Fri Jul 14 10:50:07 PDT 2006 T0379 Make started Fri Jul 14 10:53:11 PDT 2006 Running on shaw.cse.ucsc.edu Fri Jul 14 15:42:12 PDT 2006 Kevin Karplus BLAST has excellent full-length hit to 2b0cA (25% over 180 residues, e-value 7.5e-10), also good hits for 1o08A, 1c56A, 1zd3A. HMMs seem to agree (not finished yet). Fri Jul 14 18:11:06 PDT 2006 Kevin Karplus The HMMS puts 2fdrA first, with 2b0cA 9th and 1o08A 6th. Try1 seems to be based on 1mh9A (after first favoring 2b0cA). Fri Jul 21 11:39:36 PDT 2006 Crissan Harris I started looking at the try1 model and the top models in the best-models.pdb. It seems that the models agree pretty well except for residues 15-85. This region seems mostly helical but with some disagreement in the placing of the helices. The region 15-85 in try1-opt2 seems ugly, I think that the helices should be a little longer and thus a little tighter to eachother. I am going to do a new run adding in my own helix constraints that I picked up from the best-models. The helix constraints were added as follows: //helix constraints determined from best-models HelixConstraint (T0379)N17 (T0379)A31 3 HelixConstraint (T0379)M36 (T0379)E57 3 HelixConstraint (T0379)R60 (T0379)L71 3 HelixConstraint (T0379)Q74 (T0379)L84 3 Then I also included the other helix and sheet constraints from try1-opt2 as well as commenting out the dssp-ehl2 and the undertaker-align constriants. try2 is currently running on shaw. Mon Jul 24 14:16:56 PDT 2006 Cynthia Hsu Looking at try1 and try2 superimposed, it seems that Crissan's added helix constraints made little change to the overall structure of the regions addressed. Examining try2-opt2 in the near, burial, and rr scripts, I found that the region in question actually did not have any rr constraints, was almost entirely predicted to be helices, but existed mostly as coils. For try3, I modified the helical constraints as shown below, in hopes of producing a slightly unique fold: //helix constraints determined from best-models HelixConstraint (T0379)H15 (T0379)E57 3 HelixConstraint (T0379)R60 (T0379)L71 3 HelixConstraint (T0379)Q74 (T0379)E86 3 However, our best plan of action may be to produce a chimera of regions 15-86, which I will do so momentarily. try3 is currently running on abyss. The make on subdomain H15-E86 is currently running on squawk. Tue Jul 25 10:58:30 PDT 2006 Crissan Harris, Cynthia Hsu The subdomain alignment appeared identical to our first three generated models. Because of this, we decided to refrain from making a chimera. When examining the Undertaker alignments in best-models.pdb.gz, we found that the helices that were most similar to our automatically generated tries were those of the second most favored alignment. Because of this, we decided to, for try4, generate a model using only the second alignment, 1x42A. try4 is currently running on lopez. Wed Jul 26 08:28:35 PDT 2006 Crissan Harris try4 scored much worse than all of the previous tries in both undertaker's unconstrained and constrained scores as well as Rosetta's scores. It also had several very large breaks. I am going to try running it again from this same alignment(1x42A) trying to close the breaks and adding a few rr-constraints to help with the burial. The dry weights were changed as follows: dry5 20 dry6.5 25 dry8 20 dry12 7 "sidechain" was reduced to 4, "n_c_ac" was raised to 6, "soft_clashes" was raised to 30 and "break" to 80. Then I removed "maybe_metal" and "maybe_ssbond". The following rr constraints were added: Constraint F153.CB T179.CB -10. 7.0 14.0 0.5 Constraint F169.CB A180.CB -10. 