Hi Francis, Detwinning of perfectly twinned data is done with reference to Fcalc (determined from from the model): it is not possible to arithmetically deconvolute the contribution of the twin domains to the perfectly twinned reflection data.
Because your model is incomplete you will be introducing (or subject to) significant bias by refining against the resulting incompletely/inaccurately detwinned data. The solution is to use minimize_twin.inp/anneal_twin.inp (CNS) or phenix.refine twinning.twin_law="xxx" (PHENIX) to perform the refinement in the "true" space group on the original data, taking into account the twin operator. The map produced by model_map_twin.inp (and presumably phenix.refine) will still be based on data detwinned with reference to the model, but at least this occurs post-refinement. One way to limit a molecular replacement search to one twin domain is to perform the rotation search in the "apparent" space group and the translation search in the "true" space group. If the result is more than one molecule in the asymmetric unit, NCS is certainly useful because perfectly twinned data halves the number of independent reflections (and hence the data-to-parameter ratio). Test set (rfree) reflections should also be chosen taking into account the twin operator, for example with make_cv_twin.inp in CNS. HTH, James
