Section 21.1.1. Introduction

Owing to the limited resolution and imperfect phase information that macromolecular crystallographers usually have to deal with, building and refining a protein model based on crystallographic data is not an exact science. Rather, it is a subjective process, governed by experience, prejudices, expectations and local practices (Brändén & Jones, 1990; Kleywegt & Jones, 1995b, 1997). This means that errors in this process are almost unavoidable, but it is the crystallographer's task to remove as many of these as possible prior to analysis, publication and deposition of the structure. With high-resolution data and good phases, the resulting model is probably more than 95% a consequence of the data, although even at atomic resolution, subjective choices must still be made: which refinement program to use, whether to include alternative conformations, whether to model explicit H atoms, how to model temperature factors, which restraints and constraints to apply, which peaks in the maps to interpret as solvent molecules and how to treat noncrystallographic symmetry (NCS). Once the resolution becomes worse than ∼2 Å, this balance shifts and some published protein models appear to have been determined more by some crystallographer's imagination than by any experimental data.

Subjectivity is not necessarily a problem, provided that the crystallographer is experienced, knows what he or she is doing and is aware of the limitations that the experimental data impose on the model. However, even inexperienced people can avoid many of the pitfalls of model building and refinement. Supervisors have a major responsibility in this respect: education is an important factor (Dodson et al., 1996). Students who have built and refined a previously determined structure from scratch as a training exercise will have met most of the problems that can be encountered in real life (Jones & Kjeldgaard, 1997). Apart from hands-on experience, there are many other methods to reduce or avoid errors. These include (1) the use of information derived from databases of well refined structures in model building (Kleywegt & Jones, 1998) [e.g. to generate main-chain coordinates from a C^α trace (Jones & Thirup, 1986) and side-chain coordinates from preferred rotamer conformations (Ponder & Richards, 1987)]; (2) the use of various sorts of local quality checks (to detect residues that for one or more reasons are deemed `unusual' and that require further scrutiny and perhaps adjustment; Kleywegt & Jones, 1996a, 1997); and (3) the use of global quality indicators [e.g. the use of the free R value (Brünger, 1992a, 1993) to signal major errors, to prevent over-fitting, and to monitor the progress of the rebuilding and refinement process (Kleywegt & Jones, 1995b; Kleywegt & Brünger, 1996; Brünger, 1997)].