On Fri, Nov 30, 2012 at 2:03 AM, vincent Chaptal
I am using the "reflection file editor" in phenix 1.8.1-1168 to generate my test set (with the GUI). I have NCS so I ticked the box "assign test set in thin resolution shells". I also have twinning, and want to have my Rfree not biased by twin related reflections. I guess the option "use lattice symmetry to generate test set" is the one I should use for this purpose? I couldn't find much information online about this option. No log file has been generated so I can't check what happened.
Yes, you want to use lattice symmetry - this is the default in most programs in Phenix, since there's no reason not to use it. All it does is group together reflections which might be related by twin operators or pseudo-symmetry. I will see if our documentation can be expanded (and log output, sorry!). According to Randy and Paul, the thin shell method doesn't really help very much in practice. Pavel claims it can screw with the maximum likelihood equations which assume a more-or-less even distribution of the test set, although I haven't noticed this myself - it is probably more of an issue for poorer data. (A secondary problem in Phenix is that we don't have a sensible method to extend the test set to higher resolution when it's created in shells.)
I am supposing xtriage is ran in the background and if a twin law is found, then used to generate the test set.
Nope, the relationships between reflections are dictated by the lattice, and are independent of twinning (or the specific twin law) - if you don't have twinning or pseudosymmetry, then of course these reflections are not actually correlated, but it doesn't hurt to keep them together anyway (especially since you might later collect high-resolution data from a crystal that *is* twinned). In other words, if you have a lattice like this: P21 10 20 30 90 90.1 90 then the Phenix will generate the flags in P222, and expand to P21.
I would like to be sure my Rfree is generated correctly. Alternatively, what would happen when multiple twin laws are available? Is the most probale used? Is it possible to specify the editor which law to use?
It looks like you're doing the right thing. Just keep in mind that a) the problem of NCS restraints biasing your test set is not going to completely disappear with the use of thin shells, and b) as Garib Murshudov likes to point out, twinning and twinned refinement affect the calculation of R-free in other ways, so you need to exercise a little more caution than usual interpreting that statistic. -Nat