Thu Jun 22 10:29:40 PDT 2006 T0341 Make started Thu Jun 22 10:30:24 PDT 2006 Running on lopez.cse.ucsc.edu Thu Jun 22 11:53:34 PDT 2006 Kevin Karplus BLAST gets excellent hits on 1zjjA, 1vjrA, 1pw5A, 1yv9A, 2c4nA, 1wviA, 1ys9A, 1ydfA. (Best evalue is 1.5e-13, 27%id for 262 residues 1zjjA). HMMs get good hits in 1vjrA, 1zjjA, and generally on c.108.1.* Thu Jun 22 16:58:46 PDT 2006 Kevin Karplus Best hit by HMMs is 1zjjA at 6e-26. There are several others before 1vjrA at 2e-18. All 64 hits with E-value < 0.01 are either c.108.1.* or not in SCOP. Fri Jun 23 05:04:48 PDT 2006 Kevin Karplus try1-opt2 is in close agreement with alignments 1,2, and 5 from 1zjjA, 2c4nA, and 1wviA. This probably just needs some polishing to close the two bad gaps T0341.try1-opt2.pdb.gz breaks before (T0341)K173 with cost 4.32184 T0341.try1-opt2.pdb.gz breaks before (T0341)T107 with cost 2.82824 and to settle the sidechains in tighter. try2 started on camano. Tue Jul 11 23:20:42 PDT 2006 Zack Sanborn I was curious why there were no try2-opt* models in decoys. I checked the try2.log and found that it couldn't read the "read-pdb.under" file. Well, the reason for that was that it didn't exist. So, I made the read-pdb.under file (using the command 'make decoys/read-pdb.under') and restarted try2. try2 was started on camano. When try2 is done, I will see how well it does on breaks and go from there. Wed Jul 12 10:54:59 PDT 2006 Zack Sanborn try2-opt2 is the highest scoring structure with its cost function. The big reason for this is that the breaks got much better with try2-opt2: Conformation[9] T0341.try2-opt2.pdb.gz has 19 breaks T0341.try2-opt2.pdb.gz breaks before (T0341)T107 with cost 1.80099 T0341.try2-opt2.pdb.gz breaks before (T0341)A174 with cost 1.28534 T0341.try2-opt2.pdb.gz breaks before (T0341)Q106 with cost 0.924823 T0341.try2-opt2.pdb.gz breaks before (T0341)P119 with cost 0.537565 T0341.try2-opt2.pdb.gz breaks before (T0341)P110 with cost 0.516856 The remaining 14 breaks are below 0.5. The usual thing to do in this case would be to optimize a GROMACS optimized version of try2-opt2, since GROMACS does pretty well with breaks and gaps. However, this time, it has worse breaks than try2-opt2: Conformation[7] T0341.try2-opt2.gromacs0.pdb.gz has 16 breaks T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)T107 with cost 1.90308 T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)A174 with cost 1.42392 T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)Q108 with cost 1.28272 T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)Q106 with cost 1.08682 T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)D109 with cost 0.943809 T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)E64 with cost 0.606576 T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)P119 with cost 0.60321 T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)R5 with cost 0.58749 T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)Q87 with cost 0.529283 T0341.try2-opt2.gromacs0.pdb.gz breaks before (T0341)T225 with cost 0.527175 The remaining 6 breaks are below 0.5. I think I will try to optimize this GROMACS optimization anyway. At the same time, I'll do a polishing run for T0341.try2-opt2. Hopefully I'll be able to get rid of the worse breaks this way. try3 is a polishing run for try2-opt2 with breaks turned up in the costfcn and Gap conformation commands doubled, started on lopez. try4 is a similar run as above but is starting with T0341.try2-opt2.gromacs0.pdb. It was started on lopez. Wed Jul 12 11:49:48 PDT 2006 Zack Sanborn I was worried that the beta-sheet in this model pairing L92-D95 and V113-G116 was wrong since it was predicted to be helix. Looking at the t06 str2 and ehl2 sequence logos, I see that these helix predictions are VERY weak. I like the look of this beta sheet, so I believe it and it's a relief to see that the HMMs don't exactly disagree. When try3 and try4 finish, I'll try to work on the packing of these models since they appear a little foamy at the moment. Also, there's a potentially bad mismatch between the secondary structure prediction and the model at residues T149-L155 where a beta-hairpin is in the model but a helix was moderately predicted. I'd like to see if I could get a model that gets a helix there instead of the beta-hairpin to see how it scores relative to the other. To do this, I'll restart from the alignments and up the helix constraint for this region. ALSO, if I could get the helix at G160-T168 to do a half turn, the hydrophobic residues would be buried against the beta sheet. More importantly, there's a phenylalanine at position 131 in the helix neighboring this helix that is currently exposed. So, what I'd like to do is twist both of the helices toward each