Tue May 30 09:09:49 PDT 2006 T0303 Make started Tue May 30 10:42:14 PDT 2006 Running on lopez.cse.ucsc.edu Tue May 30 11:15:58 PDT 2006 Kevin Karplus Good BLAST hit to 2ah5A (E-value 2.3e-12) 39 PDB sequences in t06 multiple alignment. Tue May 30 12:41:55 PDT 2006 Kevin Karplus lopez crashed (or something) at 11:12 this morning. I'll try directing this job to the farm cluster, though the machines there are slow, they aren't subject to the power problems in PSB today. Make started Tue May 30 12:44:49 PDT 2006 Running on farm07.cse.ucsc.edu The template-lib scoring is very slow on the farm cluster. Top hits with t06-template are 1zjjA 1x42A 1swvA 1qyiA 1l7mA ... Thu Jun 15 18:36:01 PDT 2006 Kevin Karplus All the models are quite similar. I'm starting try2 as a polishing run for our try* models, and try3 as a polishing run for the models from top servers: Pmodeller6, ROBETTA, and 3Dpro. Both started on orcas. Thu Jun 15 21:13:21 PDT 2006 Kevin Karplus try3, based on the server models, does better than try2, based on our models :-( Even rosetta prefers repacking the try3 models, even though the were based on the Pmodeller6_TS1 model, not the robetta models. There may have been some crossover, but it looks like most of the gains came from gap-closing operators. Both clashes and gaps were substantially reduced. Thu Jun 15 21:23:44 PDT 2006 Kevin Karplus I see what the problem is in try2 and try1: The inserted domain has been opened up in the optimization process somewhere, and never put back together. try3-opt2, which looks a lot like the alignments, is the only decent complete model. Do we need to pick up helix-packing constraints from alignments as well as SheetConstraints? Fri Jun 16 08:49:41 PDT 2006 Kevin Karplus Two directions to go in: 1. do a subdomain for N17-R95, perhaps adding constraint N17.CA R95.CA 5.0 5.7 6.4 constraint N17.CB R95.CB 4.5 5.1 5.7 Hbond N17.N R95.O Hbond N17.O R95.N to hold the ends together (if necessary) [Fri Jun 16 08:56:59 PDT 2006 Kevin Karplus make started on lopez, but without attempting to add constraints ] 2. extract some constraints from the top alignment by hand and rerun from the alignments: Constraint A38.CA E69.CA 7 8.3 10 Constraint A29.CA F82.CA 4 5.5 7 Constraint V43.CA S55.CA 9 10.3 11.3 [Fri Jun 16 09:17:30 PDT 2006 Kevin Karplus started as try4 on shaw ] Fri Jun 16 15:03:58 PDT 2006 Kevin Karplus The N17-R95 subdomain predictions looks pretty good, but the constraints will be needed on the end to make it superimposable. Started N17-R95/try2 on lopez with the added constraints for the ends, still just optimizing from the alignments. We may want to do one more round of optimization (with just the end constraints) of N17-R95 before making a chimera with the rest of the protein. Fri Jun 16 15:24:22 PDT 2006 Kevin Karplus The try4-opt2 optimization looks very much like the try1-opt2 model---it has not fixed the problems in the inserted domain. Fri Jun 16 18:10:41 PDT 2006 Kevin Karplus The N17-R95 try2 almost forms the necessary H-bonds at the connection point between the two domains. I'll optimize from the existing models with those constraints as the *only* constraints, and see if we can fix the remaining minor problems. After that, I'll have to do a superposition with the whole chain, create a chimera, and optimize the chimera. (N17-R95 try3 running on cheep) Sat Jun 17 03:05:19 PDT 2006 Kevin Karplus I made chimera1 from try3-opt2 and N17-L90 of N17-R95/ I made chimera2 from try4-opt2 and N17-L90 of N17-R95/ Both have some breaks that need patching, but score moderately well with try4.costfcn and unconstrained.costfcn try5 will try to polish chimera1 try6 will try to polish chimera2 [Sat Jun 17 03:15:54 PDT 2006 Kevin Karplus started on orcas] I expect to submit try5, try6, try3, try4, and one other. Sat Jun 17 07:51:34 PDT 2006 Kevin Karplus Nope, both try5 and try6 mangled things in order to close the small gaps. I'll have to try again, perhaps using the try4 costfcn. Sat Jun 17 08:00:27 PDT 2006 Kevin Karplus try7 and try8 started on orcas. Sat Jun 17 08:56:24 PDT 2006 Kevin Karplus I think try7 and try8 will have problems also, as they keep inserting alignments, which throws away much of the chimera. I'll try again with InsertAlignment turned off. Sat Jun 17 09:08:39 PDT 2006 Kevin Karplus To keep track of the history: server models => try3 => chimera1 => try5,try7 alignments => try4 => chimera2 => try6,try8 Nope: the problem with try7 and try8 was a typo in the ReadConformPDB command, so that the runs were starting from a random model, then inserting alignments. retarted try7 and try8 on orcas. Sat Jun 17 12:43:34 PDT 2006 Kevin Karplus OK, try7 and try8 did OK, though I like try7 better. Both need some more polishing to close gaps. Perhaps they are ready for the try5=try6 