[ccp4bb]: Rfree and R value differences in maximum likelihood refinement

[Andrew Torelli]

Professor Dodson (and the ccp4 community),       I apologize for dragging up old messages, but I'm interested in a response you gave to a previous post and I needed time to research it myself.  On September 11th, in the "question about refinement" thread, you commented that Rfree and R values of 24% and 20.5%, respectively for a 2.4angstrom structure are "unreasonably close".       I understand that there are theoretical values for the Rfree/Rfactor ratio that are expected for refined structures (at a given resolution) using a least-squares target (thanks to Jack Tanner's July 5th post with references from Ian Tickle's group).  However, for maximum likelihood (which I assume was being used to refine the structure in question), I would expect the difference between the Rfree and Rfactor to be the same or smaller compared to a least-squares refinement.  I reference:     Adams P.D., Pannu N.S., Read R.J. & Brunger A.T. (1997) Cross-validated maximum likelihood enhances crystallographic simulated annealing refinement. Proc Natl Acad Sci U S A, 94, 5018.       Assuming a properly chosen test set (i.e. sufficient size, NCS is accounted for properly, etc.), what considerations lead you to expect (and desire)a much larger difference in the Rfree/Rfactor values?  Under what circumstances should one interpret large differences as overfitting?  I look forward to learning more about useful studies, criteria and guidelines you and other members of the ccp4bb community rely upon during (maximum likelihood) refinement.   Thank you in advance for your time, Best Regards, -Andy Torelli   Andrew Torelli, M.S. Ph.D. Candidate, Dept. of Biochemistry & Biophysics University of Rochester School of Medicine & Dentistry Box 712, 601 Elmwood Ave, Rochester, NY 14642