Gene expression profiling by DNA microarrays has become an important tool for studying the transcriptome of cancer cells, and has been successfully used in many studies of tumour classification and of identification of marker genes associated with cancer [e.g. [1–3]]. With an increasing number of microarray data becoming available, the comparison of studies with similar research goals, e.g. to identify genes being differentially expressed in normal versus tumour tissue, has gained high importance. In general, the evaluation of multiple data sets promises to yield more reliable and more valid results since these results are based on a larger number of samples and the effects of individual study-specific biases are weakened. However, the comparison of results from different microarray studies is hampered by the fact that different studies use different protocols, microarray platforms and analysis techniques. The question whether the results of gene expression measurements obtained by different platforms can be compared has been addressed in several studies [4–7]. It has been found that results derived from the measurements like lists of tumour subtype marker genes [5] or measures of intra-study correlation of gene expression patterns [6] can be compared and thus inter-validated between different platforms. However, the measures of gene expression themselves could not be directly compared between different platforms [4, 7]. Some studies propose methods for meta-analysis of microarray data with the goal to identify significantly differentially expressed genes across studies by using statistical techniques that avoid the direct comparison of gene expression values [8–14].

The goal of this study is to investigate the benefit of performing supervised classification analyses across disparate sources of microarray data. Methods of supervised classification analysis render it possible to automatically build classifiers that distinguish among specimens on the basis of predefined class label information (phenotypes), and in many cancer research studies [e.g. [1–3]] the application of these methods has shown promising results of improved tumor diagnosis and prognosis. However, as pointed out by several authors, there is a strong need for independent validation of these results, and an increase in sample size is recommended for future studies [15, 16]. We therefore chose to explore how gene expression data from different studies can be directly combined, especially for an integrated classification analysis. Such an integrated analysis promises to be a valuable tool for validation of classification results obtained in a single study, and might yield improved results because it is based on a larger number of samples.

Recently, Wright et al. [17] have proposed a statistical method based on Bayes' rule to classify cancer specimens by their gene expression profiles. They were able to classify oligonucleotide microarray data from one study with a predictor derived from cDNA microarray data from a different study. Here, we evaluate the feasibility of building predictors from and classifying microarray data independent of the platform used for expression profiling. The general approach to first derive numerically comparable measures of gene expression from different platforms (data integration) and then to apply supervised classification on the integrated data was successfully applied in first attempts to classify cancer microarray data generated with multiple array platforms [18, 19].

We adopt this approach and demonstrate the use of two data integration methods, namely median rank scores, which has already been successfully applied for comparability assessment of five different breast cancer microarray data sets [19], and quantile discretization which has not been used in the context of microarray data analysis before. For supervised classification analysis, we use support vector machines (SVM), a well-established machine learning technique for classification of microarray data [20, 21]. Integrated cross-platform classification of cancer is demonstrated for three pairs of publicly available data from microarray studies on different types of cancer [22–27]. To investigate the hypothesis that an integrated analysis of data from different microarray studies can yield results not obtained by a single study, we chose to investigate two leukemia data sets in more detail and studied differences in gene expression profiles between the cytogenetically defined subgroups t(15;17), t(8;21) and inv(16), all associated with a favourable prognosis [28, 29], and samples with normal karyotype lacking mutations in FLT3 or RAS, thought to belong to an intermediate risk group [30–32]. While differences between the first three groups are prominent and were detected in multiple studies [33, 26, 27], evidence about the homogeneity of the normal karyotype group and the associated genes is still lacking. The list of genes selected in an integrated analysis of both studies is compared to the lists of genes selected in two analyses performed separately on either study.

We investigated six publicly available cancer microarray gene expression data sets to perform cross-platform supervised classification analysis. We selected three pairs of studies, each examining the same type of cancer, i.e. breast cancer, prostate cancer and acute myeloid leukaemia, respectively. All pairs of studies allowed for either classification of cancer versus normal tissue or cancer subtype differentiation. Each pair was chosen to consist of one study using cDNA arrays and one study based on oligonucleotide arrays. We studied how to combine pre-processed data sets measured with different microarray platforms for an integrated classification analysis. The process can be divided into the following main parts: First, we determined the overlap of genes common to both platforms using the UniGene database. Next, we derived numerically comparable quantities from the expression values of both platforms by application of median rank scores or quantile discretization. Then, the support vector machine algorithm, an approved method for supervised classification analysis, was applied to different classification settings.

Data integration

Figure 1 shows for all three study pairs the number of common UniGene clusters (genes) represented on both platforms. Since there is only a moderate overlap of UniGene clusters for the pairs of array platforms, many probes cannot be used for cross-platform analysis. The number of microarray features used for cross-platform analysis is further reduced by averaging expression values of probes on the same platform that map to the same UniGene cluster. As a result, only 40–50% of genes are retained for cross platform analysis.