other, burying both of their hydrophobic sides onto the beta sheet. The first thing I'll do is try to turn that beta-hairpin into a helix. try5 has taken the sheet and helix constraints from try2-opt2 with the following change: The sheet constraint: SheetConstraint (T0341)Y148 (T0341)R151 (T0341)L157 (T0341)G154 hbond (T0341)Y149 1 Has been turned into a helix constraint: HelixConstraint (T0341)Y148 (T0341)L157 1 try5 then restarts from the top-scoring alignments, started on orcas. Wed Jul 12 13:46:16 PDT 2006 Zack Sanborn try3 and try4 have finished. Both have managed to reduce the breaks by a bit. The most severe breaks are now: For try3-opt2: T0341.try3-opt2.pdb.gz breaks before (T0341)A174 with cost 1.27944 T0341.try3-opt2.pdb.gz breaks before (T0341)P110 with cost 0.520755 For try4-opt2: T0341.try4-opt2.pdb.gz breaks before (T0341)A174 with cost 1.03747 T0341.try4-opt2.pdb.gz breaks before (T0341)T107 with cost 0.546696 try3 has *slightly* worse breaks, but it scores better than try4 with their cost functions. With an unconstrained cost function, the story is the same. Actually, try1-opt2 even scores better than try4-opt2 with the unconstrained cost function. try3-opt2 < try2-opt2 < try1-opt2 try4-opt2 < try2-opt2.gromacs0 < try2-opt2 < try1-opt2 I made a new best-models.pdb.gz, taking try's 1-4. The models have *VERY* few differences, they just do better on the breaks. try5 will likely be more different with the helix constraint we put in... that is, unless the constraints aren't strong enough. Wed Jul 12 14:21:05 PDT 2006 Zack Sanborn Now, for try6, we want to try to twist those helices to get them to pack their hydrophobic residues better. One we're especially interested in packing is F131. Looking at rr.constraints, we see the following residue constraints involved F131: Constraint F131.CB L140.CB -10. 7.0 14.0 0.507604303797 Constraint F131.CB I141.CB -10. 7.0 14.0 0.507016698475 Constraint F131.CB A142.CB -10. 7.0 14.0 0.491426543268 bonus Constraint F131.CB A165.CB -10. 7.0 14.0 0.462460269377 bonus With the exception of the last constraint (which based on the current model, looks hard to satisfy... but it is, after all, a bonus), these constraints look like they could help. Looking at the model, there's one more constraint we could try to get that helix to twist: Constraint F131.CB V114.CB -10. 7.0 14.0 To get the other helix to twist in, we should place another set of residue constraints for the hydrophobic residue L166. Looking at rr.constraints for L166, we see the following constraints. Constraint L140.CB L166.CB -10. 7.0 14.0 0.500183253814 Constraint I143.CB L166.CB -10. 7.0 14.0 0.475811163974 bonus Constraint L127.CB L166.CB -10. 7.0 14.0 0.385029759587 bonus The first two constraints would help twist the helix in the correct orientation. The last constraint, however, would not help do what we want, so we'll remove it. Wed Jul 12 16:16:06 PDT 2006 Zack Sanborn Okay, try5 has finished and it DIDN'T make the helix I wanted. So, I've started a try6 that increases the weights of the constraints overall and also specifically increases the helix constraint I wanted to make this whole time. Also, started a try7 that attempts what I discussed above by twisting the helices of the existing models to get better packing of the hydrophobics. It used the following residue constraints in addition to the helix and sheet constraints used before: Constraint F131.CB L140.CB -10. 7.0 14.0 5 Constraint F131.CB I141.CB -10. 7.0 14.0 5 Constraint F131.CB A142.CB -10. 7.0 14.0 5 bonus Constraint F131.CB V114.CB -10. 7.0 14.0 5 Constraint L140.CB L166.CB -10. 7.0 14.0 5 Constraint I143.CB L166.CB -10. 7.0 14.0 5 bonus We'll see if it does anything I want. I also increased the weight of the constraints to 50... really trying to get it to work. Wed Jul 12 16:22:55 PDT 2006 Zack Sanborn I've now had a chance to look at try5-opt2 a little more closely. It doesn't score well at all with its own cost function, but with the unconstrained cost function it scores second best... just below try3-opt2. Also, looking at best-models.pdb.gz that includes try5-opt2, it is definitely is more different than the other models. So, it will definitely be worth polishing up try5-opt2. Unfortunately, try5-opt2.pdb has one fairly severe break. T0341.try5-opt2.pdb.gz breaks before (T0341)V105 with cost 5.18116 T0341.try5-opt2.pdb.gz breaks before (T0341)E64 with cost 0.636898 However, try5-opt2.gromacs0.pdb has better breaks (like it usually does): T0341.try5-opt2.gromacs0.pdb.gz breaks before (T0341)V105 with cost 1.95541 T0341.try5-opt2.gromacs0.pdb.gz breaks before (T0341)G104 with cost 0.548658 T0341.try5-opt2.gromacs0.pdb.gz breaks before (T0341)Q87 with cost 0.537003 T0341.try5-opt2.gromacs0.pdb.gz