Identification of Prognostic Genes by Combining 
Information across Different Institutions and
Array Platforms

Keith Baggerly
Section of Bioinformatics
Department of Biostatistics and Applied Mathematics
M.D. Anderson Cancer Center

Abstract:

Identification of genes that predispose individuals to 
cancer initiation or better clinical outcomes has been
the goal of many published microarray studies. In many
cases, both the raw and the clinical data are available
on the web, so it is natural to ask whether we can improve our 
identifications of "important" genes by combining information
across studies. Such meta-analysis is often difficult, in 
large part due to vagaries of processing from lab to lab 
and platform to platform which make the mappings of numerical
scores approximate at best. In this talk, we address the 
meta-analysis issue by defining a new mapping between different
types of Affymetrix platforms. Using this mapping, we reexamine
the raw data from two studies on lung cancer 
(Beer et al, Nat. Med, 2002, and Bhattacharjee et al, PNAS, 2001).
Whereas the initial studies tried to identify genes that were
associated with poor clinical outcome, we restrict our attention
further by focusing on genes found to have predictive value
over and above that obtained using readily available clinical 
covariates such as age and smoking history. We find that the
use of multiple datasets does increase our ability to find
interesting genes by increasing the sample size, and we
further find that the lists produced have a high density
of genes known to be important in cancer progression. Almost
none of the genes that we found were described in the initial 
papers.