Wed Jun 25 09:11:46 PDT 2008 T0472 Make started Wed Jun 25 09:12:37 PDT 2008 Running on peep.cse.ucsc.edu Wed Jun 25 09:43:35 PDT 2008 Kevin Karplus This looks like an easy fold-recognition for 3bidA. Sun Jun 29 16:02:41 PDT 2008 Kevin Karplus The MQAC MQA function likes RAPTOR_TS1, fais-server_TS4, SAM-T08-server_TS2/1, Zhang-Server_TS2, ... The MQAU MQA function likes SAM-T08-server_TS1, Zhang-Server_TS2, pipe_int_TS1, RAPTOR_TS1, ... The predicted GDTs are a bit lower than I expected (68.5% for MQAU and 53.4% for MQAC). I guess I'd better do the metaserver runs and see what comes out. (If the SAM-T08-server models get selected, I might want to do another metaserver run excluding them.) Sun Jun 29 16:08:54 PDT 2008 Kevin Karplus I also need to do a polishing run, to close up breaks. Sun Jun 29 21:29:19 PDT 2008 Kevin Karplus Both the MQAU1 and the MQAC1 runs are based on SAM-T08-server_TS1. try1 likes MQAU1-opt3 try2-opt3 MQAC1-opt3 try1-opt3.gromacs0 try2 likes try2-opt3, MQAU1-opt3.gromacs0 MQAc1-opt3.gromacs0 try1-opt3.gromacs0 I'll do an MQAX run to use all server models EXCEPT SAM-T08-server. Mon Jun 30 10:02:04 PDT 2008 Kevin Karplus The MQAX1 run failed with undertaker: Spot.cc:172: XYZpoint Spot::get_spot(const ChainsResiduesAndAtoms*, int, const Pointlist&, const Conformation*, bool, const Spot*) const: Assertion `return_point.mag2() < 1.e12' failed. after just doing the first round of OptConform, which eventually focussed on Pcons_dot_net_TS1 The models are all in agreement up to about I52, but the second half is packed in 2 different ways against the first half. Tue Jul 1 15:57:28 PDT 2008 SAM-T08-MQAO hand QA T0472 Submitted Tue Jul 1 15:57:28 PDT 2008 SAM-T08-MQAU hand QA T0472 Submitted Tue Jul 1 15:57:28 PDT 2008 SAM-T08-MQAC hand QA T0472 Submitted Sat Jul 5 09:51:54 PDT 2008 Kevin Karplus I'm going to retry the MQAX1 run, and hope that either previous bug fixes or random chance will get it past the crash this time, as I don't feel like debuggng undertaker today. Sat Jul 5 10:59:30 PDT 2008 Kevin Karplus MQAX1 is based on BAKER-ROBETTA_TS4 Sat Jul 5 14:25:24 PDT 2008 Kevin Karplus I now have 3 distinct lineages: try2 < try1 < alignments MQAX1 < BAKER-ROBETTA_TS4 MQAU1,MQAC1 < SAM-T08-server_TS1 There are two copies of a domain consisting of a 3-strand sheet and a helix. Perhaps I should try predicting the two domains separately, splitting around P51. There are fairly clear correspondeces of conserved residues with a spacing of 51: Y5-Y56, F17-F68, L19-L70, I27-I78, Y33-Y84, I43-I94, V46-V97 I'll do M1-R55 and P51-T110. Sat Jul 5 14:43:28 PDT 2008 Kevin Karplus Makes started in both subdomains. Note: the 3bidA template has a domain consisting of 3 strands a helix and a 4th strand that is domain swapped with a dimeric pair. Furthermore, these dimers are in turn dimerized AB to GH to make a continuous sheet for 3bidB-3bidH, so we should try dimerizing T0472. I should make a chimeric sequence 3bidAB and try aligning to it. It may also be that this target is domain swapped so that the target corresponds to HA or BG in 3bid. That would allow a shorter linker than 3bidAB. The linker is a problem, since 3bid has a 57-residue unit, but the repeat distance in the target is only 51. Sat Jul 5 21:32:17 PDT 2008 Kevin Karplus I think that one of the edge strands is missing from the second sheet, but there is a final strand that comes back to the frist sheet, making a 4-strand sheet and a 3-strand sheet. I made a crude chimera-N1-C1 from the try1 models of the two subdomains, and I'll superimpose that model on 3bidA and 3bidB, to make a better chimera. Sat Jul 5 22:03:05 PDT 2008 Kevin Karplus chimera-N1-C1-3bid is a superposition of the try1-opt3 models from M1-R55 (up to S50) and P51-T110 on 3bidA and 3bidB respectively. try3.costfcn is an attempt to make the last strand parallel to the first one. I may be off by 2 on the phase. I'll do two optimization runs: one from alignments and one from chimera-N1-C1-3bid. I'll include the alignments from the subdomains. Sat Jul 5 