|
|
Prognostic model for glioma patients Yi-Hong Zhou, Kenneth R. Hess, W. K. Alfred
Yung
University
of Arkansas for Medical Science Arkansas Cancer Research Center, Little Rock, AR
and University of Texas M.D. Anderson Cancer Center, Houston, TX. E-mail: yzhou@uams.edu
According to the most recent CBTRUS report, more than one-third of primary brain
and CNS tumors are astrocytic gliomas, from which about three-fourth are
malignant forms, anaplastic astrocytoma (AA) and glioblastoma multiforme (GBM).
For over thirty years there are no significant improvements on the clinical
outcome of these patients, whose survival times are highly variable, especially
for those of the same grade.
We hypothesized that it was possible to stratify
patients for optimal treatment plans from prognosticating survival times based
on expression level of genes in tumor specimens that have been shown to be
associated with critical process of cancer development.
To establish such a
powerful statistical prognosis model, we used 84 cDNA from retrospective glioma
patients (43 AA and 41 GBM) as a training set, and included multiple prognostic
gene markers that have been previously determined mainly by immunohistochemistry
for glioma patients and several new markers identified by us.
All the expression
variables were quantified by real-time quantitative RT-PCR.
The value of each
expression variable was expressed as a ratio of the
copy number of the marker gene transcript to the copy number of β-actin
transcript, which has been shown to be a fair internal control gene for gliomas.
Our current prognostic model for gliomas has been greatly improved by
including the expression variates of eight genes (PAX6, PTEN, EGFR,
VEGF, CDK4, MMP2, IGFBP2, and RPS9) in
addition to three clinical variates (age, histology, and recurrence status).
The
proportion explained by the model (R^2) has reached up to 62%, in contrast to
the model including only clinical variabes (R^2 = 30%).
We will further improve
the model by identifying new prognostic expression variables and including them
to the model.
We will perform regression diagnostics to verify assumptions of
our models, and recursive partitioning analysis and hierarchical cluster
analysis to identify additional patterns in the data.
An independent set of
retrospective glioma specimens shall be used to validate our model.
Copyright © 2004 American Association for Cancer Research. All rights
reserved.
Source: http://aacr04.agora.com/planner/displayabstract.asp?presentationid=6846
|
