Tue Jun 22 09:43:23 PDT 2004 T0212 DUE 27 July Tue Jun 22 12:54:40 PDT 2004 Kevin Karplus This one looks like a new fold. Tue Jun 22 20:26:59 PDT 2004 Kevin Karplus try1 looks like it is trying to form a beta sandwich. I'll try to help it along a bit in try2 with sheet constraints. The try1 model looks better than any of the 5 alignments that were used as starting points. I hope that using all-align.a2m will give an even richer and more productive starting set. Tue Jun 22 22:25:31 PDT 2004 Kevin Karplus try2 does not look any better than try1. It may even be worse. It looks like a big part of the model is coming from the alignment to 1o91A, which is not classified in SCOP 1.65 (but may be in the beta-release version of SCOP---should we take the time to look?) I've modified try3.costfcn to be more compatible with this template, and will run again. Wed Jun 23 07:38:45 PDT 2004 Kevin Karplus I don't like try3 all that much---the hairpin around L46-I53 seems particularly poor. I thought, at first, that there was a knot in the structure, but it just misses being a knot. For try4, I'v tried a little harder by hand to put together reasonable strand pairings. Mon Jul 5 21:28:33 PDT 2004 Kevin Karplus Been a long time since I looked at this target! Try4 is looking sort of ok---the strands aren't quite forming, but they're laying out in a feasible way. I'll see if there is a fold that looks like this by using VAST job: VS59501 password: T0212try4 Nope, VAST did not find any structure neighbors. If the fold in try4 is right, this is a new fold. From martina@soe.ucsc.edu Tue Jul 6 04:21:49 2004 MIME-Version: 1.0 Date: Tue, 6 Jul 2004 04:21:43 -0700 (PDT) From: Martina Koeva To: karplus@soe.ucsc.edu cc: Martina Koeva Subject: T0212 Hi Kevin, as I was looking at T0212, I saw in the README file that you had mentioned 1o91, which cannot be found in neither SCOP, nor CATH. I looked at PDBSum and PubMed and there is a paper that was published last year on the structure of 1o91A. I wasn't sure whether you had looked at it or not. 1o91A appears to be a collagen alpha1(viii) NC1 trimer, where each of the monomers is a ten-stranded beta sandwich. The formation of the trimer buries a solvent-accessible surface and as a result a central solvent-filled channel is formed. "The biological function of collagen VIII remains unclear. Initial work suggested that collagen VIII networks may function as a matrix that resists compression and maintains an open porous structure, and is closely associated with basement membranes." (excerpt from the paper) >From what I can see in the t04 dssp-ehl2 prediction, there are exactly 10 predicted beta strands and some of the other secondary structure predictions also show lots of beta strands. The t04 CB-burial-14-7 prediction is showing a large number of strongly predicted exposed patches. Also there are a number of residues mentioned in the paper that I believe I can see in our target, but I am not sure whether a mapping between them would be valid... Here is a link to the paper: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=pubmed&dopt=Abstract&list_uids=12782141 I can also put it in the directory for the target tomorrow (today) and add it to the README file. I will also bring the paper to the meeting this morning. Sincerely, Martina ------------------------------------------------------------ Tue Jul 6 07:24:44 PDT 2004 Kevin Karplus 1o91A is in the casp6 beta release of scop: unknown-sid 1o91 A: b.22.1.1 -3282 cl=48724,cf=49841,sf=49842,fa=49843,dm=69230,sp=-3281,px=-3282 I have to fix the Make.main file (and probably some scripts, so that they can use the beta release of SCOP). Thu Jul 8 15:01:08 PDT 2004 Martina Koeva I have