Wed Jun 14 09:11:07 PDT 2006 T0330 Make started Wed Jun 14 09:11:50 PDT 2006 Running on lopez.cse.ucsc.edu Wed Jun 14 09:13:20 PDT 2006 Kevin Karplus BLAST gets a moderate hit on 2ah5A (23% over 211 residues, 5.5e-04) Wed Jun 14 10:39:36 PDT 2006 Kevin Karplus 2ah5A is getting excellent scores with the HMMs also. Wed Jun 14 15:07:14 PDT 2006 Kevin Karplus 2ah5A is our top hit, but many c.108.1.* templates score well. There seems to be general agreement on the outer domain, but the insertion L15-L94 seems to vary more. This may just be hinging motion between the two domains, or it may be higher variability in the inserted region. Wed Jun 28 18:32:52 PDT 2006 Martin Madera This target is very similar to T0324, which I'm also doing. See the T0324 README first, because the problem seems to be the same. In a nutshell, try1-opt2 is again based on 1x42, which is wrong. The top BLAST hits have a conserved crevice which undertaker doesn't like, so it uses a more distant template (1x42) which closes it somewhat, and undertaker further modifies it to fill it up as much as possible. All the machines are busy at the moment, so I'll wait with try2 till later on tonight. Wed Jun 28 21:25:46 PDT 2006 Martin Madera Hmmm, but looking at the HMM alignments using: less */T0330-*-t2k-local-adpstyle1.a2m the following look the best: >1aq6A -MIKAVVFDAYGTLFDVQSVADATERAYpgRGEYITQVWRQKQLEYSWLR ALMGRYADFWGVTREALAYTLGTLGLEPDESFLADMAQAYNRLTPYPDAA QCLAELAP---LKRAILSNGAPDMLQALVANAGLTDSFDAVISVDAKRVF >1jud mdy--IKGIAFDLYGTLFDVHSVVGRCDEAFpgRGREISALWRQKQLEYT WLRSLMNRYVNFQQATEDALRFTCRHLGLDLDARTRSTLCDAYLRLAPFS EVPDSLRELKR-RGLKLAILSNGSPQSIDAVVSHAGLRDGFDHLLSVDPV >1lvhA -MFKAVLFDLDGVITDTAEYHFRAWKALAEEIGINGVDRQFNEQLKGVSR EDSLQKILDLADKKVSAEEFKELAKRKNDNYVKMIQDVSPADVYPGILQL LKDLRS-NKIKIALASAS--KNGPFLLERMNLTGYFDAIADPAEVAASKP >1o08A -MFKAVLFDLDGVITDTAEYHFRAWKALAEEIGINGVDRQFNEQLKGVSR EDSLQKILDLADKKVSAEEFKELAKRKNDNYVKMIQDVSPADVYPGILQL LKDLRS-NKIKIALASAS--KNGPFLLERMNLTGYFDAIADPAEVAASKP >1qq5A -MIKAVVFDAYGTLFDVQSVADATERAYpgRGEYITQVWRQKQLEYSWLR ALMGRYADFWSVTREALAYTLGTLGLEPDESFLADMAQAYNRLTPYPDAA QCLAELAP---LKRAILSNGAPDMLQALVANAGLTDSFDAVISVDAKRVF >1qq7A -MIKAVVFDAYGTLFDVQSVADATERAYpgRGEYITQVWRQKQLEYSWLR ALMGRYADFWSVTREALAYTLGTLGLEPDESFLADMAQAYNRLTPYPDAA QCLAELAP---LKRAILSNGAPDMLQALVANAGLTDSFDAVISVDAKRVF >1zrn mdy--IKGIAFDLYGTLFDVHSVVGRCDEAFpgRGREISALWRQKQLEYT WLRSLMNRYVNFQQATEDALRFTCRHLGLDLDARTRSTLCDAYLRLAPFS EVPDSLRELKR-RGLKLAILSNGSPQSIDAVVSHAGLRDGFDHLLSVDPV ... the insertion is 16-92. This is a different set from what I used for T0324, and moreover doesn't agree with the BLAST hits (where 2ah5A and 2fdrA are the top two). Interestingly, 1lvhA and 1o08A -- the two best matches, no insertions or deletions in our region -- have one type of structure, let's call it -A-, the rest have a different structure, let's call it -B-. Importantly, -A- is basically the same as the structure of 2ah5A and 2fdrA (the top BLAST hits) -- there's an insertion of 7 residues wrt 2ah5A which extends two adjacent helices and moves the loop between them, but the rest of the structure is virtually identical. (The superpositions are in T0324/align.) All other structures have that short 'pg' insertion, and -B- is basically the same as 1x42 of T0324 fame. Most of them are dimers, and those that aren't look like they should be. Hmm... but this is irrelevant, really, as I want to edit the .under files, which contain the following structures one or more times: 1jud 19 -B- 1l7mA 7 strange structure, another insertion in the main domain 1rdfA 9 strange structure 1rqlA 77 -A- 1swvA 96 -A- 1te2A 443 -A- 1x42A 655 -B- 1zrn 6 -B- 2ah5A 311 -A- 2fdrA 232 -A- 2fi1A 8 -A- 2gfhA 524 -B- 2go7A 88 -A- ... the numbers indicate positions in the T06, T04 and T2k alignments. Basically -A- predominates, but lengthwise -B- is also plausible and makes a reasonable showing. Sigh. I think I'll do two runs, one based on the -A- structures, one based on the -B- structures, and we'll see how they work out. Try2 = -A- = 1rqlA|1swvA|1te2A|2ah5A|2fdrA|2fi1A|2go7A Try3 = -B- = 1jud|1x42A|1zrn|2gfhA Sun Jul 2 12:11:52 PDT 2006 Kevin Karplus For some reason, try3 did not do the repacking (perhaps try2 deleted the repack.res file that try3 needed), so I remade try3, which just made the missing try3-opt2.repack-nonPC model and rescored. I also scored the existing models with unconstrained.costfcn. It is looking like try2 is being more successful than try3 in finding models. Sun Jul 2 12:16:53 PDT 2006 Kevin Karplus Part of the problem may be that try3 was given very few alignments to try. It might be a good idea to put all the interesting chains in MANUAL_TOP_HITS in the Makefile (before the include) then do make extra_alignments make read_alignments and use the InfilePrefix 1xxxX/ include read-alignments-scwrl.under in the try*.under file. Sun Jul 2 20:00:04 PDT 2006 Martin Madera I agree that try3 is a disaster. Try2 looks ok; it's roughly the structure I want, but compared to what T0324 looked like, the insertion isn't very well packed. I will run the insertion by itself. I'm not sure it's going to work (it's short, so the iterative procedure may not work), but it's worth a try. If the HMM works, it would stop undertaker from trying to pull the insertion towards the main domain. So: 1) edit Makefile: first 15, last 93 => length 79 2) make subdomain ... worked! L15-L93 3) (make -k >& subdomain.log; gzip -9f subdomain.log) & Running on orcas. I can see a blastpgp process, so it's doing something sensible. Sun Jul 2 21:51:05 PDT 2006 Martin Madera Problem #1: it got stuck asking me for a password to apache. So I typed it in, but then Problem #2: WARNING: This version of the Acrobat Distiller may not be fully compatible with previous versions of Acrobat Exchange and Acrobat Reader due to new functionality based on recent PDF language additions. To ensure compatibility with previous versions of Acrobat Exchange and Acrobat Reader, the DEFAULT for compatibility can be set inside your personal preference file. Setting this compatibility switch, however, will result in the disabling of new features of the Acrobat Distiller. The compatibility switch may easily be overridden at run-time through the specification of -compatlevel and in no way represents an actual permanent loss of features for Acrobat Distiller. NOTE: Running the Acrobat Distiller with the -noprefs option will effectively disable the preference chosen below. Enter Acrobat Distiller personal preferences file modification [1, 2, 3] [1] use 3.0 new features [2] use 2.1 compatibility [3] leave compatibility undefined Invalid response. Please reenter response [1|2|3] Invalid response. Please reenter response [1|2|3] Invalid response. Please reenter response [1|2|3] Invalid response. Please reenter response [1|2|3] Invalid response. Please reenter response [1|2|3] Invalid response. Please reenter response [1|2|3] Invalid response. Please reenter response [1|2|3] Invalid response. Please reenter... OK. So need to set up ssh... done. And figure out what's going on with Acrobat Distiller. Ssh'd to apache, ran distill T0330.t06.w0.5-logo.eps chose [3]... and it worked. Then ran distill again, and this time it didn't ask. GOOD. Problem solved. Mon Jul 3 01:24:38 PDT 2006 Martin Madera The subdomain run finished. There were errors, so I ran it again, and there were