The SAM-T08 hand predictions use methods similar to SAM_T06 in CASP7. We start with a fully automated method (implemented as the SAM-T08-server): Use the SAM-T2K, SAM-T04, and SAM-T06 methods for finding homologs of the target and aligning them. Make local structure predictions using neural nets and the multiple alignments. These neural nets have been newly trained for CASP8 with an improved training protocol. The neural nets for the 3 different multiple sequence alignments are independently trained, so combining them should offer improved performance. We currently use 15 local-structure alphabets: STR2 an extended version of DSSP that splits the beta strands into multiple classes (parallel/antiparallel/mixed, edge/center) STR4 an attempt at an alphabet like STR2, but not requiring DSSP. This alphabet may be trying to make some irrelevant distinctions as well. ALPHA an discretization of the alpha torsion angle: CA(i-i), CA(i), CA(i+1), CA(i+2) BYS a discretization of Ramachandran plots, due to Bystroff PB de Brevern's protein blocks N_NOTOR N_NOTOR2 O_NOTOR O_NOTOR2 alphabets based on the torsion angle of backbone hydrogen bonds N_SEP O_SEP alphabets based on the separation of donor and acceptor for backbone hydrogen bonds CB_burial_14_7 a 7-state discretization of the number of C_beta atoms in a 14 Angstrom radius sphere around the C_beta. near-backbone-11 an 11-state discretization of the number of residues (represented by near-backbone points) in a 9.65 Angstrom radius sphere around the sidechain proxy spot for the residue. DSSP_EHL2 CASP's collapse of the DSSP alphabet DSSP_EHL2 is not predicted directly by a neural net, but is computed as a weighted average of the other backbone alphabet predictions. We make 2-track HMMs with each alphabet with the amino-acid track having a weight of 1 and the local structure track having a weight of 0.1 (for backbone alphabets) or 0.3 (for burial alphabets). We use these HMMs to score a template library of about 14000 (t06), 16000 (t04), or 18000 (t2k) templates. The template libraries are expanded weekly, but old template HMMs are not rebuilt. The target HMMs are used to score consensus sequences for the templates, to get a cheap approximation of profile-profile scoring, which does not yet work in the SAM package. We also used single-track HMMs to score not just the template library, but a non-redundant copy of the entire PDB. This scoring is done with real sequences, not consensus sequences. All the target HMMs use a new calibration method the provides more accurate E-values than before, and can be used even with local-structure alphabets that used to give us trouble (such as protein blocks). One-track HMMs built from the template library multiple alignments were used to score the target sequence. Later this summer, we hope to be able to use multi-track template HMMs, but we have not had time to calibrate such models while keeping the code compatible with the old libraries, so the template libraries currently use old calibrations, with somewhat optimistic E-values. All the logs of e-values were combined in a weighted average (with rather arbitrary weights, since we still have not taken the time to optimize them), and the best templates ranked. Alignments of the target to the top templates were made using several different alignment settings on the SAM alignment software. Generate fragments (short 9-residue alignments for each position) using SAM's "fragfinder" program and the 3-track HMM which tested best for alignment. Residue-residue contact predictions are made using mutual information, pairwise contact potentials, joint entropy, and other signals combined by a neural net. Two different neural net methods were used, and the results submitted separately. CB-CB constraints were extracted from the alignments and a combinatorial optimization done to choose a most-believable subset. Then the "undertaker" program (named because it originally optimized burial) is used to try to combine the alignments and the fragments into a consistent 3D model. No single alignment or parent template was used as a frozen core, though in many cases one had much more influence than the others. The alignment scores were not used by undertaker, but were used only to pick the set of alignments and fragments that undertaker would see. The cost functions used by undertaker rely heavily on the alignment constraints, on helix and strand constraints generated from the secondary-structure predictions, and on the neural-net predictions of local properties that undertaker can measure. The residue-residue contact predictions are also given to undertaker, but have less weight. There are