7.0 14.0 0.5 try5 is currently running on vashon. Wed Jul 26 10:04:49 PDT 2006 Crissan Harris I was looking at the score-all+servers.unconstrained.pretty file and saw that SAM_T06_server_TS1, and ROBETTA_TS1 are scoring pretty well compared to our alignment models. We may want to consider taking a look at these if they continue to score well. Wed Jul 26 10:41:22 PDT 2006 Crissan Harris, Cynthia Hsu Looking at the server models, they aren't remarkably different from the models we generated, so there doesn't seem to be a need to make any changes to our models at the moment. We'll keep this for future reference if we don't have five unique models to submit. Wed Jul 26 11:03:08 PDT 2006 Cynthia Hsu In examining try5, there are still very large breaks before D78 (3.13) and E181 (1.31). If we want to keep working with this particular alignment, we will need to add distance constraints to keep residues 77 and 78 together. The added rr constraints did seem to have the intended effect. I copied try5.costfcn to try6.costfcn, then raised "break" to 100 and "soft_clashes" to 40. I then added the following constraint, hoping to close the break: Constraint Y77A.CA D78.N -1. 0.5 2 1 Like try5, try6 used the 1x42A alignment. try6 is currently running on lopez. Wed Jul 26 14:27:55 PDT 2006 Cynthia Hsu try6 seems to have made the breaks worse, instead of better. I've decided to leave out the distance constraint from try7.costfcn, seeing if the higher weight on "break" alone would do the trick. I also lowered "constraints" to 9, and "n_ca_c" to 5. As did tries 5 and 6, try7 used the 1x42A alignment. try7 is currently running on camano. Wed Jul 26 20:38:28 PDT 2006 Cynthia Hsu try7 closed the breaks, as hoped. All breaks have been reduced to a cost of less than 0.5. I can't view it in RasMol, but viewing the unconstrained scores, it seems to have done worse than tries 4 and 5. This is specifically in the areas of "dry8", "dry12", and "n_ca_c", as well as "phobic_fit" although I still don't know a particularly good way of improving that, so I will just leave it at the moment. I raised "dry6.5" to 30, "dry8" to 25 and "dry12" to 8. "n_ca_c" was raised to 10 and "bad_peptide" was raised to 12. I also commented out dssp-ehl2 constraints, untertaker alignments, and rr constraints. try8 was a polsihing run. try8 is currently running on lopez. Thu Jul 27 10:46:29 PDT 2006 Crissan Harris try8 scored better than tries 4, 5, 6, and 7, but still worse than tries 1, 2, 3. The places where try8 could still use some inprovement are dry8, dry12, phobic-fit, n_ca_c. I will do another polishing run copying the costfcn from try8 and increasing the weights on each of these sligtly. try9 is currently running on orcas. Thu Jul 27 13:35:13 PDT 2006 Crissan Harris, Cynthia Hsu try 9 performed better than try8. We've decided that for the sake of generating a new model, and because it was listed in the first README entry by Professor Karplus, to do an alignment that is just based on the top BLAST match (2b0cA). We copied the basic try4.costfcn to try10.costfcn, with no changes. try10.underincluded the alignments from 2b0cA. try10 is currently running on lopez. Thu Jul 27 16:05:38 PDT 2006 Crissan Harris try10 created a fold that was similar to our current top scoring models except with some variation in the C-terminal helix. This new conformation actually looks really good in rasmol under the near script and pretty good under the burial script. Although try10 is in the top five best models according to Undertaker's unconstrained scores, it had a few significant breaks. I am going to do another run with this same alignment (2b0cA) trying to close the breaks. The dry weights were changed as follows: dry5 20 dry6.5 25 dry8 20 dry12 6 "bad_peptide" was raised to 12, "soft_clashes" to 60 and "break" to 120. I also removed"maybe_ssbond" and "maybe-metal". try11 is currently running on peep. Fri Jul 28 