn't updated the included alignment; it was still set to try13. So try17 is now focusing on improving try16. I have great expectations for try17 - now running on ribbit. Wed Aug 25 17:44:58 2004 George Shackelford Ok, I surrender. Now I'm going to see if I can polish try17 enough to get a better score that try13! drop constraints cost raise break costs. Start getting it to get together. Still leaving the requested sheet constraints in place. Try18 on peep. Thu Aug 26 10:18:15 2004 George Shackelford Try18 didn't seem to work any better. I've decided to go ahead and do a general polishing; I've removed all the constraints and cranked up the break cost. That had better make a change. This is suppose to be getting ready for release on Sat! Try19 running on ribbit. Thu Aug 26 11:44:59 PDT 2004 Kevin Karplus The best scorers on the unconstrained costfcn are try13-opt2 try12-opt2 try3-opt2 all of which are slight variants of the same somewhat mangled sandwich. The next best is try2-opt2 which is the dreaded barrel (also appearing in try1-opt2). I'm going to try directly polishing robetta-models (currently model7 scores best). The try20 cost function currently favors try13-opt2, and of the robetta models likes best robetta-model7. Thu Aug 26 13:34:18 2004 George Shackelford While robetta model7 does score well, I like model6 because it buries some strands that need to be buried. If he can just polish a robetta model, then I'm going to polish robetta model6! So there! NO! I've just looked at the results of try20 and I see the future - it is try20! It already scores the best in unconstrained. It's very foamy so I'll apply the usual polishing pressure on it. I'm using my try21 to do it. try21 running on ribbit. Thu Aug 26 15:55:18 PDT 2004 Kevin Karplus I agree that try20 looks the best of any we've seen. It is heavily based on robetta-model7, but it does seem to meet some of the rr constraints also. In general, it does better with constraints than any of the robetta models. It does better than robetta-model7 on all terms except breaks and the general hbond term. I think that George cranked up the dry6.5 weight too high on try21---it isn't THAT good a measure. It doesn't seem to have made much difference though: try21-opt1 has only tiny rotamer changes from try20-opt2. More useful might be to attach strand P85-E91 to one of the sheets, perhaps antiparallel to P161-G166 85> PDAAFVE GELRLPP < 160 I'll try setting this up as try22. Thu Aug 26 16:44:30 2004 George Shackelford Try21-opt2 is out and scores better. There are a couple of changes I'd like to make. Strand s6 (113-117) is now close to strand s7 (s6||s7). It's in the right place just not forming. The strand P85-E91 probably would work better if it were also part of that sheet and also parallel (s5||s6). P161-G166 is predicted to be more exposed. If the helix 44-64 could be turn slightly, it's predicted buried face could look away from the surface and not be exposed as much. Perhaps better than that would be to move the sheet 62-76 over to cover it and close up the gap in the surface of the protein. There's a loop from 166-186 that just hangs there looking ugly. Should do something with that. What I'm going to do is two different runs. Both will attempt the s6||s7 but only the second will attempt the s5||s6. Try23 running on ribbit. Try24 running on peep. Thu Aug 26 21:27:18 PDT 2004 Kevin Karplus Best unconstrained costs are for try25-opt2 try24-opt2 try23-opt2 try21-opt2 which are all essentially the same model---the differences are extremely small. In avoiding the rip-the-model-apart problem, George ran into the do-nothing-to-the-model problem. Thu Aug 26 21:42:06 2004 George Shackelford I had done a quick run of try25 with higher(?) constraints but no go. I seemed to have lost some info because my home system has been unstable. Ok. Try26 will up constraints to 30 and the first sheet (s6||s7) to 100, leaving the s5||s6 at 40. We'll see about this. try26 running on peep. I'm also going to run a try using the limits that 'broke' for Kevin, again focusing on the first sheet. try27 running on ribbit. Thu Aug 26 22:48:56 2004 George Shackelford What?? With those high settings in try26 NOTHING HAPPENED! Is it because there were too many 'good' models in the include section of .under?? Why? Lower values than this blew up for Kevin. Is it me? Is it my deoderant? Is it something I said? Fri Aug 27 10:06:30 PDT 2004 Kevin Karplus All of try20,21,23,24,25,26,27 are essentially robetta-model7. It takes some pretty powerful constraints to overcome the local minimum of a good packing. Perhaps we should try manipulating the model by hand and then reoptimizing. I'll try putting in George's sheet constraints (fixing the inconsistent Hbond) in try28, with very heavy constraints. I'll also remove the highly polished models (everything from try20 on, except try22), so that it is mainly going back to robetta-model7. Interestingly, it is NOT starting from robetta-model7, but from try19-opt2, which scores much better on the try28 costfcn. I'll try try29, with the identical costfcn, but starting from the robetta models. Let's see who does better in the end. Fri Aug 27 13:10:48 PDT 2004 Kevin Karplus try29-opt2, which polished robetta-model7 scores almost as well as try26 and try27, despite much less polishing time. try28-opt2 does not score as well, but is not