Section 21.1.4. Fixing errors

The object of model rebuilding is generally twofold: (1) to make the model as complete and detailed as the data will allow one to do confidently (e.g. to add previously unmodelled loops, ligands, water molecules etc.) and (2) to remove errors. At first glance, it may not seem all that important to fix each and every side chain and to correct all peptide O atoms that are pointing in the wrong direction, but one should keep in mind that an error in the scattering factor (atom type or charge), position or B factor of even a single atom will be detrimental to the entire model. Particularly in the early stages of model rebuilding and refinement, one often finds that after an extensive round of rebuilding followed by more refinement, the density improves dramatically and new features become clear. One should also keep in mind that incorrect features of a model may be very persistent and become `self-fulfilling prophecies', a phenomenon known as `model bias' (Ramachandran & Srinivasan, 1961; Read, 1986, 1994, 1997; Hodel et al., 1992). This is particularly relevant in cases where unbiased phase information (e.g. SIRAS, MIR or MAD phases, or phases obtained after NCS or multiple-crystal averaging) is not available.

For error detection to be effective, it is best not to approach the rebuilding process in a haphazard way (Kleywegt & Jones, 1997). O users can employ a program called OOPS (Kleywegt & Jones, 1996a) to carry out this task in a systematic yet convenient fashion. This program uses information calculated by O (e.g. pep-flip and real-space fit values) and retrieves or derives other information from a PDB file of the current model (e.g. temperature factors, Ramachandran plot, changes with respect to the previous model). Moreover, results from a coordinate-based quality check by the WHAT IF program (Vriend, 1990; Hooft et al., 1996) can be included. In all, several dozen quality indicators can be used and plots and statistics for many of these can be produced by the program. The program's most useful feature, however, is that it will generate O macros that when executed in O will take the crystallographer on a journey to all the residues that may require attention because they are outliers for one or more quality criteria. This makes the rebuilding process often faster and certainly more efficient and focused than a residue-by-residue walk through the model. In addition, it teaches inexperienced crystallographers to recognize and diagnose common model errors.

If a residue is an outlier for a certain criterion, the crystallographer has to inspect the local density and the structural context and decide the course of action. If the residue is in a region of the model in which many residues are outliers for many criteria, there may be something seriously wrong locally (for instance, there could be a register error), possibly because the density is poor. If there is poor density for several residues in a row, the crystallographer might consider leaving these residues out of the model for the next refinement round or cutting off the side chains at the C^β atoms. Sometimes local errors are correlated, such as a pep-flip error in one residue and a Ramachandran violation for its C-terminal neighbour, or a residue with a non-rotamer conformation and high temperature factors in conjunction with a poor real-space fit. O contains many tools to manipulate individual residues and atoms (Jones et al., 1991; Jones & Kjeldgaard, 1994, 1997; Kleywegt & Jones, 1997), e.g. to flip a peptide plane, to replace a side chain by a rotamer conformation, to change side-chain torsion angles in order to optimize the fit to the density, to move groups of atoms, to use real-space refinement on a single residue or a zone of residues, to `mutate' a residue to alanine etc. Together they constitute a toolbox with which many problems, once recognized, can be fixed relatively effortlessly (Kleywegt & Jones, 1997).