Nate, the common wisdom seems to be that setting aside 1000-2000 reflections is enough to ensure proper statistics anything more is overkill. I am not sure about the exact reference for these numbers, but you could start with Axel Brunger's original paper about refinement with FreeR. Also a couple of Ian Tickles latest papers deal with the issue. HTH Carsten

-----Original Message----- From: phenixbb-bounces@phenix-online.org [mailto:phenixbb- bounces@phenix-online.org] On Behalf Of Nate Nicely Sent: Tuesday, January 19, 2010 12:29 PM To: phenixbb@phenix-online.org Subject: [phenixbb] Reflection file conversion

I use d*TREK to reduce my data and phenix.reflection_file_converter to import the data. I seem to have had a bit of trouble generating the free R flags. It seems that the cap for the max number of free R reflections takes precedence over the percentage, even when the percentage is given expressly on the command line. I find myself setting the cap artificially high on the command line so that a full 10% is set aside.

I was taught to set aside 10% for the free R set, only going lower if setting aside the full 10% compromised completeness. I'm just wondering what is the rationale for the numeric cap? Am I making any error(s) with my methodology?

Nate N

_______________________________________________ phenixbb mailing list phenixbb@phenix-online.org http://phenix-online.org/mailman/listinfo/phenixbb

I use d*TREK to reduce my data and phenix.reflection_file_converter to import the data. I seem to have had a bit of trouble generating the free R flags. It seems that the cap for the max number of free R reflections takes precedence over the percentage, even when the percentage is given expressly on the command line. I find myself setting the cap artificially high on the command line so that a full 10% is set aside.

Yes, that's what you need to do.

I was taught to set aside 10% for the free R set, only going lower if setting aside the full 10% compromised completeness. I'm just wondering what is the rationale for the numeric cap? Am I making any error(s) with my methodology?

For the alpha/beta (or sigmaa) calculation you need a certain number of reflections per resolution bin to get reasonable estimates. Say you have 10 bins, 2000/10 = 200 is a sufficiently large sample per bin. I have to admit I'm uncertain about the pros and cons of setting aside a larger number of reflections. Ralf