Prognostic and predictive markers

A tumor marker is valuable only if it provides information above and beyond that provided by classic clinicopathologic features. A prognostic marker is one that is associated with clinical outcome, usually irrespective of the treatment given. Examples of prognostic markers include tumor size, number of positive lymph nodes, and tumor grade. A predictive marker is one that predicts clinical benefit from a specific therapy. Examples include estrogen receptor (ER) expression (predictive of benefit from endocrine therapy) and HER2/neu overexpression (predictive of benefit from anti-HER2 directed therapies). Some predictive markers are also prognostic, particularly when the therapy predicted to be beneficial is not used (eg, ER, HER2 expression). Predictive markers are more difficult to develop and validate, but are of intrinsically greater value because they are essential in selecting patients for beneficial therapies, more difficult to identify, and fewer.

Process for development of a multiparameter assay

There are several steps in the development of a marker, and a typical roadmap is summarized in Box 1 . Steps included in the process may be broadly classified as (1) conceptualization, (2) clinical development, (3) technical development, (4) validation, and (5) application. For the purpose of maker validation, prospective trials may be performed by retrospectively evaluating samples from completed clinical trials with mature clinical outcomes. Models have been proposed for appropriate strategies to validate markers prospectively in either newly initiated clinical trials, or in completed clinical trials. A critical issue is to ensure that there is sufficient sample size to conduct training and validation studies, and in particular a sufficient number of patients with the clinical event of interest (eg, recurrence). Development of an accurate marker is largely a function of the interplay between sample size and classification difficulty. It is not uncommon to find that several statistically equally good predictors can be developed for any given classification problem. In the postdevelopment process, there is potential for the assay to be less accurate and informative as a result of bias in clinical application of the assay. For example, clinicians may be more apt to use an assay in patients with intermediate clinical features, and not to use it in those with good or poor risk clinical features.

Process for development of a multiparameter assay

There are several steps in the development of a marker, and a typical roadmap is summarized in Box 1 . Steps included in the process may be broadly classified as (1) conceptualization, (2) clinical development, (3) technical development, (4) validation, and (5) application. For the purpose of maker validation, prospective trials may be performed by retrospectively evaluating samples from completed clinical trials with mature clinical outcomes. Models have been proposed for appropriate strategies to validate markers prospectively in either newly initiated clinical trials, or in completed clinical trials. A critical issue is to ensure that there is sufficient sample size to conduct training and validation studies, and in particular a sufficient number of patients with the clinical event of interest (eg, recurrence). Development of an accurate marker is largely a function of the interplay between sample size and classification difficulty. It is not uncommon to find that several statistically equally good predictors can be developed for any given classification problem. In the postdevelopment process, there is potential for the assay to be less accurate and informative as a result of bias in clinical application of the assay. For example, clinicians may be more apt to use an assay in patients with intermediate clinical features, and not to use it in those with good or poor risk clinical features.