started work on T0212. I ran DALI to look for structural similarities with other proteins and here is the list of the top proteins T0212 was found to be structurally similar to: 1: 5008-A 1jh5-A 4.9 3.5 95 144 7 0 0 8 S tumor necrosis factor ligand superfamily member 13b fr 2: 5008-A 2tnf-A 4.2 2.6 85 148 11 0 0 10 S tumor necrosis factor alpha fragment (tnf-alpha, cachec 3: 5008-A 1aly 4.2 2.7 86 146 8 0 0 9 S cd40 ligand fragment 4: 5008-A 1rj7-A 4.1 2.6 85 144 11 0 0 8 S ectodysplasin a fragment (ectodermal dysplasia protein, I have just included the list of the ones that had Z-scores higher than 4.0. I tried modeling T0212 after 1jh5, which seems to be a beta-sandwich (jelly-roll fold), just like some of the other templates. For try5, I added some sheet constraints to get the same strand alignment as the corresponding strands on 1jh5: Sheet Constraint A43 H50 V103 Y97 Hbond T99 7 Sheet Constraint Q42 K45 V68 G65 Hbond A44 4 Sheet Constraint L46 R49 V63 T60 Hbond E47 4 Fri Jul 9 14:41:38 PDT 2004 Martina Koeva The result from try5 looked pretty bad. What I really want is to take the structural alignment between T0212 and 1jh5 and try to map the strands to fit into the sheet arrangement. Here is the actual alignment, provided by DALI: NR. STRID1 STRID2 STRID1 <=> STRID2 STRID1 <=> STRID2 1: 5008-A 1jh5-A 5 - 8 <=> 15 - 18 (ASP A 5 - ILE A 8 <=> PRO A 15 - GLN A 18) 1: 5008-A 1jh5-A 9 - 12 <=> 22 - 25 (ARG A 9 - VAL A 12 <=> TYR A 22 - VAL A 25) 1: 5008-A 1jh5-A 14 - 17 <=> 43 - 46 (THR A 14 - ILE A 17 <=> LYS A 43 - VAL A 46) 1: 5008-A 1jh5-A 23 - 54 <=> 47 - 78 (PRO A 23 - THR A 54 <=> LYS A 47 - VAL A 78) 1: 5008-A 1jh5-A 60 - 68 <=> 79 - 87 (THR A 60 - VAL A 68 <=> PHE A 79 - THR A 87) 1: 5008-A 1jh5-A 70 - 77 <=> 90 - 97 (GLY A 70 - PRO A 77 <=> ARG A 90 - GLU A 97) 1: 5008-A 1jh5-A 78 - 104 <=> 99 - 125 (ASN A 78 - SER A 104 <=> LEU A 99 - GLU A 125) 1: 5008-A 1jh5-A 109 - 115 <=> 127 - 133 (HIS A 109 - LYS A 115 <=> ALA A 127 - GLY A 133) I am not sure whether this template is worth pursuing, but I want to at least try using it as a model, as the strands in T0212 seem to want to follow the arrangement of the strands in the 2 sheets of 1jh5. There is a particular mapping that I will try to get first in the T0212 model: 1: 5008-A 1jh5-A 23 - 54 <=> 47 - 78 (PRO A 23 - THR A 54 <=> LYS A 47 - VAL A 78) 1: 5008-A 1jh5-A 78 - 104 <=> 99 - 125 (ASN A 78 - SER A 104 <=> LEU A 99 - GLU A 125) Both of these segments in 1jh5 (47-48 and 99-125) incorporate 2 strands that belong to the different sheets of the beta sandwich. Also they seem to be next to each other positioned in an antiparallel fashion. In particular, the strands G50-Y60 and C104-L112 are antiparallel, as well as the strands A66-K74 and E117-I122. I want to try and see those 4 strands in T0212. Some of the constraints for those strands are already in the costraints file. For try6, I put in an additional sheet constraint with a higher weight to try and get a pairing of long strands on both sides: SheetConstraint D25 I36 P92 Y81 hbond Y32 10 SheetConstraint A43 I52 V103 G94 Hbond T99 10 SheetConstraint R9 I17 V103 I95 Hbond Y100 2 I also turned down the weight on this last sheet constraint and I removed the sheet constraints I introduced in try5. Additionally, I turned off phobic_fit. Mon Jul 12 16:42:47 PDT 2004 Martina Koeva Try6 did not really achieve what I wanted, which is making me think that the contsraints for the sheet that I introduced were wrong. However, I have decided to pursue this path a little longer and decided for try7 to remove the constraint that paired the middle strand that I am interested in (A43-I52) to its other neighbor E62-V69, where the pairing of the strands