costfcns now. Starting optimizations with try7 starting points for try9 and try8 starting points for try10. To keep track of the history: server models => try3 => chimera1 => try5,try7=>try9 alignments => try4 => chimera2 => try6,try8=>try10 Sat Jun 17 16:22:07 PDT 2006 Kevin Karplus try9-opt2 is now the best-scoring, followed by try3-opt2 and try10-opt2. Sat Jun 17 16:40:40 PDT 2006 Kevin Karplus I submitted ReadConformPDB T0303.try9-opt2.pdb # opt from chimera1 with constraints ReadConformPDB T0303.try10-opt2.pdb # opt from chimera2 with constraints ReadConformPDB T0303.try3-opt2.pdb # opt from servers ReadConformPDB T0303.try4-opt2.pdb # opt with constraints on helices InFilePrefix ReadConformPDB T0303.undertaker-align.pdb model 1 # alignment to 2ah5A with the following comments added to the method: Target T0303 was a fairly straightforward comparative modeling target, with even BLAST able to find 2ah5A as a good template. Despite this, our automatic prediction picked a different template for the inserted domain N17-R95. It also rather mangled the N17-R95 domain in attempting to close gaps. We noticed that other servers followed 2ah5A very closely, even though the alignment was not as good in this domain as in the outer domain. We optimized the models obtained from the servers, to see whether they were doing better at finding a good alignment to a template. The final optimization try3-opt2 was derived primarily from Pmodeller6_TS1, though ROBETTA_TS3 and ROBETTA_TS4 also did well in the early stages of the optimization. We also repotimized from the alignments, with some extra constraints (taken from the alignment to 2ah5A) to keep the helices of the N17-R95 domain packed. (This was try4-opt2.) We decided to optimize N17-R95 separately, then paste it back into the whole-chain models (both from the servers and from our alignments). Note: G197-P209 appears to be incorrectly placed in try4-opt2 and try10-opt2 derived from it. The models from the servers seem to have done a better job here. To keep track of the history: server models => try3 => chimera1 => try5,try7=>try9 alignments => try4 => chimera2 => try6,try8=>try10 Model 1 is try9-opt2, which is an optimizaton of the outer domain from the servers with the N17-R95 subdomain pasted in. Model 2 is try10-opt2, which is based entirely on models we generated, with the N17-R95 subdomain pasted into try4-opt2 and reoptimized. Model 3 is try3-opt2, reoptimized from the server models. Model 4 is try4-opt2, created from the alignments with some extra constraints to keep the N17-R95 domain from unfolding. Model 5 is just SCWRL-based sidechain replacement on an aligment to 2ah5A. Fri Jul 7 16:45:40 PDT 2006 Martin Madera Let me try and make sense of the models so far. try1: automatic, -B- try2: polishing try1 try3: polishing server models, -A- try4: -A- contraints on alignments, still made -B- try5: optimising chimera1, but still made -B- try6: optimising chimera2, blew up try7: optimising chimera1, different cost fcn, -A- try8: optimising chimera2, different cost fcn, -A- try9: polishing try7 try10: polishing try8 I don't like the chimera-based structures very much, the angle between the insertion and the main domain looks strange compared to other 2ah5A targets (T0324, T0329, T0330). The top five BLAST hits are: 2ah5A 2.3e-12 -A- 1x42A 1.1e-09 -B- 1te2A 7.8e-08 modified -A- 2fdrA 1.2e-05 -A- 1o08A 3.1e-04 -A- Let me try and build a model based on just the -A- structures, 2ah5A|1te2A|2fdrA|1o08A and see how well it scores. This will be try11. Running on squawk. Fri Jul 7 19:12:19 PDT 2006 Martin Madera Oops, try11 failed -- looks like make extra_alignments make read_alignments never got done. Did the makes, restarted try11. Restarted. Fri Jul 7 19:28:08 PDT 2006 Martin Madera Hmmm, *still* no 2ah5A/read-alignments-scwrl.under. Creating them 'manually' via $ for f in 2ah5A 1te2A 2fdrA 1o08A; do make $f/read-alignments-scwrl.under; done Restarted again. [Sat Jul 8 13:47:56 PDT 2006 Kevin Karplus the problem was that MANUAL_TOP_HITS has to come *before* the include in the Makefile. I fixed that and will run make extra_alignments and make_read_alignments ] Fri Jul 7 22:51:17 PDT 2006 Martin Madera Try11 finished, and it's a -B- structure. I noticed I forgot to remove all-align.a2m from the .under file! So try12 is attempt number two, this time with all-align.a2m removed. Running on squawk. Sat Jul 8 13:09:46 PDT 2006 Martin Madera Try12 looks OK, scores better than Kevin's chimeras but nothing spectacular. The main problem is that the 72-89 helix isn't packed properly. As far as I can see, the main problem are clashes caused by F78, but there's a cavity where it could go. So, I will try to pull the following residues on the three helices together: 29, 55 and 78. # this is the current distance; don't move! Constraint A29.CA S55.CA 8.5 8.89 9.4 1 # the current distance is 7.7 -- pull it closer in! Constraint A29.CA F78.CA 4.0 6.0 7.7 1 # the current distance is 8.2 -- pull it closer in! Constraint S55.CA F78.CA 4.5 6.5 8.2 1 Otherwise same cost function as try12 (sans constraints), re-optimising try12-gromacs. This is try13. Running on lopez. Sat Jul 8 17:01:19 PDT 2006 Martin Madera I like try13, but it scores worse than just about any other model (including try12) on the unconstrained function. The distances are: # this is the current distance; don't move! Constraint A29.CA S55.CA 8.5 8.89 9.4 1 ... 