A Data Warehousing Approach for Genomics Data Meta-Analysis - Evolving Application Domains of Data Warehousing and Mining - page 140

Table 3. Summary of the methods of meta-analysis

Paper Chip Statistical Method Data Mining

(Hong and Breitling, 2008) Simulated data Fischer's χ 2 ,

Effect size,
Rank Product

(Hu et al., 2006) Affymetrix Fischer's χ 2 ,

Effect size,

(Jiang et al., 2004) Affymetrix FLD Hierarch. Clust.,

Rand. Forest

(Kim et al., 2007) ANOVA Hierarch. Clust.

(Lee et al., 2004) Affymetrix Coexpression, GO

(Rhodes et al., 2002) Affymetrix Fischer's χ 2 ,

(Moreau et al., 2003) Review P min , Fischer's χ 2 ,

RMA-Based Transformation
and Integration

In summary, in this subsection, we have dis-
cussed different approaches for analyzing com-
bined gene expression data. We can observe that
the experiments that are mentioned involve very
identical studies and micro-arrays technologies.
In the next section, we present experimental tests
that confirm these limits.

With Affymetrix micro-arrays, various methods
have been proposed for transforming and nor-
malizing gene expression arrays. Here we use
the RMA method (described previously) that is
based on quantile transformation for the transfor-
mation of distribution. This transformation gives
the same distribution to each array by taking the
mean quantile and substituting it as the value of
the data item in the original arrays. We applied it
for transforming combined Affymetrix datasets
in order to obtain the same distribution. Let us
assume that we have J datasets from identical
Affymetrix chips, then the algorithm for data
integration and transformation is described by
the following steps:

experimental tests

As seen above, the variability in experimental
environments such as RNA sources, micro-array
production, or the use of different platforms,
can cause biases because of variations among
distributions, scale of intensity expression, etc.
Such systematic differences present a substantial
obstacle to the co-analysis of multiple micro-array
data, resulting in inconsistent and unreliable
information.
In this section we present experimental tests
that combine data from identicalAffymetrix chips.
It is organized in three subsections. In the first and
the second subsections we show that in this special
condition, we may apply a quite simple technique
to combine data. Then in the third subsection, we
describe interesting tests on combinedAffymetrix
data sets; we present comparative tests between
individual analyses and meta-analyses combining
different expression datasets that demonstrate how
carefully the last ones have to be considered.

1. build the X array of dimension N x J where

each array is a column,

2. sort each column of X to give X sort

3. compute each row mean in X sort and assign

this mean to each element in the row to get
X' sort

4. compute X normalized by rearranging each col-

umn of X' sort to have the same ordering as
original X

The example below illustrates the method.