Sat Aug 2 13:35:11 PDT 2008 TR462 Sat Aug 2 13:36:02 PDT 2008 Kevin Karplus CASP organizer say REFINEMENT TARGET TR462 (the best submitted model according to GDT_TS). GDT_TS=67; RMSD=2.16A. Target contains 2 domains (approx. 1-75 and 77-143). Each of these domains can be refined separately (if desired). Please, submit the refined domains in one file. The refinement assessor will be recommended to assess the whole structure and every domain separately. C-term trimmed in the model from res.144 to the end. Make started Sat Aug 2 13:38:03 PDT 2008 Running on moai11.kilokluster.ucsc.edu Sat Aug 2 16:57:58 PDT 2008 Kevin Karplus The provided model does not correspond to any of the server models. The closest server model is Zhang-Server_TS3 (GDT 77%) Sat Aug 2 18:22:16 PDT 2008 Kevin Karplus My closest model is T0462/try7-opt3 (GDT=68.2%), which is not one I submitted, next closest were MQAU1, MQAU3, try6 (my model 2, at GDT=65.9%), MQAC1 (my model5 at 63.6%) I've started subdomain predictions for M1-G74 and E77-V143. I should also take my subdomain models from before and see how well they superimpose---the CASP organizers may have picked the model with the best packing of the two domains rather than the best individual domains. Mon Aug 4 10:39:42 PDT 2008 Kevin Karplus I started a try2 optimization from just the provided model, which has bad breaks and clashes. Mon Aug 4 12:10:46 PDT 2008 Kevin Karplus Even the try2 model has breaks and clashes---bad enough that gromacs improves the model even with the try2 costfcn. Some of the sheets in the C-terminal domain are getting lost, so I may need to turn up the constraints. Mon Aug 4 12:18:58 PDT 2008 Kevin Karplus I added the subdomains in T0462 to the search in find-TR.under, to see if I previously generated any models similar to the provided one for either domain. Thu Aug 14 14:13:48 PDT 2008 Kevin Karplus Both the prior subdomains have huge rmsd_ca values from the provided model, which is claimed to be within 2.16A, so I'm not going to get any help from using them. The Pcons_dot_net_TS3 model is close (in rmsd_ca) to the first domain, as are several other first-domain only server models. Thu Aug 14 14:34:35 PDT 2008 Kevin Karplus I've started doing optimizations for both subdomains, with moderate constraints to try to keep the sheets and helices intact while reducing clashes and breaks. Thu Aug 14 15:37:45 PDT 2008 Kevin Karplus The try2 runs finished for both subdomains, optimizing the provided model. The M1-G75 domain did not change much, but some bad clashes were removed. The E77-V143 domain did not change much either, but I think I might be better off grabbing some of the sheet constraints from the try1 model, to get cleaner hbonds. I've already started the try3 runs, so the extra constraints will be put in the E77-V143/try4 run, which I'll start even before try3 finishes. Thu Aug 14 17:40:10 PDT 2008 Kevin Karplus E77-V143/try4 does a better job of getting the strands, but messes up the helix. I'll make a try4-try3 chimera that takes the helix (G95-G107) from try3 and the rest from try4. Thu Aug 14 17:50:25 PDT 2008 Kevin Karplus E77-V143/try5 started to try to close gaps and polish chimera-try4-try3 I don't think that there is as much to do for the first domain. Perhaps I could do a polishing run starting from the gromacs optimized models. Or I could try to extract constraints from the initial model and use them to find a model from alignments---I doubt that that would improve the initial model much. Thu Aug 14 20:54:47 PDT 2008 Kevin Karplus E77-V143/try5 worked fairly well, so I polished it a bit more for try6. I doubt that I'll get much better than E77-V143/try6-opt3. Hmm, gromacs increases the breaks substantially---perhaps it is seeing a clash that undertaker misses? Conformation[40] TR462.try6-opt3.gromacs0.pdb.gz has 2 breaks TR462.try6-opt3.gromacs0.pdb.gz breaks before (TR462)I128 with cost 1.51364 TR462.try6-opt3.gromacs0.pdb.gz breaks before (TR462)E127 with cost 0.635892 Looking in the E77-V143/ TR462.try6-opt3.repack-nonPC file, I see high energy for I90, I108, A83, R142, E127, I128, Y121, L106, L130. E127 has high E_dun (so bad rotamer). I128 has high E_rep (clash) as do I90, I108, and A83. I90 and I208 bump on the CG1, CG2, but underaker sees no problems with E127 and I128. Eyeballing the before and after (try6-opt3 and try6-opt3.gromacs0.repack-nonPC) does not show an obvious difference in the backbone. The biggest motion seems to be for E127.CA, so I suspect that the problem is the N-CA-C bond angle there. I'll do another optimization from just the gromacs-optimized models, with n_ca_c costs turned up. Thu Aug 14 21:23:25 PDT 2008 Kevin Karplus E77-V143/try7 started to clean up the gromacs-optimized models. I'm seeing the same thing happening in M1-G75/try4, with gromacs opening up breaks. Conformation[26] TR462.try4-opt3.gromacs0.pdb.gz has 4 breaks