This is a very important stage because it may result in a significantly improved model. It mostly involves visual inspection of the superposed structures on a graphics terminal, using a program such as RASMOL (Roger Sayle, Glaxo, http://www.umass.edu/microbio/rasmol/), QUANTA or INSIGHTII. However, the CHECK_ALIGNMENT command of MODELLER should also be used. You can obtain files with multiply superposed structures by using the MALIGN3D command with WRITE_FIT set to on (you can use any alignment for this purpose). The aim at this stage is to study the family fold, to establish the relationships between various members of the family, and to determine which regions are more conserved and which regions are more variable. This information is used to improve the initial automatically derived alignment. For example, if necessary, gaps are removed from helices and strands; they should be moved into those exposed regions that show large variations in the family of known structures and to the tips of loops. The role of disulfides and cis-prolines, if any, is noted. MODELLER will try to deal with those two features automatically but it is prudent to be careful. It has to be decided whether or not to build models for multi-subunit assemblies and whether or not to include various ligands, such as water molecules, cofactors, metal ions, inhibitors, or substrates. QUANTA and INSIGHTII have a set of tools that facilitate inspection and editing of multiple alignments.
In order to obtain the best possible model, it is very important to understand how the alignment is used by MODELLER[Šali & Blundell, 1993]. In outline, for the aligned regions, MODELLER tries to derive a 3D model for the target sequence that is as close to one or the other of the template structures as possible while also satisfying stereochemical restraints (e.g., bond lengths, angles, non-bonded atom contacts, ...); the inserted regions, which do not have any equivalent segments in any of the templates, are modeled in the context of the whole molecule, but using their sequence alone. This way of deriving a model means that whenever a user aligns a target residue with a template residue, he tells MODELLER to treat the aligned residues as structurally equivalent.