First simulated annealing using torsion angle dynamics is used to improve the model. The use of torsion angle dynamics reduces the number of parameters being refined and hence reduces the degree of overfitting of the data. The experimental phase information (the Hendrickson-Lattman coefficients from MAD phasing) is included in the refinement by the use of the MLHL target. This information is also used in subsequent map calculations to reduce model bias. For an initial model with relatively large errors (due to manual building or misplaced atoms) a starting temperature of 5000K is recommended. In order to decrease the computational time required the cooling rate can be increased from 25K to 50K. The simulated annealing refinement task file includes energy minimization both before and after the simulated annealing. Multiple refinement trials can be performed, each with different initial velocities for the molecular dynamics. It can be useful to run 5 or 10 trials if the there are some serious errors in the model - greater variation is usually seen in these areas. The structure factors from the multiple models can also be averaged to reduce the noise in the electron density maps (see the CNS task file optimize_average.inp). The simulated annealing refinement is performed with the CNS task file anneal.inp:
cns_solve < anneal.inp > anneal.out [94 minutes]
The result of the simulated annealing refinement is a new coordinate file (anneal_1.pdb). In the majority of CNS refinement task files information about the refinement procedure is written out at the top of the output coordinate file (as REMARK statements):
REMARK coordinates from simulated annealing refinement REMARK refinement resolution: 500 - 1.9 A REMARK starting r= 0.3689 free_r= 0.3752 REMARK final r= 0.3286 free_r= 0.3598 REMARK rmsd bonds= 0.012463 rmsd angles= 1.41808 REMARK wa_initial= 2.49661 wa_dynamics= 2.30846 wa_final= 2.38108 REMARK target= mlhl md-method= torsion annealing schedule= slowcool REMARK starting temperature= 5000 total md steps= 100 * 6 REMARK sg= P6 a= 115.996 b= 115.996 c= 44.130 alpha= 90 beta= 90 gamma= 120 REMARK parameter file 1 : CNS_TOPPAR:protein_rep.param REMARK molecular structure file: nsf.mtf REMARK input coordinates: nsf.pdb REMARK reflection file= nsf.cv REMARK reflection file= mad_phase2.hkl REMARK ncs= none REMARK B-correction resolution: 6.0 - 1.9 REMARK initial B-factor correction applied to fobs : REMARK B11= 7.156 B22= 7.156 B33= -14.312 REMARK B12= 2.812 B13= 0.000 B23= 0.000 REMARK B-factor correction applied to coordinate array B: -0.239 REMARK bulk solvent: density level= 0.389055 e/A^3, B-factor= 37.9744 A^2 REMARK reflections with |Fobs|/sigma_F < 0.0 rejected REMARK reflections with |Fobs| > 10000 * rms(Fobs) rejected REMARK anomalous diffraction data was input REMARK theoretical total number of refl. in resol. range: 52358 ( 100.0 % ) REMARK number of unobserved reflections (no entry or |F|=0): 4913 ( 9.4 % ) REMARK number of reflections rejected: 0 ( 0.0 % ) REMARK total number of reflections used: 47445 ( 90.6 % ) REMARK number of reflections in working set: 42732 ( 81.6 % ) REMARK number of reflections in test set: 4713 ( 9.0 % ) CRYST1 115.996 115.996 44.130 90.00 90.00 120.00 P 6 REMARK FILENAME="/tmp_mnt/Net/franklin/u0/pvm/tmp/nsf_ref/anneal_1.pdb" REMARK DATE:14-Jun-99 18:00:38 created by user: paul REMARK VERSION:0.5 ATOM 1 CB PHE 19 -4.710 35.834 61.046 1.00 31.63 ATOM 2 CG PHE 19 -6.024 35.216 60.646 1.00 31.63
This information provides a summary of the refinement and also a record of the input data and parameters used to generate this structure.
The initial model has the same B-factor for all atoms. In order to take account of the possible variation in B-factors throughout the model, group B-factors are then refined. Two B-factors are refined for each residue, one for the main-chain atoms and one for the side-chain atoms. The small number of parameters being optimized allows this to be performed even when the resolution of the data is low (even below 3A). As before, experimental phase information is included (and will be included in all subsequent refinement steps). The group B-factor refinement is performed with the CNS task file bgroup.inp:
cns_solve < bgroup.inp > bgroup.out [4 minutes]
This first round of refinement has reduced the free R-value from 37.5% to 32.5%. The next step is to locate any major part of the model which is currently missing (see the next section).