Mon Jul 17 09:35:48 PDT 2006 T0382 Make started Mon Jul 17 09:37:23 PDT 2006 Running on shaw.cse.ucsc.edu Mon Jul 31 12:02:55 PDT 2006 George Shackelford T0382 - looks like a new fold to me. It looks like all helical as well. I wonder if that is a break at G31? I don't have any indication from t2k.ehl2 and t06.ehl2. None of the notors or seps have any indication. They all seem to stress a "weakness" in predictions at that residue. I need to load the servers and do a score-all+servers.unconstrained. We come out on top above Robetta, because of doing better on alpha predictions and soft clashes. Robetta does much better on phobic_fit. I'm a bit wary of our lead. Next I need to check the close templates. Am I really getting someplace?? Both t2k and t06 seem to agree on predictions. Ok, we'll try alphabetmatch. # length: 123 # length range: 114 to 135 # alphabets used: # ehl2 burial # id score per residue 5S 10N 10N 1hlvA 322.881 2.62505 ,1.10.10.60-58,1.10.10.60-60 6paxA 321.118 2.61071 ,1.10.10.10-68,1.10.10.10-49 1gqaA 316.281 2.57139 1.20.120.10-130 1orsC 315.78 2.56731 1.20.120.350-132 1t6uA 314.753 2.55897 1.20.120.400-117 1ngkA 313.805 2.55126 1.10.490.10-126 1gcwB 313.458 2.54844 1.10.490.10-129 1k40A 313.32 2.54732 1q0gA 312.962 2.5444 1.20.120.400-117 2mhr 312.896 2.54387 1.20.120.50-118 Eexamining the first few hits and looking at the consensus ehl2, I find that alphabetmatch thinks the first helix is actually two helices. That is throwing off the match. I could go in and modify the results to make it look more like one helix; that should help. On the other hand, there is possibly no easy match out there. I can look at the top SAM hits and see if there is something that looks good. One long helix and four small ones; the long one seems bent. Interesting. So what were the best scorings in the try1? 2a9uA, 1gmpA, 1trrA, 1cwyA, 1mgrA. The last seems to be the template for try1. Not 1gmpA. 2a9uA looks nice but it has a long helix at the end. 1trrA and 1cwyA don't work. 1mgrA scarcely has a helix!! Damm I don't think I'm finding a good hit anywhere. I think I need to re-enforce the long helix and see what I can get from ANY of the current templates! Mon Jul 31 16:52:31 PDT 2006 Sylvia Do Okay, I put the best-scoring server models in the superimpose-best file. InFilePrefix decoys/servers/ ReadConformPDB SAM_T06_server_TS1.pdb ReadConformPDB ROBETTA_TS2.pdb I'm just going to see how the other models line up with these three. Wow. Looks like a mess. There is only one helix position that all three seem to agree on. Tue Aug 1 12:36:11 PDT 2006 George Shackelford Of course it looks like a mess, Sylvia. It's a new fold! When we're finished with this, they'll still look different from each other. That helps our chances of getting the right structure. If we had agreement, it wouldn't be a new fold. After reviewing the predictions in our target meeting, I need to ensure a break around residue 31. This should make the formation of alternative structures easier. After I adjusted the constraints for t06.dssp-ehl2.rdb, I got the following alphabetmatch predictions: 1hlvA 327.009 2.65861 ,1.10.10.60-58,1.10.10.60-60 6paxA 324.265 2.6363 ,1.10.10.10-68,1.10.10.10-49 1oqcA 318.215 2.58711 1.20.120.310-112 1ngkA 318.045 2.58573 1.10.490.10-126 1orsC 317.475 2.5811 1.20.120.350-132 1gqaA 316.802 2.57563 1.20.120.10-130 2ccyA 316.189 2.57064 1.20.120.10-127 1a7vA 314.733 2.5588 1.20.120.10-125 1k40A 314.503 2.55694 1t6uA 313.259 2.54682 1.20.120.400-117 The biggest change is the appearance of 1oqcA out of nowhere. The 1.20.120.10 of 1gqaA drops down I've changed the included constraints in the costfcn file where I can make changes necessary. # include T0382.dssp-ehl2.constraints # we split this next one.. # HelixConstraint T11 K52 6.399 # ... into two helices HelixConstraint T11 L29 .6399 HelixConstraint S35 K52 .6399 I've also turned down the need for beta bonds. I'm going to a fresh run again. try3 running on peep. Tue Aug 1 14:57:03 PDT 2006 try3 looks better than try2 even thought it scores about the same. It is less foamy and more "natural" looking (to my eyes). for the record: try1 <- all-align.a2m:1mgrA try2 <- all-align.a2m:1mgrA try3 <- all-align.a2m:1h5zA (it appears) Ok, now for something really different. I've added my alphabetmatch predictions to MANUAL_TOP_HITS and run extra_alignments, and read_alignments. I've taken 1hlvA and 6paxA and focused on them in try5 try4 running on camano try4 <- 1hlvA-t2k-local-dssp-ehl2-1.0+0.3-adpstyle5.a2m Tue Aug 1 18:54:05 PDT 2006 George Shackelford Well try4 looks nice and protein-like. I don't know why there is a wedge opening in the structure, it looks like it should be able to close that gap. I could put some constraints in to pull it together, but I'll leave that for now. First I'd like to investigate other possibilities. 6paxA gets a decent starting score. I'd like to push it by itself and see if it can converge to a good structure. I could take a few more of the next ones and see if undertaker can find a good structure in them. If I get really crazy, I could look at this as two domains and do a domain search. First we do the easy stuff. try5 will focus on 6paxA. try6 will group the next three for a run. Assuming there are free machines (the 505 residue monster is in towm). try5 running on peep. Wed Aug 2 11:45:18 PDT 2006 Try5 looks bad, not like a protein at all. Big gaps. I think I'll take a look at the 1.20.120 possibilties. I'll use the following: 1oqcA 318.215 2.58711 1.20.120.310-112 1orsC 317.475 2.5811 1.20.120.350-132 1gqaA 316.802 2.57563 1.20.120.10-130 2ccyA 316.189 2.57064 1.20.120.10-127 1a7vA 314.733 2.5588 1.20.120.10-125 I'm keeping the bonuses from rr.constraints in. The helices won't be bent just to get a bonus. try6 running on peep Wed Aug 2 13:42:40 PDT 2006 George Shackelford Try6 actually has an appealing shape to it, the way that it fits the helical parts together. I don't think it is right, though it scores second best according to Rosetta. In our scoring it is near the bottom. We MIGHT consider it as an alternative. Looking at sources: try1 <- all-align.a2m:1mgrA try2 <- all-align.a2m:1mgrA try3 <- all-align.a2m:1h5zA (it appears) try4 <- 1hlvA-t2k-local-dssp-ehl2-1.0+0.3-adpstyle5.a2m try5 <- 6paxA Try6 <- 1orsC global.a2m I'm running a try7 based on checking out the 1.20.120.10 group. I suspect we're not going to get anything interesting out of it. try7 running on peep try7 <- 1gqaA-t2k-global Wed Aug 2 23:41:59 PDT 2006 George Shackelford Ok, I need get a fix on the best. by unconstrained: try2, try3, try4, try1, try7, try6 by try7 cost: try2, try3, try4, try1, try7, try5 by rosetta: try3, try7, try6, try2, try1, try5 try1 and try2 come from 1mgrA. I like try2. try5 is ugly. And I like try3 over try2 so I get: try3, try2, try4, try7. try6 (or try1) superimposed, best is built. I wonder if I could close that gap in try4... I'm going to redo try4 but put in the rr.constraints and see if that closes the gap. try8 running on peep Thu Aug 3 13:47:51 PDT 2006 George Shackelford Ouch. try8 turned out poorly. It has to be the use of rr.constraints! I'm going to redo those runs that were using them and see what results I can get. try9 (a remake of try5) running on peep. try10 (a remake of try6) running on vashon. try11 (a remake of try7) running on vashon. Fri Aug 4 19:03:11 PDT 2006 George Shackelford Well, it's time to wrap this one up. The remakes mentioned above generally didn't do as well as the original. The only one that did better was try9, and it ended up looking worse than try5. I decided to do two more runs using ony 2a9uA as the main template. One run (try12) used a costfcn with the rr.constraints, and the second (try13) did not. Try12 scored very well! I'm going to include it in the group! ================================== This target got weak hits all around. The top three consensus hit were 2a9uA, 1h5zA and 2a5dB. E-values started at .5 and quickly went up. The BLAST results had a best e-value of .021 for 1vlyA and 1nrkA. We concluded this was a weak fold-recognition or new-fold. The EHL2 predictions (similar to standard EHL) indicated an all helical structure. One sticking point was the