also a number of built-in cost functions (breaks, clashes, burial, ...) that are included in the cost function. The automatic script runs the undertaker-optimized model through gromacs (to fix small clashes and breaks) and repacks the sidechains using Rosetta, but these post-undertaker optimizations are not included in the server predictions. They can be used in subsequent re-optimization. After the automatic prediction is done, we examine it by hand and try to fix any flaws that we see. This generally involves rerunning undertaker with new cost functions, increasing the weights for features we want to see and decreasing the weights where we think the optimization has gone overboard. Sometimes we will add new templates or remove ones that we think are misleading the optimization process. We often do "polishing" runs, where all the current models are read in and optimization with undertaker's genetic algorithm is done with high crossover. Some improvements in undertaker include better communication with SCWRL for initial model building form alignments (now using the standard protocol that identical residues have fixed rotamers, rather than being reoptimized by SCWRL), more cost functions based on the neural net predictions, multiple constraint sets (for easier weighting of the importance of different constraints), and some new conformation-change operators (Backrub and BigBackrub). We also created model-quality-assessment methods for CASP8, which we are applying to the server predictions. We do two optimizations from the top 10 models with two of the MQA methods, and consider these models as possible alternatives to our natively-generated models. For this REFINEMENT model, we did a standard prediction for the residues included in the model, then tried optimizing both from the predicted model and from the provided starting point. Since we were directed to focus on Y85-G92, I also tried cutting and pasting that region from other models (particularly server models). This provided slightly different starting points and a bit more flexibility in rebuilding that loop. My tools are not optimized for fine-grain placement of atoms, so I'm not sure that I can make any improvement over a 1.34 Angstrom CA_RMSD model. All the models submitted have a disulfide constraint: C79-C101. The CYS residues seemed too close not to be interacting, though with this being part of a cytosolic human protein a disulfide somehow seems wrong. Since the cys are on the edge of the model, so I wonder if a metal-binding site in the protein has somehow been broken by taking out just this fragment. A disulfide here could correct for the missing metal. Model 1 TR432.try7-opt3.pdb # < chimera-try6-try5 # best undertaker score (with try3,try4,try7.costfcn) chimera-try6-try5: mostly from TR432.try6-opt3.gromacs0.repack-nonPC.pdb I111-R130 is from try5-opt3 chimera-try6-try5 was an attempt to use the tighter packing of the C-terminal helix in try5 with the more standard loop in try6. 2 TR432.try7-opt3.gromacs0.repack-nonPC.pdb # < chimera-try6-try5 # best rosetta energy Rosetta and undertaker disagree somewhat on clashes, so running an undertaker-produced model through gromacs energy minimization to relieve tiny clashes, then repacking the sidechains (except PRO and CYS) with Rosetta produces the best Rosetta energy scores, even though few atoms move much. 3 TR432.try6-opt3.pdb # < chimera-try4-MULTICOM chimera-try4-MULTICOM: mostly TR432.try4-opt3.repack-nonPC.pdb A82-L95 from MULTICOM-REFINE_TS2 This chimera was an attempt to get a slightly different loop as a starting point for further optimization. Sometimes doing this sort of cut-and-paste introduces breaks into the backbone and has slightly different shape to the fragments, kicking the optimization out of a local optimum and allowing it to find a different one. 4 TR432.try5-opt3.pdb # < try3-opt3 < try2-opt3.gromacs0.repack-nonPC < chimera-init-try1 chimera-init-try1: mostly TR432.pdb Y85-R97 from try1-opt3 try1-opt3 was the automatically generated prediction for TR432 It had a rather different idea of what the interesting loop should be. Initially I favored this model, but it looks like the loop now blocks access to the putative active site: Y43,N81,Y85,N86, so I downgraded the model. The try5 run had some extra constraints to try to pull the C-terminal helix in tighter, which were fairly successful, so I copied that portion of the model into chimera-try6-try5 to make the final model. 5 TR432.try4-opt3.repack-nonPC.pdb # < TR432 # optimized directly from provided model, using just undertaker. # helix constraints taken from the initial model and from try1-opt3 # disulfide C79-C101