A valid concern particularly in this case, and related to a question someone asked off-list. The model came from a collaborator when it was at early stages of refinement (Rwork/Rfree 0.244/0.296). I followed up with steps I knew would additionally confound cross-validation:
1. I assigned a new free flag in thin shells to allow for the effect of NCS on free set independence (Kleywegt & Jones, 1996) <http://dx.doi.org/10.1107/S0907444995014983>. 2. I selected one of the NCS mates from coordinates I received and used PHASER to position 6 copies of it in the asymmetric unit. Afterward, I annealed the model (PHENIX.REFINE) to “remove the memory” (Brunger, 1993) <http://dx.doi.org/10.1107/S0907444992007352> of free reflections, to the extent possible given inherent interdependencies of between reflections and NCS. Further to several suggestions about the space group, I am refining the structure in R3 (hexagonal setting). XDS’s CORRECT.LP “SYMMETRY OF REFLECTION INTENSITIES” shows Rmeas(146) as 21% v Rmeas(155) 53%. The fact that I have experimented with both SCALEPACK and XDS intensities throughout my refinement should not matter in this context as the MTZ files have consistent indices and the same FREE flag. Additional testing has shown that mere omission of TLS parameterization will give Rwork/Rfree of 0.209/0.215 (versus 0.207/0.206 with 10 cycles TLS refinement). For that comparison, I had to adjust REFMAC’s WEIGHT MATRIX to achieve similar bond/angle RMSDs. The latter is relevant here since, as another off-list respondent suggested, Rfree positively correlates with those RMSDs. I further suspect (but have not tested yet) that omission of explicit NCS restraints would also cause Rfree to rise. I tend toward following Gerard Bricogne’s intuition about sample variance of Rfree. It is mentioned briefly by Brunger (1993). Would this be the variance corresponding to Cruickshank’s expected value <http://dx.doi.org/10.1107/S0907444995010638>? On Tue, Jun 30, 2015 at 2:07 PM, Robbie Joosten <robbie_joos...@hotmail.com> wrote: > Hi Wolfram, > > > > You didn’t tell us where your model came from but 10 cycles of TLS and 10 > cycles of restrained refinement is not enough for a refinement to converge > if you just picked your test set. Try resetting your B-factor and doing > 30-40 cycles refinement in REFMAC. > > > > Cheers, > > Robbie > > > > *From:* CCP4 bulletin board [mailto:CCP4BB@JISCMAIL.AC.UK] *On Behalf Of * > wtempel > *Sent:* Tuesday, June 30, 2015 18:59 > *To:* CCP4BB@JISCMAIL.AC.UK > *Subject:* [ccp4bb] Rfree below Rwork > > > > Hello, > > my question concerns refinement of a structure with 6-fold NCS (local > automatic restraints in REFMAC) against 2.8 A data. The size of my free set > is 1172 selected in thin resolution shells (SFTOOLS) and corresponding to > 4.3 % of reflections. > > A refmac run of 10 cycles of TLS and 10 cycles of CGMAT starts out at > Rfree/Rcryst 0.271/0.272. After the 10th TLS cycle I have 0.227/0.224. Yes, > Rfree < Rcryst. At the end of CGMAT I have 0.2072/0.2071. > > I understand that NCS stresses the independence assumption of the free > set. Am I correct in believing that Rfree *may* be smaller than Rcryst even > in the absence of a major mistake? My hope is that the combined wisdom of > ccp4bb followers can point out my possible mistake, suggest tests that I > may perform to avoid them and, possibly, arguments in defense of a > crystallographic model with Rfree < Rcryst. > > Many thanks, > > Wolfram Tempel >