10:25:12 PDT 2006 Crissan Harris, Cynthia Hsu try11 did much worse than try10, although its scores in breaks were improved slightly. We've decided to try to do one more run closing the breaks, as there are still some fairly large ones (with cost of 1 or 2). try12.costfcn was a copy of that of try11, but with "dry5" at 25, "dry6.5" at 30, and "sidechain" at 4. "n_ca_c" was raised to 10, "soft_clashes" to 80, and "break" to 150, and the weight in "constraints" was lowered to 8. try12 is currently running on lopez. Fri Jul 28 12:07:43 PDT 2006 Crissan Harris, Cynthia Hsu try12 did much worse than try10 and try11. We decided to copy its costfcn to try13.costfcn, then perform a polishing run on try11-opt2. try13 is currently running on lopez. Fri Jul 28 13:51:04 PDT 2006 Cynthia Hsu try13 did slightly worse than try10, but was much improved over the previous polishing runs. The main difference between the two models was in terms of "n_ca_c". Because of this, I raised both that and "bad_peptide" to 15. I also remembered to comment out the dssp-ehl2 constraints, the undertaker alignments, and the rr constraints this time, unlike in previous runs. This may make a difference in the score. try14 is currently running on lopez. Fri Jul 28 14:17:18 PDT 2006 Cynthia Hsu I decided, while we wait for try14, it may be worth it to do a general polishing run (using include read-pdb.under). For try15.costfcn, I copied try2.costfcn, then raised "dry5" to 20, "dry6.5" to 25, "phobic_fit" to 2.5, "n_ca_c" to 10, and "bad_peptide" to 17. I also raised "soft_clashes" to 60 and "break" to 150. "hbond_geom" was raised to 6, "hbond_geom_beta" to 60, and "hbond_geom_beta_pair" to 120. I also removed "maybe_metal", "maybe_ssbond", and commented out rr constraints. try15 is currently running on vashon. Fri Jul 28 15:31:58 PDT 2006 Crissan Harris, Cynthia Hsu try14 scored better than try10 (in other words, achieved the desired goal). We decided that while we were waiting for try15 to finish, we ought to go ahead and polish the other top-scoring models (namely try1 and try3). We copied try1.costfcn to try16.costfcn, then commented out the dssp-ehl2, the undertaker alignment, and rr constraints. We also removed "maybe_metal" and "maybe_ssbond". We raised "dry5" to 20, "dry6.5" to 25, "dry8"to 20, "n_ca_c" to 8, "bad_peptide" to 15, "soft_clashes" to 80, and "break" to 200. We also raised "hbond_geom" to 7, "hbond_geom_backbone" to 15, "hbond_geom_beta" to 60, and "hbond_geom_beta_pair" to 120. try16.under used ReadConformPDB T0379.try1-opt2.pdb try16 is currently running on camano. We considered doing a polishing run on try3, but looking at it in our best-models, we found that it was very similar to try2. Because of this, we decided that we would probably not be submitting it anyways, so there was no point in polishing try3. Sun Jul 30 14:51:43 PDT 2006 Cynthia Hsu try16 is currently our top scoring model. In view of this, I decided to polish try2. try17.costfcn was a polishing run of try2. I commented out the included rr constraints, raised "dry5" to 20, and "dry6.5" to 23. "maybe_metal" and "maybe_ssbond" were both removed. "phobic_fit" was raised to 7, "bad_peptide" to 15, "soft_clashes" to 40, and "break" to 150. "constraints" was raised to 12, and "bond_geom" to 7, "hbond_geom_backbone" to 15, "hbond_geom_beta" to 60, and "hbond-geom_beta_pair" to 110. try17.under used try2-opt2 as the input PDB. try17 is currently running on Whidbey. Mon Jul 31 10:20:54 PDT 2006 Cynthia Hsu While reviewing our top scoring models, I realized that try9 had not been polished enough. I copied over the try9 cost function to try18, then made the following changes: "dry6.5" to 25, "dry8" to 20, "n_ca_c" to 10, "bad_peptide" to 