Wed Jun 28 10:17:27 PDT 2006 T0350 Make started Wed Jun 28 10:18:57 PDT 2006 Running on shaw.cse.ucsc.edu Wed Jun 28 16:50:04 PDT 2006 Kevin Karplus No good hits with BLAST in PDB. (best E-value=1 for 2bkiA=2bkhA) No good hits with HMMs (best is 1t6sA, evalue 3.7) Top few alignments pull out a hairpin or short meander or strand-helix strand----little bits of supersecondary structure. The try1-opt2 prediction has a lot of strand----too much strand, in fact, since all the strand after P86 was predicted to be helix. We might want to keep the sheet constraints for the meander, combined with the helix constraints from the neural nets to make try2.costfcn. These sheets may not be quite right. t06.o_notor2 and t06.n_notor2 predict H-bonds for H63.O--L66.N (or maybe D64.O--N67.N), but try1-opt2 has H63.O--(N67.N+E68.N) Wed Jun 28 17:13:24 PDT 2006 Kevin Karplus try2 started on farm cluster. We'll probably have to play with the sheet constraints quite a bit. Wed Jul 5 08:11:02 PDT 2006 Kevin Karplus I picked up the server models and scored them with unconstrained.costfcn. SAM_T06_server_TS1 scores best (better than try1-opt2 or try2-opt2). Next is Pmodeller6_TS4, Pmodeller6_TS2=ROBETTA_TS2, ... I'm redoing the make on T0350.do2, since gromacs doesn't run properly on the farm cluster. Fri Jul 14 15:28:29 PDT 2006 Martin Madera Random notes while looking at summary.html: - last six residues (112 onwards) are clearly a His tag - region 22-29 is: NQKDNDNE ... extremely hydrophilic! - the HMMs are pretty similar; t04 and t06 seems slightly better than t2k - according to t06/t04 str2, it's: junk - helix - strand - strand - strand - helix - ?helix and the strands look anti-parallel. Try1-opt2 made the strongly-predicted terminal helix into a strand, which is probably wrong. Same for try2. Mon Jul 17 16:48:59 PDT 2006 Martin Madera Try3: same as try1 but increased the weight for the last two helices to 10. Running on abyss. Tue Jul 18 12:15:26 PDT 2006 Martin Madera Make started Tue Jul 18 12:26:02 PDT 2006 Running on camano.cse.ucsc.edu ... that's re-running the make because of the alignment bug. Try3 worked: the strands got made into helices. But I don't like the (lack of) packing. I think I'll start two more runs from the alignments, with the same cost function as try3, to see what else undertaker can come up with. Top server models: SAM_T06_server_TS1 159.01 Pmodeller6_TS4 178.43 ROBETTA_TS2 181.31 Pmodeller6_TS3 181.94 Pmodeller6_TS1 185.44 The server models all look pretty much like try3: a sheet out of three anti-parallel strands and the helices all 'packed' (not very well!) on one side of it. With the exception of SAM-T06, all servers stick out the C-terminal helix and try to wrap the N-terminal helices around it in ways that range from unrealistic to outright ridiculous. The last two servers (Pmodeller TS3 and TS1) are the best (=least worst). Hmm. The question is what to do with the helices. While I think about it, try4: the same as try3, just seeing what else undertaker can come up with... running on peep try5: starting from try3 and the server models, see what undertaker can do to pack the helices a bit better... Tue Jul 18 15:56:51 PDT 2006 Martin Madera Try4 and try5 have finished. The arrangement of the helices is different, but no better than try3. Hmmmmpf. The ~57-~63 strand is clearly amphipathic according to both near and burial. So the guess that all the servers have been making, that one side of the sheet is exposed, sounds reasonable. Though maybe the N-terminal helices should loosely cover parts of the first strand. Sheet constraints. Kevin says that the notor alphabets think there should be an H-bond between: H63.O--L66.N Grant says: ---------------------------------------------------------------------- Sure, if you look at the sequence logos, there is a "G" at H63 for the o_notor alphabet and at L66 for the n_sep alphabet, indicating a likely hydrogen bond/hairpin there. However, you may want to look at the sequence logos for the sep alphabets, because they are predicting a three strand anti-parallel sheet. Residue E61 is predicting a separation of 9 for the o_sep and -9 for the n_sep which matches Residue Y70 for sep 9 for n_sep and -9 for o_sep. Residue E68 is predicting separation -5 for o_sep and matches to residue H63 for -5 for n_sep. You are also seeing a prediction for E80 for -5 and I82 for -9. Also V 73 is getting a +9 for the o_sep alphabet. n_sep sees the corresponding residues as well. Check out the sep alphabets as well! ---------------------------------------------------------------------- ... to which I replied: ---------------------------------------------------------------------- > Sure, if you look at the sequence logos, there is a "G" at H63 for the > o_notor alphabet and at L66 for the n_sep alphabet, indicating a likely > hydrogen bond/hairpin there. Oops, obviously! I kept focusing on the beginnings/ends of the sheets and somehow managed to confuse n_notor with o_notor in the process. OK, I