Data Analysis: Supervised and Unsupervised Methods

Each microarray experiment generates thousands of data points and reports are written in a dense technical jargon. It is easy to feel lost when trying to make sense of it all. For this reason, it is important to clearly define certain technical terms as well as goals of microarray experiments. To understand how microarrays are used, the jargon “class” and, more specifically, “known class” must be first defined. A class refers to any characteristic shared by one group of samples but not other samples: e.g., cancer versus normal tissue, metastatic versus primary tumor, responders to cancer treatment versus nonresponders. A “known class” is any differentiating characteristic that the researcher will use to label the tumor samples under study a priori the data analysis. The two main goals of microarray studies are: 1) to identify molecular signatures associated with known classes, and 2) to discover new classes. To achieve those goals, two different approaches to data analysis are taken, the Supervised method (first goal above) and the Unsupervised method (second goal)¹ (Fig. 1). To read and understand microarray-based studies, knowledge of these different methods, will greatly help to understand the authors' hypothesis and data interpretation.

Figure 1

Analysis of microarray data to discover new tumor classification and to build gene predictors for cancer diagnosis. a) For class discovery, biopsies, which we hypothesize belong to different tumor classes (black biopsies), are analyzed by unsupervised (more...)

Supervised methods of analysis are used predominantly to identify the differences at molecular level between known classes (Class Comparison) and to diagnose or “predict” to which class a new tumor sample belongs (Class Prediction). By contrast, in unsupervised methods, the samples are not labeled to belong to different clinicopathologic classes before data analysis (i.e., “unknown class”). When the purpose of the experiment is to test the hypothesis that the samples are composed of different classes, the approach is called Class Discovery. Class Discovery attempts to identify new sub-classes of tumors in cancers where the actual classification needs more definition: for instance, when the classification does not explain the different patient survival after cancer treatment.

In Class Comparison studies, the purpose is to understand the differences in gene expression that might be responsible for the differences between compared classes of tumors and to, perhaps, find hints on the genes that might deserve further study. In cancer diagnosis, however, Class Comparison is usually incorporated into Class Prediction.

Class Prediction studies build a gene “predictor” based on the genes whose expression differs between the different classes of tumors under study. A predictor is a gene expression-based multivariate function that will use the genes identified in Class Comparison to assign new tumor sample(s) into the correct class (Fig. 1). However, this method suffers from one major limitation called over fitting (1). This means that the classification algorithm performs well on the samples from which it was built but poorly on independent samples. Therefore, the validation of the gene predictor is necessary for future clinical applications. The ideal predictor is built with a “training” group of samples and then validated on a “test” group of samples.² Moreover, samples in the test and training groups should be preferentially collected and analyzed at different time points in order to ensure independency between them and to validate the predictor in similar conditions as it might be applied in the future. An important caution should be taken into account. Since the samples are a priori classified based on the currently accepted diagnostic tests, which are neither 100% sensitive nor specific, this may decrease the accuracy of the gene predictor by including in the training set a few misclassified samples. Some good examples of this type of study are discussed below.^3-5

In Class Discovery studies, the samples are grouped depending on their global gene expression without reference to tumor type, grade or any other characteristic. It analyzes the expression of thousands of genes to try to discover new taxonomic groups within the samples (Fig. 1). As an unsupervised method, it will uncover the predominant relationship of the samples' gene signature, which not always corresponds with the potential clinically relevant relationship. Examples of this kind of studies are discussed later in the chapter.^3,5-9

Examples from the Literature

Over the last few years, many studies have focused on the classification and diagnosis of cancer using microarray technology. Here, we will discuss a few examples that show how this technique can improve on the information given by classical diagnostic methods. The first three studies are in hematological malignancies, for a more extended review on the microarray advances in this area you can read Eber and Golub,¹⁰ and the next four studies are in solid tumors.

