I'm happy to change the column titles if it makes it clearer. Actually
the "I/sigma" column in the Scala output is not very useful:
it is <I> / RMSscatter, ie the mean intensity/mean error, for
individual observations, not taking into account multiple
measurements. Because it is ratio of means (rather than a mean of
ratios), it can behave oddly depending on the distribution of
intensities, for instance giving an overall value which is outside the
range of values in resolution bins. It is the ratio of the previous
two columns.
On the other hand the column labelled "Mn(I)/sd" is the mean of ratios
for each reflection, ie< <I>/σ(<I>) > and does take into account the
multiplicity of measurements, so is much more relevant as an indicator
of data quality
see
http://www.ccp4wiki.org/~ccp4wiki/wiki/index.php?title=Scaling_experimental_intensities_with_Scala
Scala also outputs a convenient "Table 1" summary
On 23 Mar 2009, at 15:50, James Holton wrote:
I guess when I talk about signal-to-noise I assume the one that is
most relevant to the task at hand. So, to me, I/sigma(I) at the
phasing step would be the average intensity (I) divided by the sigma
(standard deviation) assigned to it AFTER scaling/mergeing. I admit
that the "I/sigma" column from SCALA is potentially confusing, but
if you are dealing with spot intensities, this is the first I/sigma
you think about, so I guess this is what Phil was thinking.
Personally, I think the descriptions of the columns in this table
are clear if you read the caption before the table in the SCALA
output, but Tassos is right that an alarming number of people have
never done this. RTFM?
When it doubt, use mtzdmp to see what the average values of the data
columns really are.
-James Holton
MAD Scientist
Anastassis Perrakis wrote:
I like to think of things in terms of signal-to-noise, and one can
use a
rearrangement of the Crick-Magdoff equation to tell you what the I/
sigma
of your data set needs to be for delta-F to be greater than
sigma(delta-F):
I/sigma(I) > 1.3*sqrt(Daltons/sites)/f"
where:
I/sigma(I) is the signal-to-noise ratio of the data set required to
solve it by MAD/SAD
Daltons is the molecular weight of the protein in amu
sites is the number of Se sites
f" is the f" of those sites (in "electrons")
let me see .... we recently solved a 200 residues protein, 4 mol
AU, with 2 Se per mol, total 8 Se.
Since 160 residues were ordered, I will make for you a discount,
18,000 D/monomer, 70,000 in AU.
I truncated data to 4.2 for Se search.
1.3*sqrt(70000/8)/6.5= 19
Statistics from Scala:
N 1/d^2 Dmin(A) Rmrg Rfull Rcum Ranom Nanom Av_I SIGMA I/
sigma sd Mn(I/sd) 1 0.0098 10.11 0.048 0.049 0.048
0.036 349 4967 419 11.9 342 25.8 2 0.0196
7.15 0.050 0.044 0.049 0.031 707 5360 462 11.6
372 28.8 3 0.0293 5.84 0.089 0.062 0.057 0.047
975 1634 224 7.3 177 19.4 4 0.0391 5.06 0.065
0.048 0.059 0.039 1140 2107 207 10.2 218 21.0
5 0.0489 4.52 0.061 0.043 0.060 0.034 1315 2523 227
11.1 253 21.9 6 0.0587 4.13 0.072 0.051 0.062 0.035
1470 2142 223 9.6 242 20.2 7 0.0685 3.82 0.091
0.061 0.066 0.042 1605 1566 203 7.7 219 16.1 8
0.0782 3.57 0.128 0.086 0.071 0.052 1737 1034 186
5.6 199 12.3 9 0.0880 3.37 0.189 0.137 0.077 0.074
1859 667 181 3.7 187 8.9 10 0.0978 3.20 0.314
0.224 0.085 0.129 1940 374 170 2.2 178 5.5
So, that would support your argument.
HOWEVER that would mean looking at the M(I/sd) in the table!!!
"Mn(I/sd)" is not the same as "I/sigma" in Scala notation!!!! Most
people think of I/sigma(I) in your notation,
to be the I/sigma in the scala output, or the I devided by sd in
the Denzo output. These are (very) different.
I am not sure which you meant since I/sigma(I) is not the full
notation (place the <> in the favorite place first ...), but it
seems correct if you meant Mn(I/sd) which most people do not quote
or use much ;-)
Greetings,
Tassos
A.
