Gene
expression profiling in breast cancer challenges the existence of
intermediate histological grade
Sotiriou
C, Wirapati P, Loi S, Harris A, Bergh J, Smeds J, Farmer P, Praz V,
Haibe-Kains B, Lallemand F, Buyse M, Piccart M and Delorenzi M
Background: The
histological grade (HG) in breast cancer provides important prognostic
information. However, its interobserver variability and poor
reproducibility, especially for tumours of intermediate grade, have limited its
clinical potential. We hypothesized that molecular characterization of the
grade may allow for full exploitation of the association between the
grade and relapse beyond the ability of traditional grading
procedures.
Methods: Six
datasets totalling about 700 primary breast cancers, mostly publicly
available data, were used in the analysis. Gene expression profiles
(GEP) from Affymetrix U133A GeneChips were contrasted between HG 1
(low grade) and HG 3 (high grade) tumours on a training set of 64
estrogen-receptor-positive breast cancer samples. A set of genes
positively and negatively correlated with grade was identified on this
training set and chosen as grade reporting genes. A scoring system
called the ‘gene-expression grade index’ (GGI), which essentially
summarizes the grade reporting genes by their average expression
level, was introduced. The GGI was applied to patients not used in the
gene selection to test its prognostic value.
Results:
Using 33 HG 1 and 31 HG 3 ER-positive breast carcinomas, 112 Affymetrix probe
sets were significantly upregulated in grade 3 and 16 in grade 1, at a
stringent and objective cut-off P value of 0.05 for a false discovery
count >0. These 128 probe sets represent 97 different reporter
genes. Quantifying the level of expression of these reporter genes with the
GGI, many tumors in the HG 2 (intermediate grade) populations
assume values typical for the HG 1 and HG 3 groups in the same
study. The HG 2 tumors can therefore be naturally split into a
‘HG 1 like’ group and a ‘HG 3
like’ group, to which we attribute a gene
expression grade (GG) of 1 and 3, respectively. Their survival curves follow
the GGI and are similar to those of the HG 1 and HG 3 groups,
respectively, splitting HG 2 into a good prognosis group and a poor prognosis
group (Fig. 1). Similar observations
were made in the different datasets analysed, in untreated as well as in
systemically treated patients, and on the three different main types of
microarray platforms, with substantial variability in the number of
reporter genes available. Almost all known clinicopathological
variables were significantly associated with clinical outcome in univariate
analysis, while in a multivariate model only the GG, tumour size and
nodal status were significant factors. Replacing the HG with the GG
significantly improved the prognostic two-group classification obtained
with the Nottingham Prognostic Index.
Conclusion:
Gene-expression-based grading has the potential to significantly improve
current grading systems by rendering them more objectively measurable
and improving their prognostic value. The superior performance of
the two-grade GG system challenges the purpose of classifying
tumors as of intermediate grade. Reproduction of these findings in
four independent datasets, and across different platforms and with a
simple computational system, gives hope that the approach will prove
robust and reliable.