breaks before (T0341)K182 with cost 0.531061 T0341.try5-opt2.gromacs0.pdb.gz breaks before (T0341)A9 with cost 0.507058 So, to improve the try5-opt2 model, I will attempt to polish the GROMACS model with try8. try8 started on lopez. Wed Jul 12 20:01:13 PDT 2006 Zack Sanborn Well, try6, try7, and try8 have completed. try8-opt2 is the second-best scoring model, right below try3-opt2 and above try5-opt2. I've added these models to superimpose-best.under and have looked at the best-models.pdb.gz. The variation in models is less than what I expected, but there are at least two different types of models. The biggest difference lies in the beta-hairpin vs. helix around residues Y149 - A156. Models try6-opt2 and try7-opt2 have helices (matching the secondary structure prediction), while models try3-opt2, try8-opt2 and try5-opt2 have beta-hairpins. Interestingly, the models with the beta hairpin score better. I believe this is due to a greater number of breaks in the other models, although I haven't had time to check that yet. Also, I believe try3-opt2 has done the best job packing the hydrophobics I was trying to pack with try7. I think I have some more work to do with the constraints in the try7 run. So, I believe the 5 models to submit for the soft deadline: try3-opt2 try8-opt2 try7-opt2 try5-opt2 and maybe a model based off the highest scoring alignment (1wviA) or try6-opt2 There's definitely more work needed to get some more variety in these models. Wed Jul 12 21:08:28 PDT 2006 Kevin Karplus I submitted a slight modification of Zack's list, with comment: For the preliminary submission, we have two groups of models: those wih a helix around Y149-A156 (predicted by our neural nets) and those with a hairpin there. Model 1 is try3-opt2, the best-scoring model with the latest cost function, and with the best packing of hydrophobics. Model 2 is try8-opt2.gromacs0.repack-nonPC, rosetta's favorite model of the ones it repacked sidechains for. Model 3 is try7-opt2, our best-scoring model with a helix. Model 4 is try5-opt2, yet another hairpin model. Model 5 is try6-opt2, yet another helix model. Fri Jul 14 12:05:38 PDT 2006 Zack Sanborn I'm going to make a chimera combining the majority of our top-scoring model (try3-opt2) that has the best hydrophobic packing, but replacing the beta-hairpin with the helix predicted by the neural networks. So, using the best-models.pdb.gz superposed model, I will copy-and-paste the helix from try7-opt2 (residues A146 - G160) into the try3-opt2 model. I will then optimize the model with Undertaker and hope that this will produce a model that combines the best-packing with the best agreement to the secondary structure predictions. Fri Jul 14 13:13:46 PDT 2006 Zack Sanborn I've made the chimera as described above, it's named T0341.chimera-try3-opt2.pdb.gz and can be found in the decoys directory. I've started a try9 that is optimizing this chimera. It was started on camano at 13:10. The penalties for breaks and soft_clashes were turned up. Also, the constraints from try7 remain to try to keep that helix as a helix. However, the cost for constraints were reduced back to their normal level (10). I've also started two more polishing runs for try3-opt2 and try7-opt2 to try to get them both to pack a little better as their current structures are a little foamy. For these runs, I've increased dry5 and phobic_fit weights. I've started try10, polishing try3-opt2 on orcas at 13:28. I've started try11, polishing try7-opt2 on vashon at 13:29. I've started try12, polishing try9-opt2, the chimera based optimized model. Sat Jul 15 02:22:01 PDT 2006 Zack Sanborn All above runs have completed, and I've rescored all models using the unconstrained cost function. The top models are now (ignoring *-opt1's and rosetta repacks): try10-opt2 (polishing of try3-opt2) try12-opt2 (polishing of try9-opt2, chimera of try3-opt2 and helix from try7-opt2) try11-opt2 (polishing of try7-opt2) try3-opt2 (original model with best packing and beta hairpin) try9-opt2 (first optimization of chimera) try8-opt2 (decent model with beta hairpin) try5-opt2 (decent model with beta hairpin) I've put the following models into superimpose-best.under T0341.try10-opt2.pdb T0341.try12-opt2.pdb T0341.try11-opt2.pdb T0341.try3-opt2.pdb T0341.try8-opt2.gromacs0.repack-nonPC.pdb.gz That last one is still the best scoring model in the eyes of rosetta (from grep-best-rosetta). I think the above are the five models to submit, with possibly replacing try3-opt2 (which is similar to try10-opt2) with try6-opt2 a poorly scoring model, but did contain the helix from the secondary structure predictions. Sat Jul 15 09:22:24 PDT 2006 Kevin Karplus Zack's suggestions look good. I've done the final submission using them.