/support/technical/whitepapers/agcc_impact_libraryfile_change_whitepaper.pdf. It's a snapshot of the page taken as our search engine crawled the Web.
The web site itself may have changed. You can check the current page or check for previous versions at the Internet Archive. Yahoo! is not affiliated with the authors of this page or responsible for its content.
Negligible data impact due to Ecoli and Ecoli_ASv2 library file change Minor change in library files for Ecoli and Ecoli_ASv2 has
negligible effect on data
A minor difference between the historic and modern library files for the Ecoli and
Ecoli_ASv2 commercial arrays has been identified. (For information about the four affected
Custom Arrays please contact Affymetrix Support). The modern library files use a pre-
computed feature pitch (i.e., grid spacing) from the CIF file, that differs slightly from the feature
pitch computed by either GCOS or AGCC when it is not provided in the CIF file. This is due to
small rounding differences between the two computation methods. Most summarization routines
(e.g., MAS 5, RMA, and PLIER) are robust to individual feature changes, and therefore, even
less change is observed at the signal level. As a result this small change in feature pitch has a
negligible effect on the biological results.
To confirm this, an analysis was performed using the Latin Square data on U133A
expression array where the feature pitch was intentionally removed from the library files before
analysis to recreate the issue. It is important to note that the U133A array used is not affected by
the feature-pitch issue (see list of affected arrays), and is used merely as a representative array
type to illustrate the scale of the problem. Using the different feature pitches resulted in some
intensity differences in all 42 CEL files examined. About 60% of the features on the array
showed small intensity differences, but only 5% of the features had a relatively large
proportional change in intensity (10% or greater) (Figure 1). Furthermore, those with the largest
intensity differences were mostly confined to the lower intensity features. Again, we note that
most signal summarization techniques average out changes to any individual feature. The typical
correlation of Signal (summarized over probe sets using RMA) is 0.998 between the original
feature pitch setting and the deliberately altered setting. This compares very favorably to the
correlation between technical replicates of experiments which is normally between 0.97-0.99
(data not shown). While the % present call rate produced by the MAS5 algorithm is not always
an easily-interpretable metric, people generally monitor its behavior. Figure 2 displays the %
present for the feature pitch defined in the CIF file in blue, and for calculated value in red. The
largest % present difference observed using the two difference values for the feature pitch was
0.7% (average 0.21%) compared to 4.6% (average 1.27%) for the largest % present difference
between the technical replicates. After comparing the signal values and %P calls using the pre-
computed feature pitch and the feature pitch calculated by either GCOS or AGCC we conclude
that there is negligible impact on expression performance.
In summary, a small rounding difference has been identified between the modern and
historic library files for the Ecoli and Ecoli_ASv2 arrays for the feature pitch value. This results
in small changes in the calculated feature intensities in the CEL files. Since summarization
routines (e.g., MAS 5, RMA, and PLIER) robustly combine multiple features into a single
probeset value, minimizing the observed effect at the signal level. Examination of the signal
values and the %P calls revealed very minimal differences resulting from using the two different
values for feature pitch. In fact, the minimal technical variation seen in the experiments was
greater than the variation observed using the different values for feature pitch. Therefore we
conclude that the impact of using the modern library files is negligible on expression
performance.
Figure 1: Comparison of technical variation and the intensity difference resulting from
using the pre-computed feature pitch from the CIF file and the dynamically generated
feature pitch. RMA signals were generated from a CEL file using a CIF file with the undefined
feature pitch are plotted against RMA signals from a CEL from the same DAT file generated
using a CIF file with the feature pitch defined in the file A. The R
2
value is 1.00 and the linear fit
is y=-0.008+0.999*x indicating that the two arrays are very highly correlated and there is no
overall bias (i.e., no significant impact on derived fold change). In B a similar plot for two
technical replicates (including the one used in A) with the defined feature pitch. The R
2
value is
0.999 and the linear fit is y=-0.055 + 1.008*x. C and D show MvA plots for arrays in A and B
respectively. The absolute value of fold change in both cases is less than 1.4 (red line). The
variation introduced by the small difference in the value used for feature pitch is less than that
observed for the technical replicates, and the ratios.
Minor change in library files for Ecoli and Ecoli_ASV has negligible effect on data.
p. 2
40%
42%
44%
46%
48%
50%
52%
54%
56%
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Expe rime nt Numbe r
%
P
c
a
ll
Pitch in CIF
Pitch not in CIF

Figure 2: Comparison of the MAS5.0 %P calls using the pre-computed feature pitch from
the CIF file and the dynamically generated feature pitch. The MAS5.0 algorithm was used to
generate the %P calls for the 42 arrays. The y-axis contains the %P call and the x-axis is the
individual experiments, with the three technical replicates grouped together by the vertical
gridlines. The %P for arrays analyzed with the pre-computed feature pitch in the CIF file are
shown in Blue, and those with the dynamically calculated feature pitch are shown in Red. The
largest difference observed difference using the two difference feature pitches was 0.7% (black
arrow). In general the differences in %P between technical replicates for the same experiment
were larger than the difference between the gridding algorithms (red arrows).

Minor change in library files for Ecoli and Ecoli_ASV has negligible effect on data.
p. 3