As the next step, we applied the median rank scores (MRS) method or quantile discretization (QD). In order to check whether the comparability of the data from different platforms is improved after data transformation by these methods, we compared the distribution of gene expression values per microarray between arrays of different studies. We selected one microarray per study and produced a quantile-quantile plot (QQ-plot) for every pair of microarrays from corresponding studies as shown in Figure 2. In every QQ-plot the quantiles of all gene expression values from a first microarray are plotted against the quantiles of all gene expression values from a second microarray. If the gene expression values of the two different microarrays share the same distribution, the points in the plot should form a straight line. As can be seen in Figure 2, the distribution of expression values of microarrays of different studies is much more similar after application of MRS in comparison to non-integrated data. As an effect of QD, the quantiles of the expression values of all microarrays in the integrated studies are equal by definition, resulting in points in the plots forming a straight line.

Classification analysis

After data integration by the median rank scores method or quantile discretization, respectively, two different types of cross-platform classification analyses were performed: training of a classifier on only one data set of a pair followed by classifier evaluation on the other data set, and classifier training and testing on data instances randomly chosen by a cross validation from the combined data set.

The first type of analysis was performed on non-integrated data and on integrated data, respectively. Evidently, without data integration, a classifier created on one set cannot correctly classify data instances of the other set (Figure 3). This is clearly indicated by prediction accuracies being similar to or worse than the prior prediction rates, i.e. the prediction accuracy of a classifier which always predicts a data instance to be an element of the dominating class. The only exception is the prostate cancer data, where high classification accuracy was achieved after training on the data set of Welsh et al. and classification of the data of Dhanasekaran et al. Data integration improves the results in cases of the prostate and breast cancer studies (p-values < 0.01, except for classification of the data of Dhanasekaran et al, where a high classification accuracy was already achieved on the non-integrated data set). We conclude for these two pairs of studies that data integration enables the successful application of classifiers trained on one data set to a comparable data set generated with a different platform. This conclusion does not hold for the AML studies. Here, only the result for building a classifier based on the data of Bullinger et al. and classifying the data of Valk et al. improved after application of median rank scores or quantile discretization (p-value < 0.1).

Except for the pair of breast cancer microarray data sets, the application of the MRS versus QD showed no significantly different effect on the achieved classification accuracies. For the training on the data set of Gruvberger et al. and classification of the data of West et al., the classification result was significantly better after application of QD in comparison to the result obtained after using the MRS method. In all other cases both methods can be considered equivalent.

In addition to the above mentioned separated training and validation, cross-validation analyses were performed on combined data sets. High classification accuracies were achieved with training and testing on data instances randomly chosen from both data sets (> 85%; see table 2). Although the integrated classifiers only operated at less than 50% of all genes, classification accuracies for integrated classifiers were nearly as high or even markedly improved in comparison with classification accuracies achieved for single data sets only. In the case of the breast cancer studies, the results were better than the accuracies achieved by cross-validation on each of the pre-processed single sets with all available microarray features.

both data sets integrated
	MRS	QD
Prostate cancer	97.67 %	97.56 %
Breast cancer	87.01 %	88.97 %
Acute myeloid leukemia	90.60 %	90.20%
original data
Prostate cancer	Dhanasekaran et al.	Welsh et al.
	95.28 %	99.09 %
Breast cancer	Gruvberger et al.	West et al.
	80.52 %	86.73 %
Acute myeloid leukemia	Bullinger et al.	Valk et al.
	68.53 %	99.90 %

In order to check whether similar classification results could be obtained with another method of supervised classification analysis, we repeated the above described experiments using the method of nearest shrunken centroids classification (also known as "Prediction Analysis of Microarrays", PAM) [34]. As presented in Additional Files 1 and 2, the classification results obtained with PAM are similar to those obtained by SVM.

Selection of genes with discriminative expression patterns

To show the potential of an integrated cross platform analysis, we generated lists of genes forming discriminative expression patterns by means of recursive feature elimination (RFE) analysis for the leukemia studies (see Methods for details). We generated six lists of genes, two lists for an analysis of the combined leukemia studies, integrated by MRS or QD, and two lists for each of the two leukemia studies analysed separately, using only samples of the either the MRS or QD data which belong to one study. A number of 512 elements was selected for each list, which corresponded to minimal cross-validated error rate in the integrated analyses of data from both leukemia studies. Interestingly, the intersection of the lists generated in analyses using only data of one of the two leukemia studies comprises only about 40 UniGene clusters, independently of whether MRS or QD was used (Figure 4). In the sets generated by an analysis of both studies together, integrated by MRS or QD, many genes were selected that were lost in the analyses based on a single study (Figure 4). These include important genes with regard to the biology of leukemia, like RXRA, PBX3, ABL2, SOCS1, and EGR2 (see Additional File 3 for annotated lists of selected genes; Additional File 4 contains all six gene lists ordered by gene ranks as determined in RFE analysis).