TR462.try4-opt3.gromacs0.pdb.gz breaks before (TR462)S20 with cost 1.20104 TR462.try4-opt3.gromacs0.pdb.gz breaks before (TR462)Q42 with cost 0.71484 TR462.try4-opt3.gromacs0.pdb.gz breaks before (TR462)L17 with cost 0.60168 TR462.try4-opt3.gromacs0.pdb.gz breaks before (TR462)V19 with cost 0.597376 Conformation[28] TR462.try4-opt3.pdb.gz has 1 breaks TR462.try4-opt3.pdb.gz breaks before (TR462)S20 with cost 0.585555 Thu Aug 14 21:39:01 PDT 2008 Kevin Karplus M1-G75/try5 started to optimize from the gromacs-optimized models with more weight on n_ca_c. Thu Aug 14 21:57:46 PDT 2008 Kevin Karplus The differences between E77/V143/try5, try6, and try7 are mainly different rotamers for surface sidechains. I'll superimpose try7-opt3 on the full-length model. Fri Aug 15 05:44:58 PDT 2008 Kevin Karplus Working now on the full chain, try3-opt3 looks good, but has changed the relationship between the two domains, which is probably incorrect. I'll have to add constraints to keep the two parts in the probably correct orientation. From the provided model: picking residues in 1-70 and 80-999 (to avoid along-chain closeness) close interface (CB within 8.0 of an atom of the other domain) M1.CB L82.CB 5.30 M1.CB A83.CB 9.42 M1.CB D84.CB 12.57 K2.CB I81.CB 9.90 K2.CB L82.CB 8.39 K2.CB D84.CB 12.26 R5.CB I81.CB 5.93 R5.CB L82.CB 6.79 R5.CB A83.CB 11.04 V7.CB I81.CB 9.24 V7.CB L82.CB 11.74 V7.CB A83.CB 16.37 V7.CB D84.CB 17.96 CB within 12.0 of a CB of the other domain: P45.CB Y132.CB 9.05 P45.CB R119.CB 12.38 S4.CB R119.CB 13.86 S4.CB A135.CB 16.78 V7.CB A135.CB 11.26 V7.CB T136.CB 8.98 V10.CB T136.CB 13.45 V10.CB W139.CB 11.07 A12.CB W139.CB 16.36 A12.CB I81.CB 14.06 R5.CB I81.CB 5.93 R5.CB L82.CB 6.79 M1.CB L82.CB 5.30 M1.CB A83.CB 9.42 K2.CB A83.CB 11.74 K2.CB D84.CB 12.26 L6.CB D84.CB 16.16 L6.CB L80.CB 11.82 L3.CB L80.CB 16.64 L3.CB R142.CB 19.15 M1.CB R142.CB 11.91 M1.CB K137.CB 16.97 V7.CB K137.CB 11.18 V7.CB Y132.CB 11.54 distant pairs N56.CB R126.CB 50.46 R57.CB R126.CB 48.57 R57.CB E127.CB 46.47 E22.CB E127.CB 46.48 E22.CB N105.CB 51.13 E25.CB N105.CB 53.13 E25.CB K104.CB 52.36 V28.CB K104.CB 50.66 V28.CB S110.CB 46.58 N56.CB S110.CB 43.05 Fri Aug 15 06:36:46 PDT 2008 Kevin Karplus For try4, I'll again start with just chimera-N5-C7, but with constraints on the close CBs ("CB within 12.0 of a CB of the other domain") and the most distant CBs ("distant pairs"). This should preserve/restore the orientation of the domains while the linker is reshaped to close the gaps. Fri Aug 15 08:05:47 PDT 2008 Kevin Karplus try4-opt3 does a reasonable job of re-connecting the two domains while holding the orientation of the provided model. I'll polish (from all models) for try5, hoping to get a model that rosetta will like as much as the try3 model, in which the domains have drifted a bit. Fri Aug 15 09:06:39 PDT 2008 Kevin Karplus try5-opt3 looks pretty good, but rosetta does not like it as much---gromacs seems to be opening up some breaks. For try6, I'll do an optimization starting from the gromacs-optimized models, with breaks and clashes turne up and the domain-relationship constraints turned down. Since these start out scoring a bit worse than try4 and try3, I don't expect try6 to beat try5, but it might be more compatible with the rosetta energy function. Fri Aug 15 10:23:55 PDT 2008 Kevin Karplus try6-opt3 doesn't come quite to try5-opt3, but it is close. Rosetta likes the try6-opt3.gromacs0.repack-nonPC best of the models that have the same domain interaction as the initial model. (It still likes try3-opt3 better, which has a different relationship between domains.) I will try an optimization run starting with just try6-opt3.gromacs0.repack-nonPC as try7, trying to close the small remaining breaks. Fri Aug 15 12:17:57 PDT 2008 Kevin Karplus Neither rosetta nor undertaker likes try7 as much as try6. (well, rosetta likes try7-opt3.repack-nonPC better than try6-opt3.repack-nonPC, but neither as well as try6-opt3.gromacs0.repack-nonPC) Fri Aug 15 12:27:40 PDT 2008 Kevin Karplus I think I may be reaching diminishing returns. I'll do one more polishing run trying to improve Hbonds and pack tighter, starting from all existing models (try5-opt3 and try6-opt3 score almost the same with the try8 costfcn). Fri Aug 15 13:39:59 PDT 2008 Kevin Karplus try8-opt3 scores better than try5-opt3 or try6-opt3, but rosetta doesn't care much for it. I think there is not much more I can do to this without a lot of work---work that I don't think will make enough of a difference to matter. I'll submit ReadConformPDB TR462.try6-opt3.gromacs0.repack-nonPC.pdb # < try5-opt3.gromacs0 < try4-opt3 < chimera-N5-C7 ReadConformPDB TR462.try8-opt3.pdb # < try6-opt3 < try5-opt3.gromacs0 < try4-opt3 < chimera-N5-C7 ReadConformPDB TR462.try5-opt3.pdb # < try4-opt3 < chimera-N5-C7 ReadConformPDB TR462.try4-opt3.pdb # < chimera-N5-C7 ReadConformPDB TR462.try2-opt3.gromacs0.pdb # < TR462