The Potential for the Transcriptome to Serve as a Clinical Biomarker for Cardiovascular Diseases

Statistical Analysis and Validation of a Genomic Classifier

An obvious use of microarrays is to discover in an unsupervised manner genes that are either up- or downregulated. The high-dimensionality of microarray data with examination of thousands of transcripts in a relatively small number of samples raises the potential concern of creating false-positive results. In addition to simple fold-changes which fail to account for variability of expression between groups, more advanced statistical tests based on multiple testing were established, including the concept of the false discovery rate (FDR) which offers a sensible balance between the number of true and false-positives and basically yields probability values with higher stringency.¹⁰ Given concerns over microarray methodology, there is a demand that results of key changes in transcript levels be confirmed by an independent molecular method. When the goal of microarray analysis is gene discovery, usually an independent method such as quantitative real-time RT-PCR has been used to reproduce mere fold-changes of up- or downregulated genes.

Classification is distinct from gene discovery as it serves to stratify a biological sample into predefined clinical categories based on the expression levels of component genes. Current oncology signatures range between 15 and 100 genes and rely on consistency and pattern association. In the present study by Morgun and Shulzhenko which focuses on discovering genomic classifiers rather than the discovery of differentially expressed genes, there is little use in confirming individual gene expression differences. Indeed, the overall usefulness of corroborative studies for individual genes has little value in the setting where the emphasis is put on “pattern” recognition.¹¹ Rather, the validation comes from testing the accuracy of the molecular biomarker to perform as a diagnostic tool in new patient populations from other centers, referred to as “external validation”.¹² For this purpose, Morgun and colleagues applied their classifier to independent datasets, for which microarray data were publicly available and found that their 98-classifier gene set achieved similarly high prediction accuracy in detecting rejection in three studies examining tissue samples from kidney transplants and bronchoalveolar lavage samples from lung transplants. As reproducibility of microarray data remains a critical issue, the good performance of this classifier in independent datasets of different organ systems seems remarkable. Not only does it strengthen the validity of the original dataset, but it also suggests similarities of rejection processes across different organs.

Interestingly, when the classifier was applied to a fourth large microarray dataset consisting of nearly 300 blood samples of cardiac allograft recipients, no significant predictions were found. Of note, this study itself failed to identify any classifier for the first year after the transplantation based on expression patterns of peripheral blood mononuclear cells.¹³ Clearly, future studies are needed to correlate gene expression patterns in myocardium to tissue specimens more readily available from individual patients and suitable as myocardial surrogates (possibly peripheral blood leukocytes). Complying with MIAME criteria, Morgun and Shulzhenko deposited their raw data files in a public database, thereby allowing for direct comparison to future studies.

Biological Interpretation

Biological interpretation of the classifier might not always be straightforward, as the classifier might contain many genes previously not associated with rejection or infection. In fact, the lack of biological interpretable information in the classifier should not imply that the genomic results are not valid. It is not the biological plausibility that distinguishes an excellent classifier, but its robustness of classification accuracy with independent data.

However, determining biological plausibility certainly increases the value of a given microarray dataset. In this sense, Morgun and Shulzhenko extended their analysis beyond class prediction and identified biological pathways in rejection and Chagas disease. As expected, transcripts encoding for immune processes were upregulated. An intriguing new finding, however, was that energy-related transcripts were predominantly downregulated in rejection and T. cruzi infection. This suggests that energy metabolism processes are depressed beyond the level previously encountered in a rat model of acute cardiac allograft rejection¹⁴ and highlights the power of an unsupervised approach to identify new pathophysiological molecular pathways which merit further investigation.

Measurement Platform

Microarray measurements can be greatly influenced by extraneous factors. Even though it is noteworthy that 20 of the 44 prospectively collected tissue samples in the study by Morgun et al were hybridized at a different institution on a different chip batch than the samples from the training set, implying a certain robustness of their classifier, reproducibility of microarray data across different laboratories raises potential challenges.¹⁵ The acceptance of MSA into clinical practice could be enhanced by conversion of the classifier to a platform which is more readily available and already validated in clinical diagnostic testing, eg, automated real-time RT-PCR. In this sense, it is important to mention that Morgun and Shulzhenko identified a subset of their 98 gene classifier for detecting rejection which consists of 14 genes and which had a similarly high prediction accuracy to that of the original classifier. Testing of stringent classifiers in a prospective way in a large number of patients from different centers will be important for translation of genomic research into clinical practice.

The study by Morgun and colleagues represents an important contribution toward achieving this goal and is apt to stimulate interest in future studies of transcriptome analysis in cardiovascular medicine. As evident by the increasing importance of transcriptome-based MSA in oncology, high throughput genomic technologies hold great promise to lead to individualization of pharmacological treatment and assessment of prognosis in clinical practice in general (personalized medicine), potentially providing new tools in the assessment and management of common cardiovascular syndromes such as atrial fibrillation and congestive heart failure.⁹