region around G31. Was this helical or was there a break? We examined our other local structure predictions and decided from the ones predicting hydrogen bond connections that there was a break. This required adjusting the helical constraints that had been automatically generated for try1. Model 1 is try3-opt2 which was based on 1h5zA (as best as we can determine). This is the results of the constraints used for try2 but with greatly reduced weight on generating beta sheet h-bonds. Try3 got the best Rosetta score and the second best in our unconstrained scoring. We chose it over Model 2 on its appearance. Model 2 is try2-opt2. This is based on 1mgrA as was try1, however try2 has a different solution. This did have helical constraints that included a break at G31 which try1 did not have, but we don't believe that made a lot of difference; undertaker simply found a different structure. Model 3 is try12-opt2 which is based on 2a9uA. 2a9uA had the best e-value in our group of best scoring templates. The results is a near bundle of helices. I don't like the way two of the helices fell between the first two but I don't think we can really fix that. We've tried before on a similar target. Model 4 is try4-opt2 which is based on 1hlvA. 1hlvA was the best scoring of pdb templates selected by "alphabetmatch" using the ehl2 and near-backbone-11 local structure alphabet predictions. It has an appealing cross-hatch structure. Model 5 is try7-opt2 which is based on 1gqaA. 1gqaA comes from the alphabetmatch program and is a representative of SCOP domain 1.20.120.10. try7-opt2's best distinction is having the second best score using Rosetta. try6-opt2 which is based on 1orsC. Try6-opt2 has the third best score using Rosetta. try1-opt2 may be better to include than try6 or try7. At least it's different. try9-opt2 scores pretty well (unconstrained) but is so foamy as to be unbelievable. try8-opt2 scores decently (unconstrained) but has a terrible looking hole in the structure. It has to be wrong. try11-opt2 is a remake of the try7 that doesn't look like try7 try10-opt2 scores poorly and doesn't even look that decent. Fri Aug 4 21:51:12 PDT 2006 Grant Thiltgen I did the submission for T0382 Model 1 is try3-opt2 which was based on 1h5zA (as best as we can determine). This is the results of the constraints used for try2 but with greatly reduced weight on generating beta sheet h-bonds. Try3 got the best Rosetta score and the second best in our unconstrained scoring. We chose it over Model 2 on its appearance. Model 2 is try2-opt2. This is based on 1mgrA as was try1, however try2 has a different solution. This did have helical constraints that included a break at G31 which try1 did not have, but we don't believe that made a lot of difference; undertaker simply found a different structure. Model 3 is try12-opt2 which is based on 2a9uA. 2a9uA had the best e-value in our group of best scoring templates. The results is a near bundle of helices. I don't like the way two of the helices fell between the first two but I don't think we can really fix that. We've tried before on a similar target. Model 4 is try4-opt2 which is based on 1hlvA. 1hlvA was the best scoring of pdb templates selected by "alphabetmatch" using the ehl2 and near-backbone-11 local structure alphabet predictions. It has an appealing cross-hatch structure. Model 5 is try7-opt2 which is based on 1gqaA. 1gqaA comes from the alphabetmatch program and is a representative of SCOP domain 1.20.120.10. try7-opt2's best distinction is having the second best score using Rosetta. Submission sent on August 4th, 9:49 pm. Mon Oct 9 12:13:10 PDT 2006 Kevin Karplus Our best model was try3-opt2.repack-nonPC. Our best submitted was model1 (which would have been improved slightly by SCWRL). There were several better server models, the best being Pcons6_TS1=ROBETTA_TS2 (which also would improve slightly by scwrling). We might have done better on the helix packing if we had used the rr constraints more. The alphabetmatch results (models 4 and 5) were much worse than the others. Models 1 and 3 did score much better.