Use of Gene Signatures to Improve Risk Estimation in Cancer

Breast cancer is one of the most prevalent human cancers and ranks second as a cause of cancer death in women. In 2006, approximately 212 000 new cases of invasive breast cancer were diagnosed in the United States, and each year more than 40 000 women die of the disease.¹ To date, therapeutic decisions for locally advanced breast cancer are mainly guided by clinicopathological parameters, such as patient age and functional status, comorbidities, estrogen receptor status, tumor grade, tumor size, and lymph node status. This results in inaccurate risk estimation in which some patients in the early stage of disease are overtreated and experience undeserved adverse effects. In addition, there is substantial variation in outcome among patients with similar clinicopathological disease characteristics. This medical challenge calls for better understanding of this disease and refined risk estimation to improve treatment efficacy and patient quality of life.

The advent of microarrays^{2 - 3} allows for a simultaneous screen of gene expression patterns on a genome-wide scale and has become one of the most widely used technologies to study the molecular biology of human disease. The application of this methodology to cancer research has demonstrated its potential for refining cancer diagnosis and outcome prognostication at large. Breast cancer is one of the cancers for which gene expression profiling has identified distinct molecular entities associated with differential prognosis^{4 - 10} and response to cytotoxic agents.^{11 - 12}

In this issue of JAMA, Acharya and colleagues¹³ report on the use of gene expression signatures to refine risk estimation and therapeutic decision making in a multi-institutional panel of 964 breast cancer cases. This is one of the largest studies in human cancer showing the ability of gene expression profiles to improve risk stratification beyond established risk assessment algorithms that take into account clinicopathological variables. Distinct from other breast cancer microarray studies, the study by Acharya et al did not develop another novel gene signature. Instead, by combining gene expression data for a set of oncogenic and tumor microenvironment–related genes previously identified to be associated with poor prognosis in breast cancer, the investigators showed that molecular profiles can be useful to refine risk estimation in 3 risk subgroups (low, intermediate, and high), which they have defined deliberately based on established clinicopathological variables, by using the Adjuvant! Online scoring tool in an initial training set of 573 patients. Subsets of patients with comparably poor and good prognosis were identified for each clinicopathological subgroup. The prognostic power of the gene signature was subsequently validated in an independent test set of 391 patients.

Close examination of expression patterns for each individual gene does not reveal a clear distinction between the various prognostic subgroups. However, as stated by the authors, it is the aggregation of the gene signature that provides the prognostic power. This observation is consistent with most other cancer microarray studies, which have commonly used so-called metagenes—patterns of gene expression—for risk classification.¹⁰ Although such metagenes lend support to the notion that in a complex human disease, it is not a single gene but rather multiple genes that account for the disease process, the risk associations of metagenes must be considered purely correlative and thus only of predictive value. Metagenes identify gene subsets without considering underlying mechanisms; therefore, there is low potential to aid biological interpretation. Molecular predictor models that place genes into pathways and networks allow moving from association to mechanism. The mechanistic insights that are gained are crucial for novel molecularly targeted therapies. In contrast with other microarray studies, Acharya et al have chosen to build on previous knowledge and to select a parsimonious multigene predictor that has been well characterized in breast cancer in terms of its oncogenic mechanisms. This gene set should thus represent an attractive target of future therapeutic modulation.

Although the findings are encouraging, several factors may confound the interpretation of this study. One issue is why the authors chose to apply clinicopathological risk grouping and molecular risk estimation in an iterative fashion rather than identifying a predictor model that incorporates both clinicopathological and molecular profiles up front. Risk grouping of patient populations who share similar disease characteristics groups patients into distinct classes and ignores the individual fate of each cancer disease. For example, nomograms can express the risk of individual patients relative to a combination of distinct clinical, pathological, and biological factors. Nomograms are graphical representations of statistical models that allow individualized predictions based on the characteristics of a single patient.

The authors could have used a nomogram model that incorporates all the established clinicopathological factors and the gene signature. As opposed to risk grouping, this model would use a linear scale for risk estimation and would thus assist oncologists with difficult clinical decision making. Such risk estimation is substantially easier to use and more adaptable to study tailored therapeutic options for individual patients with cancer. Besides the inherent problems of merging gene expression data sets generated by different microarray platforms,^{14 - 15} heterogeneity of patient characteristics and treatment as a result of pooling several retrospective studies hinders assessing the independent prognostic power of the gene signature.¹⁶ The robustness of the molecular risk estimation model may be challenged when considering that in contrast with the clinicopathological risk model, the authors did not formulate an algorithm to assign a risk score to each patient's metagene profile but used similarity comparison via cluster analysis in a more or less subjective fashion.

In essence, the study by Acharya et al¹³ demonstrates the potential value of using microarray-based gene signatures to refine outcome predictions. The mentioned limitations call for prospective confirmation in large independent patient populations. Such confirmation should ideally be performed using sensitive and robust methods.^{17 - 18} In an attempt to tailor risk estimation, these investigators shy away from pure metagene predictors but instead focus on genes with mechanistic implication in breast cancer. Because these genes represent potential targets for specific molecular therapy, this approach represents an advance in the changing landscape of oncology toward individualized patient management.