Finally, we used hierarchical clustering as a visualization method to display coherence in gene expression of the genes selected by RFE in the leukemia studies. We first clustered the data of both leukemia studies separately based on the genes selected by RFE on either set. As shown in Figure 5(a,b), the samples of Valk et al. were perfectly grouped according to their karyotype while in the data of Bullinger et al. samples with karyotype t(8, 21) and inv(16) were not grouped homogeneously. Then, we clustered the data of Valk et al. using only genes found to be discriminative on the data of Bullinger et al. (Figure 5c). Figure 5d shows the reverse case, a clustering of the data of Bullinger et al. based on the gene selection on the data of Valk et al. For the selected groups of genes, coherence in gene expression between samples of the same karyotype was weak when results of an analysis solely based on one leukemia data set are transferred to the other leukemia data set, as samples of the same karyotype were not grouped homogeneously. Figure 5e and 5f show clustering results on all samples of both studies using gene lists integrated either by MRS or QD. Here we can observe a much more consistent grouping of the samples according to their karyotype than that observed in Figure 5c and 5d. Still, both methods of data integration are not able to fully eliminate study specific self-similarity of samples, as the samples form clusters according to study origin.

In this study we showed that classification of cancer microarray data can be markedly improved by cross-platform classification analysis of gene expression data from different studies with similar focus. Key techniques for cross-platform classification analysis were data integration methods rendering microarray data numerically comparable across platforms in combination with well-established machine learning techniques for generation of predictive models. An obvious advantage of an integrated classification analysis is the improved generalization performance and reliability of the resulting predictive models (classifiers) since they are found and validated on a larger number of samples, thus the effect of study-specific biases can be reduced. For all study pairs used here, we achieved high classification accuracies when using data samples randomly chosen from both data sets of a comparison pair for classifier building and testing. Our findings endorse the encouraging results of first attempts of multi-platform microarray classification analysis [18, 19].

For integration of microarray measurements from different platforms, Bloom et al. [18] used a scaling approach based on measurements for one common reference RNA sample. As hybridization results for such a common reference RNA sample are normally not available for different microarray studies and platforms (especially in the case of custom made cDNA arrays), we applied the median rank scores method [19] and quantile discretization for data integration. Besides the problem of integrating microarray data that have been measured with different platforms, a general problem in combining measurements from different gene expression studies is the variability between results of different studies. This is primarily due to biological differences among the samples of different studies, differences in the technical procedures to obtain gene expression measurements, and random variation. The use of methods providing an abstraction of data like ranks or discretized values reduces this variability at the price of reduced information. Therefore, data sets processed by MRS or QD can not be considered as a suitable input for every kind of analysis purpose. However, for the aim of cross-platform classification analysis, the combination of such abstraction methods with a sophisticated machine learning technique like the support vector machine used here helps to compensate for this loss of precision, and can yield useful results.

Even when a classifier is built on one data set of a pair of compared studies and the samples of the other study are classified, good classification results can be observed for the prostate and breast cancer studies. In this case, the generalization ability of the classifier is sufficient to correctly classify most of the samples of the other study, and thus the classifier obtained on the data of one study can be validated by the data of another study. In contrast, the results for the AML studies indicate that the generated classifiers based on only one of these studies are too specific. This might be due to fact that the sample sets of either study are not representative enough to cover all characteristic transcriptional features observable for the investigated phenotypes. Indeed, the results for the cross-validation analysis using samples from both AML studies show that classifiers with better generalization performance can be obtained underlining the potential of a cross-platform classification analysis.

Selection of discriminative gene expression signatures is an important task frequently performed in microarray studies. Here, we applied RFE analysis for selecting subsets of genes with distinctive expression patterns on the data of the leukemia studies of Bullinger et al. [26] and Valk et al. [27]. For visualization of the coherence in gene expression of the genes selected by RFE in the different studies we performed hierarchical clustering. Gene sets selected only on data of one study show poor coherence in gene expression for the karyotype groups of samples on the other set. Clustering results observed for gene sets selected on the combined set are more consistent. Therefore, these discriminative gene sets are apparently of more general validity. On the other hand, cluster analysis showed that neither of the two methods of data integration was able to entirely overcome study specific self-similarity of the leukemia samples. For cross-platform classification analysis, however, the MRS and QD algorithms yielded good results.