How to Apply a Published Microarray Class Predictor to Classify New Samples

Despite great advances in discovering cancer molecular profiles, the proper application of microarray technology to routine clinical diagnostics is still unresolved. One key limitation is that an individual tumor cannot be classified independently. It needs to be compared to other samples or “standards”, whose classification is known, and which are analyzed under the same conditions as the individual tumor. For this, some points appear to be critical. First, if the predictor was created in the same lab as the sample of interest being analyzed, the sample preparation, array set up, reference sample (for two-color design), slide processing and analysis should be exactly the same as for the original set. The major limitation here might be the availability of the same reference sample. When using one-color design, it is not necessary to use reference sample for hybridization, but all other cautions are essential for the correct classification. Second, if one wants to apply predictor genes discovered in another lab, then the task is more complex. In order to obtain comparable results, usage of the originally established protocol is essential. Recently the question of interlaboratory comparability was addressed for microarray data on human tumor specimens.²² This work showed that, under similar technical conditions, a high correlation between gene expressions in repeated samples could be obtained regardless of the laboratory in which the experiments were done. However, even when using the same protocol and microarray platform, it is still necessary to analyze a set of known samples together with the unknown one/s. Furthermore, a large number of samples from several independent datasets are required to guarantee the applicability of the validated profiles.²³ Although demanding, the application of a molecular profile (previously described in one laboratory) by a second laboratory with a slightly different framework may represent an important benefit. In fact, it helps to define how general or specific to certain situation/s the profile is. For example, a set of 231 genes was described by van't Veer et al²⁴ as discriminating for prognosis in node-negative untreated breast cancer patients. However, a different laboratory found that a subset of 93 genes, out of the 231 genes forming the predictor, was valid to make the same discrimination even in a more heterogeneous population of node-positive/negative patients treated with adjuvant therapy.²⁵

Translation of microarray profiles into clinical practice is already beginning in some academic centers in the Netherlands and United States and profiles that have been validated in retrospective studies are now being applied in prospective clinical trials.²⁶

How to Select Biomarkers from a Microarray Class Predictor

Another way to explore genome-wide expression data for cancer diagnosis is to translate this information into surrogate molecular markers (Fig. 2). There are at least two important advantages for doing this. First, they can be measured by relatively cheap and widely used clinical methods such as RT-PCR, ELISA and immunohistochemistry. Second, they can be detected in serum or other body fluids permitting the establishment of noninvasive diagnostic test, which is very important especially in the cases of cancers with more difficult access for diagnostic biopsy (e.g., lung, ovary, pancreas). Usually, the biomarkers will be chosen from the list of genes that form a predictor. But genes that, when combined, were good predicting the class to which a new sample belongs are not necessary good biomarkers when used alone or combined with just a few other genes. Microarray predictors usually consist of tens or hundreds of molecules. Therefore, two of the main questions for translation of microarray classifiers into diagnostic markers are: first, which genes should be selected from microarray profiles and second, how to select the minimum number of these genes sufficient for good diagnostic classification. Although several statistical procedures were suggested for this purpose,^27,28 currently there is no consensus about the best one. Apparently, the use of multiple algorithms increases the confidence and validity of the selected genes.²⁸

Figure 2

Discovering and testing diagnostic markers. Markers can be selected from the gene predictor profile illustrated in (Fig. 1). Selected markers are measured by quantitative RT-PCR in a test set of samples to check assignments of samples to the classes. (more...)

At the moment, real time RT-PCR is the most widely used technique for validation of microarray results as well as for attempting to substitute the microarray profiles for diagnostic markers. It is important to remember that, despite its similarity to microarray measurement, RT-PCR could give slightly different results. A factor that greatly contributes to the difference is the normalization procedure, which is much more precise in microarrays than in RT-PCR. In fact, microarrays generally use global normalization including all genes expressed by the sample (usually a few thousands) since the majority of the genes don't show significant expression variation across all the samples. Consequently, the normalization is not influenced as much (as it is in RT-PCR) when one or a few control (or housekeeping) genes don't behave as such and show variation on their expression among the samples.

Gordon and colleagues²⁹ proposed an interesting solution in the form of a ratio-based method of samples classification that circumvents this problem. First, ratios between genes showing opposite expressions in the clinical groups of samples are calculated. Then, samples are assigned to one of the groups accordingly to the value of the ratio. The authors used this approach in two studies. In the first one,²⁹ malignant pleural mesothelioma and adenocarcinoma of the lung were differentially diagnosed by means of eight genes. Five were up-regulated and three were down-regulated in mesothelioma respect to adenocarcinoma giving, therefore, 15 ratios. Any individual ratio had at least 90% of accuracy discriminating the tumor samples and they reached 99% when 3 random ratios were combined. In the second one,³⁰ the same authors selected some genes from the published microarray data in prostate cancers,^31,32 created the optimal ratio-based test and examined it using RT-PCR in an additional cohort of cancer and normal tissue. A 3-ratio test using 4 genes was 90% accurate distinguishing normal prostate and cancer samples. Thus, the most important feature of this solution is that using gene expression ratios it is possible to avoid the selection of “right” housekeeping genes and the normalization process for assignment of samples to classes.