22:12:14 PDT 2008 Kevin Karplus try3 started to optimize from alignments. Sat Jul 5 22:15:39 PDT 2008 Kevin Karplus try4 started on peep to optimize chimera-N1-C1-3bid. If the try3 and try4 runs don't seem to move in the right direction, I may need to try again without the nn1000, align_constraints, and two-stage constraints, and turn up ReduceConstraint. I'm hoping that the cost function alone will be a strong enough push, but it is far from guaranteed. Sun Jul 6 06:51:26 PDT 2008 Kevin Karplus try3-opt3 < align(3bidA+1wfbA+1r17A+2bz6L+1hzfA) does not manage to for the last strand, though the other constraints seem to be reasonably well met. The try3 model has considerable trouble with the second domain, and may not be worth further attention. The try4-opt3 model attempted to form the last strand, but without success. Cleaning up the clashes and breaks might help (and adding the knot costfcn to make sure the joining is done on the correct side). The MQAX1 model, though quite different, is also worth further attention. Sun Jul 6 07:50:47 PDT 2008 Kevin Karplus try5 started on cheep to try to clean up try4. Sun Jul 6 07:58:06 PDT 2008 Kevin Karplus try6 started on moai cluster to try to improve MQAX1 Sun Jul 6 08:01:42 PDT 2008 Kevin Karplus MQAX5 stated on peep to see what the servers have that come close to the try4/try5 goal. Sun Jul 6 10:13:50 PDT 2008 Kevin Karplus For try7, I'll try moving the last strand by 2 (relative to try4 and try5) to see if that makes it easier to place. I'll run try7, like try4, from the chimera-N1-C1-3bid, and I'll also do an MQAX7 run from all servers. I'll probably have to do another polishing run afterwards to make a good model. Sun Jul 6 13:22:48 PDT 2008 Kevin Karplus MQAX7-opt3 from RAPTOR_TS1 scores best with the try7 costfcn. The try7-opt3 model makes it looks like I should back the last strand up 2 more, to get SheetConstraint G3 K9 T104 T110 hbond L7 MQAX7 is having similar problems closing the gap before the last strand. I'll try try8 and MQAX8 with this constraint. try8 will start with all models except the metaserver models (and the try6 models, which were from MQAX1). Sun Jul 6 13:37:46 PDT 2008 Kevin Karplus MQAX8 started on cheep, try8 started on peep. Sun Jul 6 16:39:16 PDT 2008 Kevin Karplus MQAX8 is based on SAM-T08-server_TS1 try8 is based on try7-opt3 (without substantially closing the gaps). Both MQAX8-opt3 and try8-opt3 come close to doing what I wanted---definitely better than previous models. But they need to have gaps closed. Sun Jul 6 16:53:14 PDT 2008 Kevin Karplus try9 started on cheep, polishing from all models (probably will use MQAX8 or try8). Sun Jul 6 18:43:22 PDT 2008 Kevin Karplus try9 looks very good to me. I'll do a polishing run for it, then see if I'm ready to submit. Sun Jul 6 19:36:37 PDT 2008 Kevin Karplus try10-opt3 is about as good as I'm going to get things. So I guess it is time to submit ReadConformPDB T0472.try10-opt3.pdb # < try9-opt3 < MQAX8-opt3 < SAM-T08-server_TS1 # a very different approach for final strand to same domain ReadConformPDB T0472.try6-opt3.pdb # < MQAX1-opt3 < BAKER-ROBETTA_TS4 # failed try at final strand (off by 2?) ReadConformPDB T0472.MQAX7-opt3.pdb # < RAPTOR_TS1 # Failed try at final strand (off by 4?) ReadConformPDB T0472.try5-opt3.pdb # < try4-opt3 < chimera-N1-C1-3bid ReadConformPDB T0472.MQAU1-opt3.gromacs0.pdb # < SAM-T08-server_TS1 Sun Jul 6 19:52:59 PDT 2008 Kevin Karplus Submitted with comments: Although T0472 has fairly easy homology to 3bidA, it was not a trivial target to model. The 3bid structure has strand-swapped dimers, and T0472 has two copies of the 3bid monomer. It is not long enough, however, to contain a full strand-swapped dimer from 3bid and the ends of the monomers there are not close enough to make a good single chain. My main line of reasoning was that the C-terminal strand would still exist and swap to the N-terminal domain, but that the last strand of the first domain would simply be missing, allowing a fairly easy connection between domains. I achieved