had already worked. If this does not work, I would have to go back and reintroduce the old sheet constraints. For try7 I also introduced two additional general constraints from the residue-residue predictions, hoping that one of them at least would pull out and unwind one of the helices that seem to be forming right now. Looking at the literature, from the templates that we have so far and from the structural alignment with DALI that I ran on T0212, it seems that most of the templates and most of the structural similarities with other proteins point to a TNF-like fold. Given that none of the t2k and t04 predictions seem to find helices anywhere in the protein and that the TNF-like fold consists of all-beta proteins that follow a beta-sandwich structure (2 sheets), forming a beta-jelly roll, I will be aiming to remove the helices. This is especially relevant to the segment at the end (C-terminal end) of the target, which is currently folding into a helix and is predicted to be a strand. Several comments regarding the literature: 1. There is an extended loop/coil between residues 70 and 80 (G70-A80) in T0212, which right now does not seem to be sticking out of the protein too much. However, the literature on TNF proteins seems to indicate the presence of such a loop as a flexible region, towering over the top (Idriss 2000), so I might try to model that in the next few tries. 2. I was worried about the general shape of the model, but it seems that a large number of the TNF proteins in their monomeric form are about 60 Angstroms long and 30 Angstroms wide (Idriss 2000), so I am getting more convinced that I should try and keep the long strands. Wed Jul 14 09:31:10 PDT 2004 Kevin Karplus We need to make the N-terminal and C-terminal strands attach to sandwich---they are curling up. Mon Jul 19 10:56:19 PDT 2004 Martina Koeva try8 introduced two new SheetCosntraints to make the N-terminal and C-terminal strands attach to sandwich. I also increased the weight for hbond-beta from 0.4 to 0.8. Tue Jul 20 15:04:00 PDT 2004 Martina Koeva So I started try9 with the intention to pursue the jelly roll fold topology. For now I mainly want to bring the strand between A43-T54 down next to the other strands. In order to do that I introduced two general constraint to try and get the turns to align more or less at the same level. I also increased the weight on the sheet constraint between strands A43-W51 and V103-I95 from 3 up to 7. I was hoping that this would further encourage the strand A43-T54 to be pulled down from where it is in try4. I also introduced a new strand constraint at the very end to try and get the helix that is forming at the C-terminal to unwind, in which case the N-terminal and C-terminal strands might actually manage to form their hydrogen bonds. I increased the weight on hbond_geom_beta -> from 0.8 to 1 as well as the weight on break -> from 20 to 40 Wed Jul 21 10:53:07 PDT 2004 Martina Koeva try9 mysteriously failed...I did not find anything in the log file and no pdb files were generated. So I think I will start it again and keep all the files the same. Somebody might have restarted the machine that I was using to run undertaker on (heehaw or bark). Thu Jul 22 15:28:59 PDT 2004 Martina Koeva So try9 failed again, but it successfully finished on the third attempt. For try10 the only change that I made was in the increase of the break weight. It seems I have been somewhat successful in getting the sheet on one side to form, which is what I was hoping for. However, I still haven't managed to get undertaker to unwind the C-terminal strand. I also think I want the strand 66-69 to shift a little bit so that the predicted segment aligns with the other