8.9 # the current distance is 7.7 -- pull it closer in! Constraint A29.CA F78.CA 4.0 6.0 7.7 1 ... 6.0 # the current distance is 8.2 -- pull it closer in! Constraint S55.CA F78.CA 4.5 6.5 8.2 1 ... 6.5 Nice! Exactly what I wanted. Except the numbers just wild guesses, and hence probably wrong. I've been looking at the model in spacefill & slab mode and there's quite a cavity in the centre. Given that undertaker seems pretty good at moving things where I want them, it shouldn't be too difficult to pack it. Except, how to pick the constraints? The colouring I will be using is: background black select 17-34 color red select 17-34 & backbone color pink select 35-49 color green select 35-49 & backbone color lightgreen select 50-66 color blue select 50-66 & backbone color lightblue select 67-95 color grey select 67-95 & backbone color lightgrey select * restrict 17-95 center 17-95 ... and then spacefill & slab mode. The order is R-G-B-grey. As things stand, there are some pretty bad clashes between red and green. To see the packing between the two, do: select backbone spacefill off backbone 100 restrict 17-49 center 17-49 Hmm... actually let's turn off the residues that are on the outside: restrict 17-49 center 17-49 spacefill on select 17,20,21,24,25,28,29,31,32,39,40,41,42,45,49 spacefill off backbone 100 Sun Jul 9 09:10:15 PDT 2006 Kevin Karplus I need a list of the top 5 models, and superimpose-best.under still has try9, try10, try3, try4, align1, which was our preliminary submission. Is this still the best set? I don't seen any comments above that would lead me to believe that Martin has found anything he likes better than what we have already submitted, so I won't do a new submission today *unless* Martin updates superimpose-best.under and sends me e-mail. Sun Jul 9 16:27:58 PDT 2006 Martin Madera Nope, so far no improvement on what was submitted for the soft deadline. I've gone back to looking at try12 rather than try13. I find the color scheme very useful, it gives you a very good sense of what's going on. I think that the largest cavity is actually between the bottom of the red helix and the grey one. And it could be closed by pulling Val33.CA towards Glu74.CA: select Val33.CA, Glu74.CA color yellow The current distance is 9.7A, I reckon 5A is what it should be. So Constraint V33.CA E74.CA 3.0 5.0 7.0 2 Running as try14 on lopez. Sun Jul 9 20:11:00 PDT 2006 Martin Madera In try14 the distance is 5.0A (spot on), but it seriously distorted the protein in doing so. Plus I think that 5.0A is way too close, the van der Waals spheres are almost touching! I'll try 7.0 next. Running as try15 on peep. I will also start two runs polishing the gromacs versions of our top models so far: try16: polishing try7 & try9-opt2.gromacs0 ... running on shaw try17: polishing try8 & try10-opt2.gromacs0 ... running on lopez Sun Jul 9 22:06:50 PDT 2006 Martin Madera Try15 finished, but it scores worse than try14 (and much worse than try7 and try8). Hmmmm, I'll wait for try16 and try17 but otherwise I'm giving up on this target. Sun Jul 9 22:35:28 PDT 2006 Martin Madera Try16 and try17 have finished. According to the unconstrained cost function, try16 is worse than try9, and try17 is worse than try10. Ho-hum. Rosetta seems to like my runs over Kevin's. But this is probably an artefact, because most of my runs were started from the gromacs versions (and Rosetta likes those). I need to do more work on T0307, so target over. Wed Mar 21 20:32:27 PDT 2007 Kevin Karplus This has two domains. On domain 1, our best model was try3-opt1, which 9 server models beat. The best we submitted was model3=try3-opt2, which was in the top 5%. Our model1=try9-opt2 was worse, but still in the top 5th. On domain 2, our best model was try13-opt2.repack-nonPC, for which only 3 server models were better. But we didn't submit this. The best we submitted was model3=try3-opt2 which was still in the top 6th. Our model1=try9-opt2 was in the middle of the pack. For the whole protein, try3-opt1 did best, beaten by 8 server models, with model3 close behind. Hmm, try3 was just a polishing of server models, so it is embarassing that it did the best of our runs.