15, "soft_clashes" to 50, "break" to 120, "hbond_geom" to 10, "hbond_geom_backbone" to 20, "hbond_geom_beta" to 80, and "hbond_geom_beta_pair" to 130. try18.under used ReadConformPDB T0379.try9-opt2.pdb. try18 is currently running on lopez. I updated the superimpose-best.under file with our current top models. While doing this, I noticed that our best models still score poorly on the breaks, so in the future, more polishing runs will be necessary. Our current top models are as follows: ReadConformPDB T0379.try17-opt2.pdb //try2, polished ReadConformPDB T0379.try16-opt2.gromacs0.repack-nonPC.pdb //try1, polished ReadConformPDB T0379.try14-opt2.pdb //try11<2bc0A, polished ReadConformPDB T0379.try15-opt2.pdb //read-pdb.under polishing ReadConformPDB T0379.try9-opt2.pdb //1x24A alignment. Mon Jul 31 10:57:02 PDT 2006 Cynthia Hsu I decided to polish the top scoring Rosetta model. try16.costfcn was copied to try19.costfcn. "dry5" was raised to 25, "dry6.5" to 30, and "dry8" to 25. I raised "n_ca_c" to 10 and "bad_peptide" to 20. "soft_clashes" was raised to 80, "break" to 240, "hbond_geom" to 10, "hbond_geom_backbone" to 20, "hbond_geom_beta" to 80, and "hbond_geom_beta_pair" to 140. try16-opt2.gromacs0.repack-nonPC.pdb was the input pdb. try19 is currently running on shaw. The high cost of "soft_clashes" and "break" was prevalent in all of our top scoring models. I decided to have try20.under use the read-pdb.under script to polish all current models, hoping that would produce satisfactory results. try20.costfcn was a copy of try17.costfcn, our top scoring model. I raised "dry8" to 20, "phobic_fit" to 3, "n_ca_c" to 10, "bad_peptide" to 17, "soft_clashes" to 60, and "break" to 140. try20 is currently running on orcas. Mon Jul 31 15:15:46 PDT 2006 Crissan Harris, Cynthia Hsu try20-opt2 still performs rather badly in breaks. We copied the try20.costfcn to try21.costfcn, raised "break" to 200, then lowered "hbond_geom" to 4 and "hbond_geom_backbone" to 8. "hbond_geom_beta_pair" was also decreased to 80. try21.under used try20-opt2.pdb as its input. try21 is currently running on shaw. For try19, we decided to polish the gromacs repacked version because the breaks there were fewer and smaller. "break" raised to 280, "hbond_geom" lowered to 5, "hbond_geom_backbone" to 10, "hbond_geom_beta" to 50, and "hbond_geom_beta_pair" to 80. try22 is currently running on orcas. try18 did fairly well on breaks, even though it ranked among the worst-scoring of our top models. The superimpose-best.under file was updated accordingly: ReadConformPDB T0379.try20-opt2.pdb //read-pdb.under, polished #ReadConformPDB T0379.try17-opt2.pdb //try2, polished - identical to try20 ReadConformPDB T0379.try19-opt2.gromacs0.repack-nonPC.pdb //try1, polished ReadConformPDB T0379.try14-opt2.pdb //try11<2bc0A, polished ReadConformPDB T0379.try15-opt2.pdb //read-pdb.under polishing - surprisingly, looks completely different from all other models, but very similar to top alignments. ReadConformPDB T0379.try18-opt2.pdb //polished try9, from 1x24A alignment Mon Jul 31 17:11:46 PDT 2006 Crissan Harris, Cynthia Hsu try22 still has one significant break. try21 has not finished running yet. We copied try22.costfcn to try23.costfcn, then raised "break" to 310. "dry6.5" was raised to 35, and "dry8" to 28. try23.under used try22-opt2 as its input. try23 is currently running on lopez. Tue Aug 1 10:16:22 PDT 2006 Cynthia Hsu try23 has a "break" score of 1.1. I am considering leaving this as is. try21 had a "break" score of 1.4. I decided to, for try24, copy try21.costfcn and raise "break" to 240, "soft_clashes" to 80, and lower "hbond_geom_beta" to 50. try24.under used try21-opt2 for its input file. try24 is currently running on lopez. In summary of