Methods for analyzing gene expression

There are several methods for analyzing gene expression, which have been reviewed extensively elsewhere and which are illustrated in Fig. 1 and summarized in Table 2 . Irrespective of the analysis platform used, messenger RNA (mRNA) is first extracted from the tissues of interest (see Fig. 1 A–C). Because mRNA is highly vulnerable to degradation, sample handling at this step is critical; surgical specimens should be frozen as soon as possible when used for analysis, or placed in appropriate preservative (eg, RNA Later, Qiagen, Valencia, CA, USA). Following mRNA purification, several platforms exist for gene expression profiling. The first gene expression microarray technology to enter widespread use was the 2-color microarray (see Fig. 1 D). mRNA from 2 samples (an experimental sample and a reference sample) is converted to fluorescently labeled cDNA. Each sample is labeled with a different color (red or green) and the samples are pooled at a 1:1 ratio and hybridized to the same microarray. For all microarrays, each spot on the microarray represents 1 gene and the fluorescence intensity at each spot is proportional to the expression level of that gene in the sample. In a single-color array (see Fig. 1 E), each tumor sample is labeled with the same fluorescent dye and hybridized to its own microarray. For both array types, after removal of nonhybridized material by washing, images are obtained using laser scanning, which detects the relative fluorescent intensity of the hybridized probe at each spot. Before statistical analysis, the data must be normalized (see Fig. 1 H) to compensate for variation in labeling, hybridization, and fluorescent detection, and filtered using specific criteria to reduce the likelihood of detecting noise. In the example shown (see Fig. 1 H), array 3 had a higher average signal intensity (red) than array 1, which, in turn, was higher than array 2. Mathematical correction by normalization results in each array having the same average signal intensity, thereby largely eliminating variation caused by technical issues and allowing detection of biologically relevant differences in gene expression between samples. When many samples are analyzed, it is often convenient to summarize the data in a heatmap (see Fig. 1 I). In this format, the patient samples are typically represented in columns, and genes in rows. Genes that are expressed at levels above the median are colored in red, close to the median in black, and below the median in green. Other color schemes are also commonly used (eg, blue-yellow), which may be especially helpful for individuals with red-green color blindness. By comparing the expression level of a large number of genes in each of the samples, the technique of hierarchical clustering can be used to determine which samples are most similar in gene expression. For example, in Fig. 1 I, the samples from 2 distinct clusters are indicated by the blue and purple branching in the dendrogram at the top of the heatmap. These groups may correspond to samples with different biologic properties (eg, ER-positive vs ER-negative tumors, or high-grade vs low-grade tumors) or groups with different clinical outcomes (see Fig. 1 J).

Summary of steps involved in sample preparation and analysis using high-throughput genomic technologies ( A–J ).

Two nonmicroarray-based technologies are also finding applications in this area. In quantitative reverse-transcriptase polymerase chain reaction (qRT-PCR), RNA from the tumor is converted to cDNA and arrayed in different wells of a multiwell plate, each well containing specific PCR primers for a particular gene (see Fig. 1 F). qRT-PCR analysis can then be used to rapidly and accurately quantify the expression level of each gene of interest within the sample. Unlike microarray analysis, which interrogates tens of thousands of genes, the qRT-PCR technology is more appropriate when a limited group of genes is being investigated, although this method may still allow analysis of several hundred genes simultaneously.

The recent advent of high-throughput massively parallel sequencing machines has opened up an entirely new possibility for gene expression profiling, termed RNA sequencing. It is now feasible to profile samples by direct sequencing of cDNAs derived from many millions of RNA transcripts in a given sample. In addition to providing absolute expression levels (as the precise number of transcripts can be counted), this technology also allows the identification of alternatively spliced isoforms, mutations, and novel transcripts arising from fusion genes. In the example shown (see Fig. 1 G), the blue gene is more highly expressed than the red gene, resulting in many more sequencing reads. Furthermore, mutations and fusion genes may be detected using this technology, which is not possible using the microarray platforms or qRT-PCR. In this example, the yellow gene has a mutation that is detected in the sequencing reactions, whereas the other transcript results from the fusion of 2 independent genes (green and purple) and is recognized from sequencing reads spanning the boundary between the 2 genes. Recent studies in breast, prostate, and leukemic cell lines show the potential of this approach.

