Microarrays, as efficient tools to simultaneously monitor the expression of tens of thousands of genes, have been widely applied in both mechanistic and decision-making research during the past decade 
. The large number of commercially available microarray platforms has expanded the use of the technology and made it more widely available to different laboratories. However, left unresolved is the issue of whether inter-platform differences may conceal or confound biologically significant information with respect to potential biomarkers and prediction models. Thus, the concern that one needs to stay within a particularly microarray platform manufacturer slows down the identification and qualification of genomic biomarkers 
The extent to which different microarray technologies influence the identification of differential gene expression has been addressed by a large number of studies and is the subject of a review paper 
. Despite the conflicting information given by a handful of early published studies where both concordance
between technologies was demonstrated, the maturation of microarray technology and data analysis methods has led to improved cross-platform correlations
. Moreover, the first phase of FDA-led Microarray Quality Control project (MAQC-Ι) has further confirmed the reproducibility of the identification of differentially expressed genes across different platforms 
. These studies suggest that similar results should be expected regardless of microarray platform if appropriate experimental and analysis protocols are applied, meaning that mechanistic research can incorporate datasets from multiple sources without significant concern about platform-specific affects.
The clinical use of array-based diagnostics is relatively late in coming; this is partially due to the demand of a substantial number of patient samples to be used for training, since estimates of a predictor's error rate during model construction are more prone to be biased for small datasets
. Therefore, an attractive approach would be the re-use of relevant pre-existing sets of expression profiles as training data. Although researchers have demonstrated that reciprocal validation can be achieved using different patient cohorts and microarray platforms
, few benchmark analyses have been carried out until recently to confirm the feasibility of re-using datasets obtained from different platforms for diagnostic models. Based on the toxicogenomics datasets generated in phase II of the MAQC project using both Rat Genome 230 2.0 Array (Affymetrix platform) and Rat Oligo 2-color G4130A Array (Agilent platform) on the same tissue samples, our recent study
evaluated and confirmed that high cross-platform concordance of predictive signature genes and classifiers can be achieved for binary classification. However, in reality, decision-making is not always binary. For example, subtype identification in disease diagnosis
, toxicant discrimination
and the stratification of toxicity severity in drug risk/safety assessment
can, in most cases, only be achieved using multiple-class prediction. Thus, the consistency of microarray platforms with regard to multiple-class prediction discussed in this study is also of importance to the future success of microarray-based predictive models in clinical application and safety evaluation.
The primary issue we addressed is the comparability of models constructed from different platforms. We then further evaluated cross-platform consistency with regard to whether predictive signature features selected on one platform could be directly used to train a model on the other platform and whether predictive models trained using one platform could predict datasets from the other platform with comparable performance. In this study, three commonly-used multi-class machine learning algorithms were applied: fuzzy k-nearest neighbors (FKNN)
, linear discriminant analysis (LDA)
and support vector machine (SVM)
. The results provide a baseline confirmation of the cross-platform consistency of multiple-class prediction.