strands. try9-opt2 and try10-opt2 seem to be scoring the best so far, a little bit better than try4-opt2. I will use the try10-opt2 structure to submit it to VAST again and see whether any structures come up from there. Meanwhile, for try11, I turned down the break weight, increased the weight for all sheet constraints and included the latest rr predictions (from T0212.280.rr.constraints file). Thu Jul 22 16:38:32 PDT 2004 Martina Koeva Submitted try10-opt2.pdb to VAST and still waiting for results. ID: VS60013 Password: T0212try10 Fri Jul 23 13:51:10 PDT 2004 Kevin Karplus The VAST run is done. The hits are for 1poqA, 1c28A, 1nqjB=1nqjA, 1bmv2, 1umhA, 1nlqA, 1gr3A, 1wkt I picked up the vast alignments (all 8), added the templates to the TOP_MANUAL_HITS list, and ran "make extra_alignments" to get a full set of alignments. The I ran "make read_alignments all-align.a2m.gz" to make the scripts and alignment file for using the alignments. I created an "unconstrained.costfcn" and scored all existing models with it. It likes try1-opt2 best, then try7-opt, try10-opt2, robetta-model2, try4-opt2, ... I created a "strands" rasmol file to identify the strands: define s1 8-14 define s2 27-38 define s3 43-54 define s4 60-69 define s5 81-89 define s6 95-104 define s7 109-114 define s8 117-120 define beta s1 or s2 or s3 or s4 or s5 or s6 or s7 or s8 try1 looks pretty good, except that s7 needs to run antiparallel to s2. try7 looks ok, but s1 is wound up into a helix, as is s8. The hbonding for the strands is not very good, so they don't for proper sheets. try10 looks ok, but s7 and s8 are wound into helices. try 10 is trying to get s1 ^v s6 ^v s3 ^v s4 and s5 ^v s2 ^v s7+s8 The jelly-roll motif is a curled up hairpin, so has s1 ^v s6 s2 ^v s5 s3 ^v s4 We are missing a strand (s0) to pair with s7+s8. It is possible that s3 and s4 are really two strands each: SheetConstraint H50 A56 G65 K59 SheetConstraint E41 L46 T74 V69 Hbond A44 Then we have enough strands: s1 ^v s6 s2 ^v s5 s3a ^v s4b s3b ^v s4a If that is a jelly roll, then we need also s1 ^v s2 s3a ^v s6 s3b ^v s5 s4a ^v s4b and s7+s8 are left out in the cold. I've put the Branden and Tooze pp.77-78 coloring scheme for the jelly roll into the strands script: # color as jelly roll select s1 or s6 color green select s2 or s5 color red select s3a or s4b color blue select s3b or s4a color yellow This is pretty close to what try1 was trying to do, except that the s3a ^v s4b pairing is missing, so it might better be seen as select s0 or s7 or s8 color green select s1 or s6 color red select s2 or s5 color blue select s3 or s4 color yellow It looks like we have a choice of going with a jelly-roll + a strand (the try10 style) or jelly-roll minus first strand (try1 style). Both seem feasible to me. I'll try improving on the try1 style, since Martina has been working on the try10 style. try12 is my first guess at strand constraints to improve on try1. I also used Rosetta to repack sidechains for try1, try8, try10, and try11. Rosetta likes try1-opt.repack-nonPC best currently, and thinks the others are terrible (probably because of clashes). (The command I use for looking at Rosetta scores is grep total *repack* | sort -g +13 in the decoys/ directory---it only works if the repack-nonPC.pdb files have not been gzipped.) Sat Jul 24 08:15:11 PDT 2004 Kevin Karplus try12-opt2 scores best with the try12 costfcn, try1 far behind, then try10. try12 looks pretty good overall, but isn't forming the H-bonds. I'll try adjusting the strands +- 1 or 2 residues to try to get the hydrophobic side buried and be consistent with the shape of try12. try12 scores best with the try13 costfcn, followed distantly by try6, try10, try4, try8, ... Sat