Quality control of gene expression analysis methods

There are multiple sources of error in clinical application of multiparameter assays, including preanalytical, analytical, and postanalytical. Approval requires meeting specific technical requirements regarding performance, reliability, and reproducibility. The Microarray Quality Control (MAQC) project was organized by the US Food and Drug Administration (FDA) to improve current and next-generation molecular profiling technologies and foster their proper applications in discovery, development, and review of FDA-regulated products ( http://www.fda.gov/ScienceResearch/BioinformaticsTools/MicroarrayQualityControlProject/default.htm ). The effort includes multiple stakeholders, including multiple centers within the FDA and other federal agencies, major providers of microarray platforms and RNA samples, academic laboratories, and others. In MAQC-I, 2 human reference RNA samples were evaluated, and differential gene expression levels between the 2 samples were calibrated with microarrays and other technologies (eg, qRT-PCR). The resulting microarray data sets were used for assessing the precision and cross-platform/laboratory comparability of microarrays, and allowing individual laboratories to more easily identify and correct procedural failures. In MAQC-II, teams developed classifiers for 13 end points from 6 relatively large training data sets, and produced more than 18,000 models that were tested by independent and blinded validation sets generated for MAQC-II. In MAQC-III, also called sequencing quality control, the technical performance of next-generation sequencing platforms is being evaluated by generating benchmark data sets with reference samples and evaluating advantages and limitations of various bioinformatics strategies in RNA and DNA analyses.

Methods for bioinformatics analysis of gene expression

The method by which gene expression data are analyzed depends on the objectives of the analysis, which may be broadly classified as class comparison, class prediction, or class discovery, as described by Simon and colleagues. A description of the statistical analytical methods is beyond the scope of this article, but has been reviewed by others. Class comparison involves determining differences in expression profiles associated with a specific known clinical characteristic (eg, BRCA mutation-associated cancer) or outcome (eg, recurrence or organ-specific recurrence). The primary goal of this type of analysis is to find an informative set of genes and to estimate corresponding population parameters, such as the individual effect of increased expression in each gene on the probability of recurrence. Given the multiplicity of testing at the gene level, gene importance is inferred by ranking all of the genes measured on the array by statistical significance, summarized by the magnitude of the test statistic, corresponding P -value or adjusted P -value. Extending the univariate approach, modeling approaches such as linear or logistic regression may be used to adjust for other genes of interest or known clinical factors such as tumor stage, age, or treatment modality. Model building must take relationships between predictors (included in and omitted from the model) into account to produce precise estimates of effect as well as valid inferences. Statistical association does not confer predictive ability; in classification, for example, a marker exhibiting an odds ratio as high as 3.0 is a poor classification tool. Similar to, but distinct from class comparison, class prediction involves developing a gene expression–based algorithm that accurately predicts group membership of a particular sample. It is well understood that high correlation between predictors in the model does not preclude a good fit; therefore less emphasis is placed on the interpretability of the final model in favor of highly accurate predictions. To this end, the error rate or mean squared error is of primary importance in assessing model performance. Predictive ability is first assessed through internal cross-validation approaches. However, testing the model in independent and more heterogeneous data sets is vital to properly evaluate its true predictive value.

Class comparison and prediction fall into the category of top-down approaches, in which gene expression data from cohorts with known clinical outcomes are compared with genes that are associated with prognosis without any a priori biologic assumption. After this unbiased evaluation, it has become standard to test the subset of important genes from the ranked list or model for enrichment of specific molecular pathways using tools such as Ingenuity IPA (Ingenuity Systems, Redwood City, CA, USA). Alternatively, a bottom-up approach may be used, in which gene expression patterns that are associated with a specific biologic phenotype or deregulated molecular pathway are first identified and then subsequently correlated with the clinical outcome. Class discovery may also be based on unsupervised statistical clustering algorithms such as hierarchical or k-means clustering. An analysis of 4 validated gene expression signatures developed either by the class discovery or bottom-up approach (intrinsic gene set, wound-response signature) or class comparison or top-down approach (70-gene assay, 21-gene assay) showed significant agreement in their outcome predictions for individual patients, suggesting that they are tracking similar biologic phenomena despite being developed by differing methodologies. Although proliferation is the strongest parameter predicting clinical outcome in the ER-positive/HER2-negative subtype and the common denominator of most currently available prognostic gene signatures, immune response and tumor invasion are predominant molecular processes associated with prognosis in the triple negative and HER2-positive subgroups, respectively.