derived from robetta-model7. Currently I favor submitting try27-opt2 best derived from robetta-model7 try28-opt2 best not derived from robetta??? try13-opt2 best unconstrained not from robetta try1-opt2 full auto There are no template alignments worth showing, so we have room for one more prediction. Which existing or new model should it be? Perhaps rosetta's favorite: try21-opt2.repack-nonPC ? Fri Aug 27 14:18 2004 Bret Barnes So, I'm running four more trys (30-33). The first try attempts to form a sheet at the very end of the protein. Str2 predictions give very small prediction for this, but near predicitons show alternating hydrophobics. The second try (try31) includes the constraints from try30 (above), but also trys to tie that sheet into the adjacent sheet at the end of the protein. Try32 is the same as try30, but it attempts to make the two helicies at the top of the protein (92-104) into a sheet. Now I don't know if this is that great of an idea, the helicies pack nicely together, but might as well try something else. The last try (try33) includes try31 sheet constraints and the sheet constraint in try32. So, bascially try30-33 is just a combination of similar things, I don't know which is best, so might as well try them all. try30 on meow PID 8273 try31 on baa PID 8069 try32 on ribbit PID 6560 try33 on croak PID 19986 Fri Aug 27 15:24:49 PDT 2004 Kevin Karplus The starting models for try30, try31, ... seem a bit strange. Why start with ReadConformPDB T0273.try21-opt2.repack-nonPC.pdb ReadConformPDB T0273.try31-opt2.pdb.gz ReadConformPDB T0273.try25-opt2.pdb.gz ReadConformPDB T0273.try24-opt2.repack-nonPC.pdb ReadConformPDB T0273.try21-opt1.pdb.gz ReadConformPDB T0273.try21-opt2.pdb.gz for try31? (try31-opt2.pdb doesn't even exist yet, so that is ignored). Bret, are these the models you wanted to started from, or did you copy another .under file and not check the starting setup? Fri Aug 27 17:03:37 PDT 2004 Martina Koeva I rescored all models with the unconstrained function, as well checked whether there were any changes in the models Rosetta likes best. Try32 and try33 (that are finished) are way down the list. There are no changes at the top. It doesn't seem that try31 and 30 will rise up that much either. Rosetta still likes try21-opt2.repack-nonPC better than any of the others. I can't seem to find any models that look quite different from the ones proposed for submission that score well. The ones that look fairly different are some of the early "barrel" models and they are not represented in the proposed models for submission (and most likely wrong). Maybe try3-opt2, or try5-opt2? Fri Aug 27 17:39:02 PDT 2004 Martina Koeva I superimposed (script name: superimpose-Bret-models.under) ReadConformPDB T0273.try27-opt2.pdb ReadConformPDB T0273.try28-opt2.pdb ReadConformPDB T0273.try21-opt2.repack-nonPC.pdb ReadConformPDB T0273.try32-opt2.pdb ReadConformPDB T0273.try33-opt2.pdb ReadConformPDB T0273.try31-opt1.pdb ReadConformPDB T0273.try30-opt1.pdb and the result was stored in Bret-models.pdb. I included the latest models, generated from the runs by Bret, as well as the earlier best models so far. There are some differences at the C-terminus as Kevin pointed out. Try32-opt2 looks like it's attempting to form a sheet at V170-L172 with F179-L181. However, it's also trying to unfold the helices at the top near residues 100 and has introduced a few breaks without straightening the segments out. Try33-opt2 has curled up part of the strand L162-E168 back into a helix. It has also introduced a break near 179-182, as well as back near residues 104-107. Try31-opt1 has tried to put in the last attempted "sheet" with the 3-strand sheet formed before. It forms strange hydrogen bonds between the strand 162-165 and the segment next to it (167-170), where the segment 167-170 looks like it's trying to be on top of the other strand (?). Try30-opt1 looks like it has shaped the C-terminus into a hairpin-like conformation, but the antiparallel segments don't form a sheet. If we were to submit one of these as a 5th model, I would suggest either try32 or try30. I like try32 better at the C-terminus, because it does look like it's trying to form a sheet there, but it still has the two small breaks near residues 100-110. Try30 does not introduce any such breaks, but it also does not form the expected strands at the C-terminus (which was the purpose of the run). To summarize: 1. If looking for a completely different model: try3-opt2 or try5-opt2 (try3 scores best in its own equivalence class - barrels) 2. Lowest Rosetta energy: try21-opt2.repack-nonPC 3. If looking for changes near the C-terminus: try32-opt2 or try30-opt2 Fri Aug 27 18:21:45 PDT 2004 Kevin Karplus The opt2 models are done, so I replaced them in superimpose-Bret-models.under. I don't see anything to get excited about, and I don't like try3-opt2. try5-opt2, though probably wrong, is an interesting barrel, so I think I'll toss it in. I'll submit T0273.try27-opt2.pdb best (from robetta-model7) T0273.try28-opt2.pdb best not from robetta T0273.try13-opt2.pdb T0273.try5-opt2.pdb T0273.try1-opt2.pdb Thu Nov 18 23:54:26 PST 2004 Martina Koeva Based on the smooth gdt scores: best sam-t04 21.2826 (try31-opt1-scrwl/try31-opt1) best submit 20.0439 (model1) model1 20.0439 auto 11.4178 align 12.7887 robetta best 22.5247 (robetta model8) robetta1 19.6287