FAQ: Difference between revisions

FAQ (view source)

Revision as of 15:54, 10 February 2011

1,081 bytes added , 10 February 2011

→‎why do the latest XDS/XSCALE versions only give a single table, with I/sigma>=-3 cutoff?

Kay

Bureaucrats

2,684

edits

@@ Line 71: / Line 71: @@
 === why do the latest XDS/XSCALE versions only give a single table, with I/sigma>=-3 cutoff? ===
-This was changed in the May 2010 version, and there is no keyword to get the old version of the tables back.
+This was changed in the May 2010 version. There is no keyword to get the old version of the tables back.
-The reason why this has been changed is that only the Signal/noise>=-3 table is meaningful because it describes the data in XDS_ASCII.HKL . These data are used "downstream" for conversion to amplitudes (XDSCONV and TRUNCATE implement French G.S. and Wilson K.S. Acta. Cryst. (1978), A34, 517.), structure solution, and refinement.
+Short explanation: the reason why this has been changed is that only the Signal/noise>=-3 table is meaningful because it describes the data that are used "downstream" for  structure solution, and refinement. These are also the numbers that should go into "Table 1" of your paper. The -3*sigma cutoff is used to ensure that strongly negative intensities - which of course should not exist - do not enter downstream steps.
-If only the positive intensities were output to XDS_ASCII.HKL then this would lead to wrong average intensities at high resolution, wrong conversion to amplitudes, wrong Wilson B factors, and ultimately to worse models.
+Longer explanation: the -3 sigma cutoff is actually imposed by [[XDSCONV]], not by [[XDS]] (which is why XDS_ASCII.HKL may also contain intensities that are below -3*sigma). [[XDSCONV]] performs the conversion from intensities to amplitudes (like TRUNCATE, it implements French G.S. and Wilson K.S. Acta. Cryst. (1978), A34, 517-525.). If e.g. only positive intensities were written out by XDS, and used by XDSCONV, then this would lead to wrong average intensities at high resolution, wrong conversion to amplitudes, wrong Wilson B factors, and ultimately to worse models.
-Consequently just these data are described in CORRECT.LP . These are also the numbers that should go into "Table 1" of your paper. By the way, SCALEPACK also uses a 3-sigma cutoff.
+Would it matter if the cutoff were not -3, but e.g. -4? Simple statistics can tell us what the percentage of reflections with intensity < -3*sigma is in a resolution shell that contains only noise (i.e. far beyond the useful resolution of your crystal): assuming a gaussian distribution of intensities around zero, we expect (100-99.73)/2 % = 0.135% of reflections to be less than -3*sigma, and another 0.135% to be larger than 3*sigma. For a useful resolution shell, however, the expected percentage of reflections with intensity less than -3*sigma would be much less. Thus, changing the cutoff (there is indeed a keyword for that - see XDSCONV documentation) would not matter much.
+By the way, SCALEPACK also uses a -3 sigma cutoff (see "SIGMA CUTOFF" at http://www.hkl-xray.com/hkl_web1/hkl/Scalepack_Keywords.html).
 == other questions ==