The analysis of gene lists obtained by RFE indicated that gene signatures can be generated on a combined set that comprise important genes that were not part of gene signatures generated on either set alone. Notably, the intersection of lists from the Bullinger and Valk data sets with the list from the combined set contained only a few genes, none of them to be known of high importance in the context of AML. Similarly, the intersection between the Bullinger and Valk data sets was not large (Additional File 3). In contrast, the list obtained from the combined data set contained many genes well known to be involved in leukemia pathogenesis, like PBX3 [35], the retinoid receptor X [36], the ABL2 tyrosine kinase [37] or early growth response 2 [38]. In addition, many genes in the combined list are prominent oncogenes or tumour suppressors, like BCL2 [39] or ERBB3 [40]. Most notable is the inclusion of human telomerase TERT, which has been found by Hahn et al. to be one of three necessary factors for transforming a normal cell into a tumour cell [41].

We compared the gene lists generated by the RFE method to the result of a meta-analysis approach as described by Rhodes et al. [8]. This method aims at identifying genes that show reproducible standardized differences in mean expression between phenotype groups across studies. For this, a p-value is calculated for every gene in both leukemia studies separately, in order to quantify the significance of differences in mean expression between phenotype groups within a study. Then, the study specific p-values are combined to a test statistic S and significance values for this test statistic by a permutation approach are calculated (for details see [8]). At a significance level of p = 0.01, 43 genes were selected by this meta-analysis approach. Of these 43 genes, 12 genes were also found in the list generated by an RFE analysis of the data of both studies integrated by MRS, 19 were also found in the list generated by an RFE analysis of the data of both studies integrated by QD. This result shows that the gene lists selected by RFE analysis also contains genes that would have been found in an independent meta-analysis, but that also many different genes are selected. This is not surprising, as there are essential differences in both approaches. The meta-analysis performed here applies a univariate statistical test to find genes with a significant difference in group means of expression values, whereas the SVM based RFE analysis is a multivariate approach which also considers combinations of genes and selects genes with maximum influence on the discriminative performance of a classifier. While interpreting a gene list generated in a RFE analysis, one has to keep in mind that the main goal of methods like the SVM based RFE approach used here is to generate signatures that allow for accurate classification of phenotypes. These gene signatures are unlikely to contain all and only genes that are most relevant to the genetic differentiation between complex disease phenotypes. The task to find the complete set of only those relevant genes out of gene expression data is much more demanding and might pose an irresolvable challenge as the changes of gene expression profiles recorded by microarrays are mostly secondary and tertiary effects and not the primary ones. With microarrays one observes the avalanche of gene expression changes, not necessarily the small pebble causing it. First promising concepts and methods to work on the task to find the set of relevant genes have been proposed [42], but their usefulness to address biological questions has still to be thoroughly investigated. However, the finding that RFE signatures generated by an integrated analysis of both leukemia studies contained genes that are described as being relevant for tumor biology in the literature, and that were not found in either single set analyses, shows the potential of cross-platform microarray data integration to be useful not only to improve results for phenotype classification but also for generation of gene signatures that contain more biologically interesting genes.

Considering integrated classification analyses in general, a limiting factor for future application is posed by inconsistencies in biological phenotype annotation across studies. In many cases, it is hard to obtain consistent annotation on the samples used that would allow to form comparable groups for classification analysis. This is due to lack of ontologies for description, or the use of categories that are based on subjective evaluation such as histological grading or borderline expression of a molecular marker as determined by immunohistochemistry. In such respect, it would be highly desirable to introduce systems for annotation of samples that are analogous to the MIAME standard for description of technical details of hybridization [43]. Until such a system exists, one has to focus on studies where consistency can be guaranteed by expert evaluation, as is the case for the data sets investigated here.

More study results will be needed to validate our findings. Cross-platform analyses have to be conducted considering more than two studies at a time. Here, the problem of having relatively few genes in common between all studies will gain increasing importance. Methods to make use of gene expression values only available on some platform(s) but not on others will be required. For this, the adaption of a recent approach by Guo et al. [44] could be a first step. Guo et al. use functional expression profiles (FEP) instead of gene expression profiles (GEP) for their classification analysis and generate the FEP by averaging the expression levels of genes mapping to the same Gene Ontology (GO) annotation. For integrating data from different microarray platforms, mapping of such functional summary measurements as FEP rather than the actual gene expression measurements between different chip platforms might result in an increased number of measurements (in terms of the number of genes) having an influence on the analysis results. However, by this approach the amount of information in the data is also reduced, as for example anti-correlated genes mapping to the same GO annotation would countervail each other. Further research is required to evaluate the impact of these two effects on the results of an integrated cross-platform classification analysis. The general improvement of matching genes between different platforms would also be beneficial in order to avoid false or missing mappings. Such developments are under way in our laboratory.

Cross-platform classification of multiple cancer microarray data sets yields discriminative gene expression signatures that are found and validated on a large number of microarray samples, generated by different laboratories and microarray technologies. Predictive models generated by this approach are better validated than those generated on a single data set, while showing high predictive power and improved generalization performance. The results presented here for the three sample study pairs indicate that this approach bears the potential to become a widely applicable technique for inter-validation of studies performing classification of microarray gene expression data.