Jul 24 15:21:45 PDT 2004 Kevin Karplus The overall shape of try13 is ok, but some strands are flipped the wrong way. I8,V10,V12 should be buried, as should M96 and M102. I probably have the hbonds spec'ed wrong. The bulge I was trying to create in strand 3 did not form, which may have contributed to the problem. I'll try a bit harder to form that bulge. Sat Jul 24 19:58:58 PDT 2004 Kevin Karplus try14 made almost no change to try13---tiny improvement in cost, but no major change to the shape or the hbonding. With the unconstrained.costfcn, the order is try1, try7, try14, try10, try13, robetta-model2 Perhaps I need to stiffen the constraints some to get some movement? The highest-cost constraints in try14-opt2 are constraint(L46.CA, R49.CA) [> 5 = 6.54 < 7] # cost= 6.132 dist= 9.34 constraint(T54.CB, E62.CB) [> 5.08 = 5.58 < 6.08] # cost= 4.915 dist= 8.53 constraint(I53.CB, M96.CB) [> 5.08 = 5.58 < 6.08] # cost= 4.502 dist= 8.35 constraint(H50.CB, A66.CB) [> 5.08 = 5.58 < 6.08] # cost= 3.829 dist= 8.05 constraint(T54.CA, E62.CA) [> 3.93 = 4.43 < 4.93] # cost= 3.491 dist= 6.69 constraint(T48.CB, V68.CB) [> 5.08 = 5.58 < 6.08] # cost= 3.385 dist= 7.85 constraint(V10.CB, Y100.CB) [> 5.08 = 5.58 < 6.08] # cost= 3.214 dist= 4.15 constraint(I53.CA, M96.CA) [> 3.93 = 4.43 < 4.93] # cost= 3.026 dist= 6.48 constraint(R49.CB, Y100.CB) [> 5.08 = 5.58 < 6.08] # cost= 2.798 dist= 7.57 constraint(T35.CB, N111.CB) [> 4.52 = 5.02 < 5.52] # cost= 2.572 dist= 3.78 constraint(I52.CB, Q64.CB) [> 5.08 = 5.58 < 6.08] # cost= 2.557 dist= 7.46 constraint(M1.N, H126.C) [> -5 = 4 < 10] # cost= 2.549 dist= 17.7 constraint(E47.OE1, T48.OG1) [> -10 = 4 < 6] # cost= 2.492 dist= 9.27 hbond constraint(T35.N, N111.O) [= 2.87] # cost= 2.131 dist= 4.23 hbond constraint(N111.N, T35.O) [= 2.87] # cost= 2.125 dist= 4.01 constraint(E47.CB, V68.CB) [> 5.08 = 5.58 < 6.08] # cost= 2.076 dist= 3.46 I can either try harder to get the bulge, or I could give up on it and slide half the strand up one. Perhaps I should give up on it for now---I don't have good ways of creating bulges. I'll pair T48-I53 with V103-G98 (hbonding I52). Sun Jul 25 07:48:50 PDT 2004 Kevin Karplus After I gave up on trying to create the bulge, suddenly the bulge appears in try15---not quite the way I wanted it, with E47 and R49 both pointing in, but plausible. Unfortunately M46-V103 seems to be flipped over the wrong way, as are V63 and V69 and I8-T14 I think that the stripes across the sheet should be S7-V103-K45-G70 I8-M102-L46-V69 R9-G101-E47-V68 V10-Y100-T48-G67 E11-T99-R49-A66 V12-G98-W51-G65 T14-M96-I53-V63 E15-I95-T54-E62 OR xxxxxxx-L46-V69 S7-V103-T48-V68 I8-M102-R49-G67 R9-G101-H50-A66 V10-Y100-W51-G65 V12-G98-I53-V63 K13-Y97-T54-E62 T14-M96-xxx E15-I95-xxx I'll try the second of these as try16, the first as try17. Maybe I should salt-bridge E47-R49. That would allow both to be buried. (or would flip both to the surface, which seems more natural). I believe that W51 on s3 is on the right side (buried), but that s4, s1, and s6 are flipped. W51 and I53 hbond to the s4 side. Sun Jul 25 14:46:16 PDT 2004 Kevin Karplus Neither try16 nor try17 forms good beta sheets. Both still have s1 and s6 flipped what I think of as the wrong way. I think that Hbond constraints aren't getting a high enough cost when they are violated---perhaps I've set the tolerances too wide so they don't affect the SheetConstraints much. I'm only getting a small cost at a distance of 4.25 and I'm getting NEGATIVE costs at distances of 7.16. Something is badly broken in the code! Sun Jul 25 16:17:28 PDT 2004 Kevin Karplus Not badly broken, but the scaling needed to be improved. The long tails are not surprising for a Lennard-Jones-style