still errors. No try1 PDB structures got generated, which is a pity. However, looking at the HMM hits, it would have been a waste of time anyway. The iterative procedures didn't really work, and to my surprise the template libraries didn't pick it up either! Well, was worth a try. -- Try4: following Kevin's suggestion on MANUAL_TOP_HITS, added MANUAL_TOP_HITS:= 1rqlA 1swvA 1te2A 2ah5A 2fdrA 2fi1A 2go7A 1jud 1x42A 1zrn 2gfhA to the Makefile, ran make extra_alignments make read_alignments and started try4 (copied from try3.under) with the following extra lines in try4.under: InfilePrefix 1jud/ include read-alignments-scwrl.under InfilePrefix 1x42A/ include read-alignments-scwrl.under InfilePrefix 1zrn/ include read-alignments-scwrl.under InfilePrefix 2gfhA/ include read-alignments-scwrl.under (just after the lines taken from all-align.a2m). Also changed the scoring function to: dry5 30 dry6.5 30 dry8 20 dry12 5 (instead of: dry5 15 dry6.5 20 dry8 15 dry12 5 ) to make it slightly more compact. Running on orcas. Tue Jul 4 00:19:58 PDT 2006 Martin Madera Try4 has done the trick for the -B- domains. I'll do two polishing runs based on try2 and try4 and call it a day. Try5: polishing try2-opt2. Break, dry5 dry6.5 x2. Running on lopez. Try6: polishing try4-opt2. Break, dry5 dry6.5 as for try5. Running on pyro. Tue Jul 4 08:49:40 PDT 2006 Kevin Karplus To do a real "polishing" run, you need to turn up clashes as well as breaks, and turn constraints way down or off. remaining clashes are still a bit too large. Please try to have something ready by 8pm, as I'd like to get some sleep tonight. Tue Jul 4 15:26:10 PDT 2006 Martin Madera Maybe "polishing" was the wrong word for try5/try6, I meant "improve upon". I wanted to get a more compact structure first, before doing the final polishing runs (along the lines of what Kevin did yesterday for T0329). Try5/try6 aren't much better than try2/try4. Hmmm. Out of T0324, T0329 and T0330, this (T0330) is the most difficult one: target best BLAST E-value T0329 4.1e-10 T0324 1.8e-08 T0330 5.5e-04 so it's not surprising that it looks the worst so far. I don't see an obvious way of improving try6. For try5 I have an idea: try and push G50 closer towards R83 in an attempt to close the cavity. So try7: start from the gromacs versions of try5 and try2, with Constraint G50.CA R83.CA 0.0 7.0 9.0 5 DryXX and break back to standard, constraints up x3 to 30. Running on peep. Now two real polishing runs (based on Kevin's try13/try14 from T0329), so that we have something decent to submit for the soft deadline: try8: polishing try2,5-gromacs ... running on squawk try9: polishing try4,6-gromacs ... running on pyro. Tue Jul 4 16:54:01 PDT 2006 Martin Madera Noticed errors in try7.log: had ReadConformPDB TT0330 ... note the double T! Fixed, restarted. Tue Jul 4 17:15:49 PDT 2006 Martin Madera Polishing is all very well, but the fact is that so far the structures for the inserted domains aren't very good. Try7 is an attempt to do something about this, but we need more. Any ideas? I am tempted to do something with constraints, to keep an open crevice, like what I did for try12 in T0329. Except that was a disaster, so I really need to go back and understand what went wrong. Tue Jul 4 18:54:27 PDT 2006 Martin Madera Try7 and try8 have finished. I can't really see any difference between try8 and try5, but I guess the devil's in the detail. Try7: the constraint was: Constraint G50.CA R83.CA 0.0 7.0 9.0 5 and the actual distance is... 7.0A! And, wow, look at the packing compared to try5!!! My only complaint is that I should have pulled it in even closer. According to the unconstrained cost function, try7 