Evaluation of inter-laboratory and cross-platform concordance of DNA microarrays through discriminating genes and classifier transferability.

Microarray technology has great potential for improving our understanding of biological processes, medical conditions, and diseases. Often, microarray datasets are collected using different microarray platforms (provided by different companies) under different conditions in different laboratories. The cross-platform and cross-laboratory concordance of the microarray technology needs to be evaluated before it can be successfully and reliably applied in biological/clinical practice. New measures and techniques are proposed for comparing and evaluating the quality of microarray datasets generated from different platforms/laboratories. These measures and techniques are based on the following philosophy: the practical usefulness of the microarray technology may be confirmed if discriminating genes and classifiers, which are the focus of most, if not all, comparative investigations, discovered/trained from data collected in one lab/platform combination can be transferred to another lab/platform combination. The rationale is that the nondiscriminating genes might not be as strongly regulated as the discriminating genes, by the biological process of the tissue cells under study, and hence they may behave more randomly than the discriminating genes. Our experiment results, on microarray datasets generated from different platforms/laboratories using the reference mRNA samples in the Microarray Quality Control (MAQC) project, showed that DNA microarrays can produce highly repeatable data in a cross-platform cross-lab manner, when one focuses on the discriminating genes and classifiers. In our comparative study, we compare samples of one type against samples of another type; the methodology can be applied to situations where one compares one arbitrary class of data against another. Other findings include: (1) using three discriminating-gene/classifier-based methods to test the concordance between microarray datasets gave consistent results; (2) when noisy (nondiscriminating) genes were removed, the microarray datasets from different laboratories using common platform were found to be highly concordant, and the data generated using most of the commercial platforms studied here were also found to be concordant with each other; (3) several series of artificial datasets with known degree of difference were created, to establish a bridge between consistency rate and P-value, allowing us to estimate P-value if consistency rate between two datasets is known.