potential. The models from try14 onward are all very similar. I wonder if I should start a new optimization from the alignments, which I haven't done since try12. Sun Jul 25 21:03:44 PDT 2004 Kevin Karplus try18 scores worse than try17 with the try18 costfcn. Other than needing to flip over strand1, I rather like it. Also---it brought the N and C terminus together the "wrong" way---I wanted the N terminus to flip over against the edge of the jelly roll, antiparallel to s7. Maybe I need to be more explicit! The unconstrained cost function orders the models try1, try7, try14, try10, try13, robetta-model2, try4, ... The strands cost function (which has strand constraints, but not sheet constraints), orders them try14, try13, try17, try12, try9, try10, try16, try5, try6, ... The try18 costfcn orders them try17, try18, try14, try13, try16, try15, ... The try14 costfcn orders them try14, try13, try17, try15, try16 The try19 costfcn, which has the hbond_geom weights far too small (Martina seems to have missed the discussion of the rescaling of the Hbond scores), orders them try9, try8, try11, try10, try4, try6, ... Currently, I favor: try17 (or should I use try14?) try10 try1-opt2.repack-nonPC try1 best model from undertaker I'm going to fix up the edge constraints of try18 for try20, and see if I can finally nail this jellyroll. Sun Jul 25 22:59:11 PDT 2004 Kevin Karplus try20-opt1 looks rather foamy and has bad gaps, and s1 is still flipped over. Unless it makes some big improvements in the second round of optimizations, I'm not sure that it is going to be competetive. I was just looking at the best-scoring of the robetta models: robetta-model2. It is a rather different choice of strand order, but quite compatible with the secondary-structure prediction. The model is awfully foamy though. Maybe I should try copying its sheet constraints and trying to make something compatible. With the try21.costfcn, all the robetta models score best, with try10-opt2 quite a way behind. Mon Jul 26 06:28:08 PDT 2004 Kevin Karplus The try19, try20, and try21 runs have finished. Try19 produced a rather ugly jelly roll, rather foamy with lots of predicted buried residues exposed. The try19 cost function scores try9 best---try19 comes after try8, try11, try10, and try4. Try20 is a rather foamy version of try17. The ends are treated a bit better but the sheets are a bit of a mess and the whole thing is a bit foamy. It scores well with the try20 cost function, but not with any of the others. Try21 is very loose and not much in agreement with the secondary structure predictions, but scores best with the unconstrained cost fcn! The try21 cost function still prefers robetta-model2. Maybe we should try polishing that robetta model by optimizing from existing models with the try21 costfcn? [That will be try23.] The unconstrained order is try21, try1, try7, try14, try10, try13, robetta-model2, ... Rosetta likes the repacked version second (after try1-opt2). We should probably include try21 in our predictions, even though it is poor---the secondary structure MAY be better than even our best jelly roll. The strands.costfcn orders them try14, try13, try17, try12, try21, try9, ... Looking more closely at strands.costfcn, I don't like the way it handles s3, so I'll roll back to the s3 definition from t04.dssp-ehl2, which changes the order to try14, try13, try17, try12, try16, try15, ... I'm starting a try22, which has a costfcn like try17, but with some attempt to bring the ends together and a bulge in strand s7. It scores try17 best and try20 second best, so I *hope* it will do something reasonable. Mon Jul 26 10:45:43 PDT 2004 Kevin Karplus try23 is done, try22 is in the