our current top models: try21 < try20-opt2 < read-pdb.under (similar to try2) try23-opt2 < try22-opt2 < try19-opt2.gromacs0.repack-nonPC < try1-opt2 < 2fghA, 2b0cA try14-opt2 < try11-opt2 < 2bc0A try15-opt2 < try9-opt2 < try8-opt2 < try7-opt2 < 1x42A try18-opt2 < try9-opt2 < try8-opt2 < try7-opt2 < 1x42A Given that Rosetta currently favors try22-opt2.gromacs0.repack-nonPC, I've decided to make one more attempt to close the break, starting from this one as the polishing run. I copied try22.costfcn to try25.costfcn, then made the following changes: "dry5" was raised to 30, "dry6.5" to 35, and "dry8" to 28. "bad_peptide" was raised to 25, and "break" to 320. try25.under used try22-opt2.gromacs0.repack-nonPC as its input PDB. try25 is currently running on orcas. Tue Aug 1 13:14:25 PDT 2006 Cynthia Hsu try24 finished with no improvement in the breaks, and the overall score is worse than that of try21. I decided to copy try24.costfcn to try26.costfcn, then raise "soft_clashes" to 100 and "break" to 280. As before, try21 was the input pdb for the try26.under file. try26 is currently running on shaw. try25 scored poorly with both Undertaker and Rosetta. I've decided to leave this model as is. Tue Aug 1 15:40:17 PDT 2006 Cynthia Hsu The polishing run of try26 produced a higher scoring model, but only just barely. The iprovement in breaks is also negligible. I copied try26.costfcn to try27.costfcn. "soft_clashes" was raised to 110 and "break" to 320. "hbond_geom" was lowered to 3 and "hbond_geom_beta_pair" to 70. try27.under is set up to ReadConformPDB T0379.try26-opt2.pdb. I'll run this later this evening, when the processors are not as full. Tue Aug 1 21:49:43 PDT 2006 Cynthia Hsu try27 currently running on lopez. Wed Aug 2 10:17:55 PDT 2006 Cynthia Hsu try27 had a "break" score of 1.1. Our current top models are as follows: try27-opt2 < try26-opt2 < try21 < try20-opt2 < read-pdb.under (similar to try2) try23-opt2 < try22-opt2 < try19-opt2.gromacs0.repack-nonPC < try1-opt2 < 2fghA, 2b0cA try14-opt2 < try11-opt2 < 2bc0A try15-opt2 < try9-opt2 < try8-opt2 < try7-opt2 < 1x42A try18-opt2 < try9-opt2 < try8-opt2 < try7-opt2 < 1x42A However, the second choice model may be the gromacs repacked version of try22, as this is most favored by Rosetta. I've decided to make one more run, using the read-pdb.under script to polish all models, and the try27.costfcn. try28 is currently running on lopez. Wed Aug 2 15:14:14 PDT 2006 Crissan Harris, Cynthia Hsu try28 is now our top scoring model. As it looks identical to try27, we've decided that it would replace it as our top model. This is a summary of our top selections: try21 < try20-opt2 < read-pdb.under (similar to try2) try23-opt2 < try22-opt2 < try19-opt2.gromacs0.repack-nonPC < try1-opt2 < 2fghA, 2b0cA try14-opt2 < try11-opt2 < 2bc0A try18-opt2 < try9-opt2 < try8-opt2 < try7-opt2 < 1x42A try22-opt2.gromacs0.repack-nonPC Our selections are as follows: Model 1 is try28-opt2. This was a polishing run of the following: try28-opt2 < try27-opt2 < try26-opt2 < try21 < try20-opt2 < try17-opt2 < try2-opt2 < 2bc0A 2bc0A was Undertaker's choice as the top alignment. This model was generated from the same cost function as try1, but with helix constraints added in an effor to create a diverse model that followed the dssp-ehl2 constraints more accurately. It is currently the top favorite of Undertaker. Model 2 is try23-opt2. This was a polishing run of the following: try23-opt2 < try22-opt2 < try19-opt2.gromacs0.repack-nonPC < try1-opt2 < 2fghA, 2b0cA try19-opt2.gromacs0.repack-nonPC was selected because at the time it was the most favored model according to Rosetta. Model 3 is try14-opt2. It was a polishing run of the following: try14-opt2 < try11-opt2 < 2bc0A This