However, even using conventional normalization of RT-PCR results, the microarray results can be translated into RT-PCR diagnostic profiles. In fact, Lossos and colleagues³³ studied by RT-PCR 36 genes whose expression had been reported to predict survival in diffuse large-B-cell lymphoma based on microarray data. Six genes that were among the strongest predictors entered the multivariate model and were able to distinguish different survival groups.

Similarly, a diagnostic discrimination between benign and malignant esophageal tissue was proposed using the expression of the most informative genes selected from microarrays and further evaluated by RT-PCR.³⁴ In this study, logistic regression and linear discriminant analysis were applied for the selection of clinically useful gene-classifiers. Continuing in the same field of gastrointestinal oncology and following up microarray experiments, Mori et al³⁵ found highly specific markers that detected minuscule amounts of cancer cells in cytology-negative peritoneal washings by RT-PCR. Importantly, they prospectively identified a proportion of patients with minimal residual disease that could not be diagnosed and treated otherwise.

Because RT-PCR can easily and fast detect mRNA levels of considerably many genes, and it does not depend on availability of specific reagents (like antibodies), it is the most rapid translation method of microarray observations into clinical practice. ELISA and immunohistochemistry, however, detect proteins in quantitative or semi-quantitative manner. Therefore, these two methods require a more complex procedure for marker selection and validation, since not only the mRNA expression needs to be validated but also the level of the corresponding protein in tissue and or serum.

To increase a chance that marker candidates selected from large scale gene expression data will pass all rigorous validation requirements, results of microarrays from different studies could be screened. An example of a tissue marker discovered using data from several microarray datasets is alpha-methylacyl CoA racemase (AMACR).³⁶ AMACR was selected using 4 independent datasets where the gene was over-expressed in prostate cancer comparing to benign prostate tissue. During validation in an independent set of samples, the same results were obtained when measuring mRNA and protein expression using RT-PCR and immunoblot, respectively. Then immunohistochemistry on tissue array was applied to analyze a large number of samples and evaluate clinical utility of AMACR. Interestingly, AMACR immunostaining showed not only good sensitivity (97%) and specificity (100%) for prostate cancer in the whole sample population, but it performed well also in diagnostically challenging cases that needed additional expert pathological review.

Similarly, the proteins villin and moesin were found to be tissue biomarkers that successfully distinguished between colon and ovarian adenocarcinomas.³⁷ To do so, the authors measured gene expression and protein levels in tumor cell lines. As a result of these experiments and also based on antibody availability, villin and moesin were selected as candidates for colorectal adenocarcinoma and ovarian adenocarcinoma, respectively. Then, after sequence verification, and corroboration of mRNA expression using Affimetrix array, they validated the protein levels. This was done by protein lysate microarrays followed by immunohistochemistry on tissue microarray where the authors obtained high sensitivity and specificity for both markers. This and the AMACR studies represent good examples on how a multi-step approach including genomic, proteomic, and tissue array profiling, results in selection of very few but efficient diagnostic tissue markers.

From a diagnostic point of view, serum cancer markers are even more important than tissue markers because of their ease of procurement for large screenings for early cancer diagnosis. However, the search for serum markers is the more challenging. In fact, a candidate for serum marker, selected from gene expression profiles, should not only be over-expressed locally in the cancer microenvironment, but also codes for a protein that is secreted to the periphery in sufficient levels to be detected in blood. In this situation, bioinformatics tools like Gene Ontologies are helpful to choose genes with the characteristics of interest (e.g., secreted molecule) among the huge amount of differentially expressed genes.

Ovarian cancer is a good example of discovery of serum markers. This type of cancer is usually diagnosed in advanced stage, when only about 28% of the patients survive 5 years.³⁸ In a series of studies, three serum diagnostic candidates (prostatin, osteopontin and creatine kinase B) were evaluated.^37-39 By microarrays, the authors found these three proteins (among many others) over-expressed in ovarian tumor versus normal cell lines and the three markers were selected based on antibodies availability. After corroboration of the results by real time RT-PCR and immunohistochemistry, the authors screened sera from patients with ovarian cancer, benign disease, and healthy controls by ELISA. The results showed a strong association between increased levels of the markers and ovarian cancer. Perhaps future works combining all three markers and in a larger sample setting will show how useful these three markers could be in diagnostic screening of ovarian cancer.

Overall, microarray expression profiles are an excellent source of useful markers, which will allow the diagnosis of different tumors by conventional techniques in clinical laboratories.

Aknowledgements

We want to thank Brandon Reines for critically reading this chapter.