get it, thanks for straightening me out. > However, you may want to look at the sequence > logos for the sep alphabets, because they are predicting a three strand > anti-parallel sheet. > > Residue E61 is predicting a separation of 9 for the o_sep and -9 for the > n_sep which matches Residue Y70 for sep 9 for n_sep and -9 for o_sep. > Residue E68 is predicting separation -5 for o_sep and matches to residue H63 > for -5 for n_sep. You are also seeing a prediction for E80 for -5 and I82 > for -9. Also V 73 is getting a +9 for the o_sep alphabet. n_sep sees the > corresponding residues as well. Thanks! I totally forgot about the sep alphabets. What do the letters mean?! Ah, I see, the library file you sent me: (/projects/compbio/lib/alphabet/hbonds.alphabet) CharName = G sep_+3 3_10_helix_or_turn CharName = H sep_+4 alpha_helix CharName = I sep_+5 sep_+5 CharName = J sep_+6 sep_+6 CharName = K sep_+7 sep_+7 CharName = L sep_+8 sep_+8 CharName = M sep_+9 sep_+9 CharName = N sep_+10 sep_+10 CharName = P sep_-3 sep_-3 CharName = Q sep_-4 sep_-4 CharName = R sep_-5 sep_-5 CharName = S sep_-6 sep_-6 CharName = T sep_-7 sep_-7 CharName = U sep_-8 sep_-8 CharName = V sep_-9 sep_-9 CharName = W sep_-10 sep_-10 Okay... so let's translate this into a sheet diagram: n_sep: E61 V = -9 H63 R = -5 E68 I? = +5 Y70 M = +9 V73 V = -9 E75 R = -5 I82 M = +9 o_sep: E61 M = +9 E68 R = -5 Y70 V = -9 V73 M = +9 E80 R = -5 I82 V = -9 ... hmmm, I take it that the n_sep numbers should be negative? So let's try that: n_sep: E61 V = +9 H63 R = +5 E68 I? = -5 Y70 M = -9 V73 V = +9 E75 R = +5 I82 M = -9 o_sep: E61 M = +9 E68 R = -5 Y70 V = -9 V73 M = +9 E80 R = -5 I82 V = -9 so I have: E61.N ... Y70.O H63.N ... E68.O E68.N ... ??? Y70.N ... E61.O V73.N ... V82.O E75.N ... E80.O I82.N ... V73.O or, in other words: E75.N ... E80.O V73.N ... V82.O V73.O ... I82.N E61.N ... Y70.O E61.O ... Y70.N H63.N ... E68.O ... which seems to agree with what you said. OK, excellent! Thanks a lot, Martin ---------------------------------------------------------------------- Adding Kevin's hairpin, I get: E75.N ... E80.O V73.N ... V82.O V73.O ... I82.N E61.N ... Y70.O E61.O ... Y70.N H63.N ... E68.O H63.O ... L66.N or... # display -> backbone # colours -> monochrome select Glu61.N, Tyr70.O color red spacefill on select Glu61.O, Tyr70.N color green spacefill on select His63.N, Glu68.O color blue spacefill on select His63.O, Leu66.N # pink for hairpin! color pink spacefill on select Val73.O, Ile82.N color green spacefill on select Val73.N, Ile82.O color blue spacefill on select Glu75.N, Glu80.O color red spacefill on ... aaand: try3 & try5 are basically spot on except Glu75.N--Glu80.O is broken, and try4 is *exactly* spot on. Try1 messed up the 73-82 turn and try2 is completely messed up. So I want the sheet constraints from try4, and try1 & try2 are useless. Now the helices. Tue Jul 18 19:37:43 PDT 2006 Martin Madera ---------------------------------------------------------------------- For T0350, I managed to sort out the sheet (or rather, try4 happens to completely agree with the n_sep/o_sep alphabets), but the helices are still a complete mess. I think a few more runs to tidy them up -- make sure the most hydrophobic of them actually touches the sheet -- could improve things a lot. Looking at the server models, their packing of the helices is completely hopeless, so we could actually do well on this target. I'll give you what I have now, but any chance of a properly documented resubmission in the morning? ---------------------------------------------------------------------- Kevin wants to submit, so here are the best models: - Try5-opt2: best scoring. Used our try3 and server models, but topologically is based on try3, which it managed to improve on by breaking up the final helix, bending it over and burying the end. The helices look implausible but score well. - Try4-opt2: this has the best sheet accoring to our o_sep and n_sep alphabets, but scores poorly, probably because the helices are a mess. - Try3-opt2: the first model that has roughly the correct sheets and helices. - Try1-opt2: third of the sheet is right, third of it is wrong, and third of it shouldn't be a sheet at all. - Try2-opt2: scores poorly and the sheet is all wrong, but who knows. Tue Jul 18 20:24:03 PDT 2006 Martin Madera Edited try4-opt2 in ProteinShop to move the helices around a bit, and saved the result in decoys/T0350.edit4.pdb.gz. Try6: same as try5 but based on edit4 and using sheet constraints from try4 ... running on peep Try7: same as try6; want to see what else undertaker can come up with ... running on lopez Tue Jul 18 21:31:48 PDT 2006 Kevin Karplus Submission done, with comment We are still trying to do fold-recognition on T0350, but may be assembling things from bits and pieces of supersecondary structure. Model 1 is try5-opt2: best scoring. Used our try3 and server models, but topologically is based on try3, which it managed to improve on by breaking up the final helix, bending