beats try5 by quite a margin. I'll start a polishing run based on try7 (the gromacs version), but it's likely to take a few hours. I will also do a repeat of try7, trying to pull it in even close. I've updated best-models based on what we've got so far. Tue Jul 4 20:06:49 PDT 2006 Martin Madera Try9 finished. According to the unconstrained cost function it's a minor improvement on try6, but Rosetta seems to like it. Updated best-models. Tue Jul 4 20:21:28 PDT 2006 Martin Madera While looking at best-models, in particular the model from alignments, I got the following idea for a constraint: Constraint A51.CA D72.CA 0.0 6.3 8.3 5 Try10: like try7, but with the above constraint rather than the try7 one, and reading in try2,5,7-opt2.gromacs. Running on peep. Try11: polishing try7-opt2.gromacs. Running on lopez. Try12: polishing try7-opt2. Running on squawk. Tue Jul 4 23:04:52 PDT 2006 Kevin Karplus try10 and try11 have finished. try12 is still running---perhaps squawk is a slightly slower machine? With unconstrained costfcn, best models are try11=opt2, try12-opt1, try7-opt2, try10-opt2 (I think that try12-opt2 will be best when try12 finishes.) try11-opt2 is best with the try11 costfcn (which has no constraints), followed by try8-opt2, try10-opt2, try12-opt1, try7-opt2 try10.costfcn (with a hefty constraints) orders them try10-opt2, try8-opt2, try5-opt2, try2-opt2 try11, try12, and try7 are all *very* similar, but try10 is a bit different---it looks more like try5. Tue Jul 4 23:37:50 PDT 2006 Kevin Karplus preliminary submission with comment: Model 1 is try11-opt2, our best-scoring model at the moment. Model 2 is try12-opt1, a very similar model. Model 3 is try10-opt2, optimized from try5-opt2.gromacs0, a slightly different helical domain. Model 4 is T0330.try9-opt2.gromacs0.repack-nonPC, an alternative helical domain, and the model that rosetta likes best after repacking. Model 5 is sidechain replacement by SCWRL on a backbone from alignment to 2ah5A. Fri Jul 7 14:37:59 PDT 2006 Martin Madera A quick summary of the models to remind myself where things stand on this target. try1: -B- structure; I thought it was a failure at the time but it's actually OK try3: attempt at -B-, blew up try4: -B- structure try6: re-running try4 with higher breaks & dry try9: polishing try2,6 (but only using try6) try2: -A- structure try5: re-running try2 with higher breaks & dry try8: polishing try2,5 try7: from the Gromacs versions of try2,5, a constraint that closed the cavity try10: a different constraint, trying to improve Gromacs versions of try2,5,7 try11: polishing Gromacs version of try7 try12: polishing try7 Try10 looks worse than try7, and try7 (and later models based on it) look very good. I don't think there's much else we can do on this target, so I'll call it a day. I agree with Kevin's choice of models for the preliminary submission. Fri Jul 7 16:58:21 PDT 2006 Martin Madera A good idea gleaned from T0303: try polishing the server models. Looking at decoys/score-all+servers.unconstrained.pretty, the top models are: SAM_T06_server_TS1 Pmodeller6_TS1 ROBETTA_TS1 : BayesHH_TS1-scwrl HHpred3_TS1-scwrl Zhang-Server_TS1 3Dpro didn't do so well. So I'll base it on Pmodeller6 & Robetta, copying T0303/try3.under. This will be try13. Running on peep. Fri Jul 7 20:25:19 PDT 2006 Martin Madera Try13 finished. According to the unconstrainted cost fuction it doesn't score very well, so I don't think we should include it in the submission. Fri Jul 7 23:43:30 PDT 2006 Kevin Karplus I see nothing in here to indicate the desire for a new final submission. I'll let the preliminary submission ride, unless Martin sends me email.