A Data Warehousing Approach for Genomics Data Meta-Analysis - Evolving Application Domains of Data Warehousing and Mining - page 144

genes is lost, the expressed genes are not coher-
ent. One explanation comes from the method of
transformation distribution that results in two
normalizations. Another reason could be on the
inadequacy of data to be combined.
Table 6 gives similar results on GSE7216 data.
We had previously conducted similar individual
analyses on GSE7216 datasets in order to extract
differentially expressed genes for the cytokines
on reconstituted skin compared to control. We
observe an important distortion again.
These results demonstrate that meta-analyses
combining pre-normalized datasets may provide
unreliable information.

• in the second column, the number of differ-

entially expressed genes elicited by indi-
vidual analyses on GSE6281 and GSE7216
data only in each of the four cases,
in the third column, the number of dif- •

ferentially expressed genes elicited by a
meta-analysis combining raw CEL data of
GSE6281 and GSE7216 series in each of
the four cases (we applied the RMA-based
transformation method),
in the fourth column, the number of dif- •

ferentially expressed genes common in the
two previous lists.

While meta-analysis results are closer to indi-
vidual ones, we still find an important distortion
in genes found in common.
This first test on the integration of only two
individual data series conducted on the same
gene chip shows how meta-analyses should be
conducted carefully. Integrating pre-normalized
data may be irrelevant. For Affymetrix series,
CEL raw data are more adequate for combination.
When datasets result from different gene chips,
further problems occur: it becomes essential to
guarantee that probe sets on different chips cor-
respond to identical (or approximate identical)
oligo sequences. In such cases, meta-analyses
must pre-process data with probeset sequence

Integration of Raw Data
and Normalization

In this test we combined raw data from GSE6281
CEL files (34 files) and GSE7216 CEL files (25
files). We used the Bioconductor 17 package affy
to read CEL files and to normalize them globally
with RMA procedure by quantile normalization
as shown previously in the subsection on “RMA-
based transformation and integration”.
On Table 7 and Table 8 that compare new meta-
analyses and previous individual analyses with the
same p -value and fold change, we have:

Table 6. Comparison of differentially expressed genes elicited by an individual analysis on GSE7216
and a meta-analysis on pre-normalized data of GSE6281-GSE7216 series

Comparison Individual analysis Meta-analysis Gene in common

IL19-control 82 1103 70

IL20-control 207 1438 192

IL22-control 606 2608 576

IL24-control 363 2020 347

IL26d-control 0 858 0

KGF-control 224 1619 197

IFNg-control 253 1875 248

IL1b-control 214 2121 196