Choosing Weights

The determination of proper weights is essentially an empirical process and therefore requires some experimentation. In other words, you will have to guess.

Basically, the importance of each kind of restraint you place on your model can be individually set using a weight. The larger the weight the more important that restraint becomes. Since all the weights are relative increasing one is the same as dropping all the others.

I recommend that you decide your weights based on trial and error. Choose weights which give you the quality of answer you want. When your TNT control file was created by from_pdb it contained the weights

WEIGHT RFACTOR  0.0010
WEIGHT BOND     0.8  ANGLE 1.3  TORSION 0   PSEUDOROTATION 1
WEIGHT TRIGONAL 2    PLANE 5    CONTACT 10  BCORREL 1

You will need to fine-tune these to match your situation. After a trial run of refinement compare the r.m.s. errors for the geometry restraints of the model to the target values in Table .

Table: Target r.m.s. Value for each Restraint Type

If the r.m.s. error for a class of restraints is too large I increase its weight. If the model is too consistent with the observations its weight must be reduced.

Usually the only problem will be the RFACTOR weight. Since there is no objective target values for the diffraction data term, we can only set the RFACTOR weight by monitoring the stereochemical restraints' r.m.s. errors. If all these restraints are too tight the RFACTOR weight must be increased. If they all are too loose the RFACTOR weight must be dropped.

Usually I change the RFACTOR weight by factors of two to ten. For example, if the bond length r.m.s. were 0.08 instead of 0.02 (and all the other geometric restraints were proportionately large) I would decrease the RFACTOR weight by a factor of ten. If the bond length r.m.s. were instead 0.01 I would double the RFACTOR weight.

When the model is in bad shape it is a good idea to adapt the model to the diffraction data first and then bring the model into agreement with ideal geometry. This is done by performing a series of cycles with a high RFACTOR weight and another series with lower weight. The final weight is chosen so that the geometry statistics at the end are just what you want.

If the starting model is good, say the R value is less than 25%, then there is no reason to follow this complicated procedure. In this situation you may simply refine with the RFACTOR weight set to give proper geometry. Usually the same weight for the RFACTOR term can be used each time you return from model building.

The B correlation (BCORREL) module restrains the B factors of the model to a library of standard restraints. These restraints were developed by an analysis of the relationship between the B factors of atoms bonded to each other in models refined to 1.6Å diffraction data. If your data extends to 1.6Å it makes no sense to apply these restraints and you should set the BCORREL weight to zero. If your diffraction data is not of this quality you should set the weight to one. With very low resolution data you may need to increase the weight to force the model to reach the target for this restraint.