it over and burying the end. The helices look implausible but score well. Model 2 is try4-opt2: this has the best sheet according to our o_sep and n_sep alphabets, but scores poorly, probably because the helices are a mess. Model 3 is try3-opt2.gromacs0.repack-nonPC: the first model that has roughly the correct sheets and helices, reoptimized by gromacs and with sidechains selected by rosetta on a fixed backbone. It is Rosetta's favorite of our backbones to repack. Model 4 is try1-opt2, the fully automatic model: third of the sheet seems right, third of it is wrong, and third of it shouldn't be a sheet at all. Model 5 is try2-opt2, which scores poorly and the sheet is all wrong, but who knows. ------------------------------------------------------------ Tue Jul 18 23:29:33 PDT 2006 Martin Madera Try6 and try7 have both improved on try5, the previous best. And looking at them I got an idea for how to improve the model further: twist the sheet (which looks a bit flat at the moment) and increase the angle between the two helices. The helices also need rotating, because the burial isn't optimal. Wed Jul 19 00:45:15 PDT 2006 Martin Madera Adjusting sheets in ProteinShop is a good way to get RSI. I can't see a way of selecting an entire sheet and making it twist / curl. And trying to adjust individual strands while keeping the sheet together is *incredibly* frustrating. Wed Jul 19 02:42:36 PDT 2006 Martin Madera Further edited try6 in PhotoShop, rotating the helices so that the buried residues are actually buried. Saved as: decoys/T0350.edit13.pdb.gz Try8 & try9: same as try6 and try7, but on edit13 rather than edit4. Running on lopez and peep. Wed Jul 19 07:45:53 PDT 2006 Martin Madera Try8 is a slight improvement on try6,7, try9 does worse. Both look funny, but then all the helices for this target tend to look funny. The helix I rotated to improve burial got rotated back in both. I like try6 the best out of the new models, and don't like try8 at all (the His tag plays too important a role). So I suggest: ------------------------------------------------------------ We are still trying to do fold-recognition on T0350, but may be assembling things from bits and pieces of supersecondary structure. Model 1 is try6-opt2: our second best-scoring model and the one where the helix packing seems the most sensible. It is based on a hand-edited version of try4-op2, improved by undertaker. Model 2 is try5-opt2. It used our try3 and server models, but topologically is based on try3, which it managed to improve on by breaking up the final helix, bending it over and burying the end. The helices look implausible but score well. Model 3 is try4-opt2: this was the first model with the correct sheet structure according to our o_sep and n_sep alphabets, but scores poorly, probably because the helices are a mess. Model 4 is try3-opt2.gromacs0.repack-nonPC: the first model that has roughly the correct sheets and helices, reoptimized by gromacs and with sidechains selected by rosetta on a fixed backbone. It is one of Rosetta's favorite of our backbones to repack. Model 5 is try1-opt2, the fully automatic model: third of the sheet seems right, third of it is wrong, and third of it shouldn't be a sheet at all. ------------------------------------------------------------ Wed Jul 19 09:28:46 PDT 2006 Kevin Karplus I resubmitted as recommended by Martin (editing the T0350.method file to reflect the changes). I wonder whether we still want try3, now that Rosetta prefers try7-opt2.gromacs0.repack-nonPC How similar is try7 to try6? to try3? Wed Jul 19 09:42:48 PDT 2006 Kevin Karplus I resubmitted yet again, with the following comment: We ended up using fold recognition, assembly of models from fragments, and hand-editing with ProteinShop for this target. Model 1 is try6-opt2: our second best-scoring model and the one where the helix packing seems the most sensible. It is based on a hand-edited version of try4-op2, improved by undertaker. Model 2 is try7-opt2.gromacs0.repack-nonPC: a similar model to our model 1, but with the undertaker-produced model reoptimized by gromacs and with sidechains selected by rosetta on a fixed backbone. It is Rosetta's favorite of our backbones to repack. It also looks most feasible to dimerize along the edge of the sheet. (L51-H63) We did not have time to attempt optimizing in a dimeric context. Model 3 is try5-opt2. It used our try3 and server models, but topologically is based on try3, which it managed to improve on by breaking up the final helix, bending it over and burying the end. The helices look implausible but score well. Model 4 is try4-opt2: this was the first model with the correct sheet structure according to our o_sep and n_sep alphabets, but scores poorly, probably because the helices are a mess. Model 5 is try1-opt2, the fully automatic model: third of the sheet seems right, third of it is wrong, and third of it shouldn't be a sheet at all. ------------------------------------------------------------