generated a somewhat diverse model, as only one of the two alignments that Undertaker had originally selected were used. It was a full-length hit on BLAST, with 25% sequence identity and an e-value of 7.5e-10. Model 4 is try18-opt2. It was polished from the following: try18-opt2 < try9-opt2 < try8-opt2 < try7-opt2 < 1x42A Although it did not score well, it did produce a slightly different model from the previous results. It was the second model used by Undertaker to produce the best-models.pdb.gz file, and the helices most resembled those that had been produced by try1-opt2. Model 5 is try22-opt2.gromacs0.repack-nonPC. It was a gromacs repacked version of the following polishing run: try22-opt2.gromacs0.repack-nonPC < try19-opt2.gromacs0.repack-nonPC < try1-opt2 < 2fghA, 2b0cA Although virtually identical to model 2, we have chosen to include it because it is Rosetta's favorite model, and because we do not have any other diverse models to include. Thu Aug 3 16:24:53 PDT 2006 Grant Thiltgen I'm getting this target ready for submission. I am going to remove model 5 from the submission just because it is virtually identical to model 2. I will look through the models and see if we have anything different, and if not, I'll just submit an alignment. Yeah, i think I'll just submit an alignment. Actually, looking at everything, I'm going to start a new try from alignments using the 2fdrA alignments, which are the favored undertaker alignments to see what results I might get before submission. Try29 started on camano. Thu Aug 3 22:47:06 PDT 2006 Grant Thiltgen The model built from the 2fdrA alignments isn't as good as the other models at all, but I think it's better than submitting just an alignment. Model 1 is try28-opt2. This model was a very polished model based on the 2bc0A alignments. This model was created using constraints based on the predicted secondary structure elements from our neural networks. This scored the best on our unconstrained costfcn for undertaker. Model 2 is try23-opt2. This model was also a polished model based on the 2b0cA alignments, but this model was initially started from try1-opt2 which is the fully automatic results from undertaker. We also chose this model to continue with because an early model scored very well with Rosetta, so we worked on polishing this model. Model 3 is try14-opt2. It was a polished model from alignments that were also based on 2b0cA. This did generate a slightly different model. This model was solely based on the 2b0cA alignments becasue it had a full-length hit on BLAST with an e-value of 7.5e-10. Model 4 is try18-opt2. It was a polished model from alignments based on 1x42A. This model scores poorly in comparison to the other three models, but it gives a little bit of diversity in the models, as it was created by the second best alignments from the undertaker alignment file, and the helices most resembled the ones from the initial undertaker run. Model 5 is built from alignments to 2fdrA. The model scores poorly in comparison to the top three models, but it is included because 2fdrA created our best scoring alignments in undertaker. It has some of the helices packed differently than the best three models. Models submitted 10:49 pm on August 3rd, 2006. Thu Nov 2 11:12:56 PST 2006 Kevin Karplus Model 5 did turn out a bit better than align1, but align4 and align3 were better than model 5. Model 1 was our best submitted (only try28-opt1 and try27-opt2 beat it of our models). The biggest problem is that we straightened out the loop for D38-K43, pushing the helix A27-D32 back and displacing the R18-I28 helix. Although there were a few better models, none of the model 1 submissions by servers was better. (Our servers did not do well on this target.)