When constructing the initial model for loop refinement, MODELLER fills in any missing atoms in both the loop and non-loop regions. These are needed in order to accurately determine the interaction between the loop and the rest of the protein. The coordinates for the missing atoms are constructed automatically using internal coordinates, so clashes between the atoms and the rest of the protein may exist (note, however, that the score of the loop does not include protein-protein internal interactions, so will not be affected).
If set to False, the default, the written-out models will contain the atoms present in the original non-loop region plus those added automatically. If set to True, the models will contain only the non-loop atoms that were present in the input model.
Note that if loopmodel.loop.write_selection_only is set to True, only the loop region is written out, so this option has no effect.