Sg. Hilsenbeck et al., Statistical analysis of array expression data as applied to the problem oftamoxifen resistance, J NAT CANC, 91(5), 1999, pp. 453-459
Background: Although the emerging complementary DNA (cDNA) array technology
holds great promise to discern complex patterns of gene expression, its no
velty means that there are no well-established standards to guide analysis
and interpretation of the data that it produces. We have used preliminary d
ata generated with the CLONTECH Atlas(TM) human cDNA array to develop a pra
ctical approach to the statistical analysis of these data by studying chang
es in gene expression during the development of acquired tamoxifen resistan
ce in breast cancer. Methods: For hybridization to the array, we prepared R
NA from MCF-7 human breast ceh tumors, isolated from our athymic nude mouse
xenograft model of acquired tamoxifen resistance during estrogen-stimulate
d, tamoxifen-sensitive, and tamoxifen-resistant growth. Principal component
s analysis was used to identify genes with altered expression. Results and
Conclusions: Principal components analysis yielded three principal componen
ts that are interpreted as 1) the average level of gene expression, 2) the
difference between estrogen-stimulated gene expression and the average of t
amoxifen-sensitive and tamoxifen-resistant gene expression, and 3) the diff
erence between tamoxifen-sensitive and tamoxifen-resistant gene expression.
A bivariate (second and third principal components) 99% prediction region
was used to identify outlier genes that exhibit altered expression. Two rep
resentative outlier genes, erk-2 and HSF-1 (heat shock transcription factor
-1), were chosen for confirmatory study, and their predicted relative expre
ssion levels were confirmed in western blot analysis, suggesting that semiq
uantitative estimates are possible with array technology. Implications: Pri
ncipal components analysis provides a useful and practical method to analyz
e gene expression data from a cDNA array, The method can identify broad pat
terns of expression alteration and, based on a small simulation study, will
likely provide reasonable power to detect moderate-sized alterations in cl
inically relevant genes.