this structure mainly by adding constraints to the undertaker cost function, after superimposing models of the two domains on 3bidA and 3bidB, to get the basic shape to measure constraints from. Model 1 T0472.try10-opt3.pdb # < try9-opt3 < MQAX8-opt3 < SAM-T08-server_TS1 This is the model I like best, with essentially no breaks or clashes, but still a compact model with the final strand where I wanted it. 2 T0472.try6-opt3.pdb # < MQAX1-opt3 < BAKER-ROBETTA_TS4 This metaserver model represents a very different approach for handling the final strand---attaching it to the C-terminal domain. It is more compatible with the secondary-structure prediction, which predicts a helix for E53-A57, but I prefer a structure that has greater similarity between the tandem repeats. 3 T0472.MQAX7-opt3.pdb # < RAPTOR_TS1 I was not certain of the phase of the last strand. This is an earlier attempt that undertaker did not manage to close. I think the final strand is off by 2. 4 T0472.try5-opt3.pdb # < try4-opt3 < chimera-N1-C1-3bid This is a still earlier attempt to get the last strand in place, built on top of a model obtained by superimposing separately optimized domains on the 3bidAB pair. I think that I was trying to place the final strand off by 4 in phase. 5 T0472.MQAU1-opt3.gromacs0.pdb # < SAM-T08-server_TS1 This is an early metaserver prediction, before I had decided to force the C-terminal strand to swap to the first domain. It is based on the same server prediction as model1, but does not have the benefit of the constraints that held the configuration to look like 3bid. It is representative of the models generated by automatic mathods, where human intution was not applied. ---------------------------------------------------------------------- Mon Jul 7 04:09:00 PDT 2008 Kevin Karplus Although I've submitted a model I'm fairly pleased with, I'm wondering whether I should optimize it in dimeric contexts, since 3bid has not only the AB pairing but also the AB:GH pairing that joins the sheets of the monomers. If I were to dimerize as in the 3bid model, I'd want the D108 residues to pair and to have a 3-residue antiparallel strand K107-L109 to L109-K107. To get a decent superposition, I should probably align the N-terminal domains on 3bidB and 3bidG, Mon Jul 7 04:50:21 PDT 2008 Kevin Karplus Oops. that should be 3bidA and 3bidG. Mon Jul 7 06:58:27 PDT 2008 Kevin Karplus OK, made a dimer and dimer/try11-opt3 does seem to improve things a bit (making a salt bridge between the halves of the dimer). Perhaps I should also make a dimer using the other domain, though that will be harder, since the edge strand is missing. Mon Jul 7 12:23:12 PDT 2008 Kevin Karplus The constraint to apply on that edge is most likely SheetConstraint Y56 V60 V170 Y166 Hbond K58 I'll do a dimer/try12 run to improve that edge, then one more run on the other edge (with the try11 costfcn or similar). Mon Jul 7 17:13:07 PDT 2008 Kevin Karplus I just noticed that the RAPTOR_TS1 model is actually quite good, for all except the first strand of the second domain. It has the same modeling of the C-terminal strand that I ended up with. Mon Jul 7 17:38:54 PDT 2008 Kevin Karplus The first two rounds of optimization in dimer/try13 may have been wasted, as the final round started over from dimer-try122-from-try11. Wed Jul 9 18:41:31 PDT 2008 Kevin Karplus THEY CANCELLED T0472, because it was released as 2k49A yesterday. I looked to see how well we did. We had a model that was better than any server model, but didn't submit it (try8-opt3.repack-nonPC, GDT=67.3% real_cost=-103.78) model1 was our best submitted model at 66.1% GDT and real_cost -91.57, but pipe_int_TS1 had 68.4% GDT and real_cost 102.74. pipe_int_TS1 was the only server model to beat our model1--it was one of the top 6 models in SAM-T08-MQAC, so the MQA seems to be working.n Somewhat surprisingly, running scwrl on model1 improved it quite a bit, to real_cost -99.06. Top servers: pipe_int_TS1 (model1) FAMSD_TS1 (model2) RAPTOR_TS1 (model4) pro-sp3-TASSER_TS1 PS2-server_TS3