super_iteration of the second round. Try22-opt1 doesn't look too promising---it has both strands s1 and s6 the wrong way out. try23 is a lot like robetta-model2, which it is based on. It has s5 as 2 strands, part of s7 turned into a helix, and part of s4 turned into a helix. It is still pretty foamy. Mon Jul 26 11:43:57 PDT 2004 Kevin Karplus Try a cut-and-paste from try17 and try18 (body of try18, ends from try17). Mon Jul 26 16:36:01 PDT 2004 Kevin Karplus Marina and I made a cut-and-paste chimera try17-try18-chimera.pdb REMARK 114-end from try17-opt2 REMARK 1-7 from try17-opt2 REMARK rest from try18-opt2 It has some very bad clashes, and some breaks around where the cut and paste was done, so try24 is trying to polish it up using the same score function as try22. Current list try17 (or try24) try10 try22 try1-opt2 try23-opt2 based on robetta Mon Jul 26 18:48:56 PDT 2004 Kevin Karplus I'll definitely go with try24 as soon as it finishes. In fact, I'll submit try24-opt1 now, just in case. (When it finishes, I should decide which Rosetta-repacked model to include.) Mon Jul 26 21:43:53 PDT 2004 Kevin Karplus try24-opt2 looks pretty good. There are a couple of bad breaks that did not get fixed, but otherwise not bad. try23 still scores best unconstrained, but try24 score best with strands.costfcn I'm rearranging which models we submit and resubmitting everything: ReadConformPDB T0212.try24-opt2.pdb ReadConformPDB T0212.try10-opt2.pdb ReadConformPDB T0212.try1-opt2.pdb ReadConformPDB T0212.try23-opt2.pdb ReadConformPDB T0212.try23-opt2.repack-nonPC.pdb Sun Sep 19 13:04:49 PDT 2004 Kevin Karplus I put 1tzaA in the Makefile as the real hit, and evaluated the whole-chain rmsd. We did beat robetta on this target with the top models being model1, robetta1, robetta3, model2, model3, model4, model5, robetta2, robetta10, robetta7, robetta5, robetta8, robetta6, robetta4, robetta9. But our best model was not one we submitted: try20-opt1 had 9.1091 Ang rmsd, while model1 had 9.7256 (best robetta was 10.4744). Our ordering of submitted models was right. Wed Sep 22 10:16:33 PDT 2004 Kevin Karplus Looking at GDT scores, the order is model1, robetta4, robetta3, robetta10, model2, robetta2, model4, model5, model3, robetta7, robetta1, robetta5, robetta8, robetta6, robetta9. Our best model is try24-opt2 with 30.645% 2nd best model is try20-opt1 with 30.444% Our best submitted was model1 with 30.242% robetta4 had 25.000% Fri Sep 24 17:39:55 PDT 2004 Kevin Karplus Looking at the GDT curves and smooth_GDT scores, it is clear that our model1 (try24-opt2) outperforms the robetta models and our other models. We picked the right one this time! name length missing_atoms rmsd rmsd_ca GDT smooth_GDT model1.ts-submitted 126 0.0000 9.7256 8.9809 -28.0242 -28.0244 robetta-model3.pdb.gz 126 0.0000 10.8583 10.5021 -23.9919 -24.2660 model2.ts-submitted 126 0.0000 11.0223 10.3426 -23.9919 -23.6344 robetta-model4.pdb.gz 126 0.0000 16.5012 16.0378 -24.1935 -23.5440 robetta-model10.pdb.gz 126 0.0000 13.3121 13.1167 -23.1855 -22.3569 robetta-model7.pdb.gz 126 0.0000 13.8078 13.6067 -21.9758 -22.1026 model5.ts-submitted 126 0.0000 12.7664 12.3428 -23.9919 -22.1020 model3.ts-submitted 126 0.0000 12.1074 11.5134 -22.3790 -22.0901 full auto model4.ts-submitted 126 0.0000 12.7534 12.3428 -22.5806 -21.9711 robetta-model2.pdb.gz 126 0.0000 12.7813 12.3835 -22.5806 -21.8718 robetta-model1.pdb.gz 126 0.0000 10.4744 10.1418 -20.9677 -20.4790 robetta-model6.pdb.gz 126 0.0000 16.1020 15.8192 -20.1613 -20.2686 robetta-model5.pdb.gz 126 0.0000 14.9270 14.6406 -21.1694 -20.1675 robetta-model9.pdb.gz 126 0.0000 18.0800 17.5784 -19.9597 -19.5864 robetta-model8.pdb.gz 126 0.0000 15.2888 14.9217 -19.9597 -19.5570