Not all patients respond equally to cancer therapeutic compounds. Recent advances in high-throughput genomic, transcriptomic, and proteomic technologies with the ever-increasing understanding of the molecular mechanisms of cancers permit uncovering genes that harbor personal variations in clinical outcomes or drug responses. Personalized medicine has revolutionized the healthcare paradigm by integrating personal genetic information, improving the drug treatment efficacy, shifting the practice of medicine, and creating opportunities to introduce new business and healthcare economic models.

The traditional standard "one-dose-fits-all" approach to drug development and clinical therapy has been ineffective, as it incurs all risks of subsequent drug toxicities and treatment failures [1]. The percentage of patients for whom a major drug is effective is presented in Figure 1 [1]. With the great variability across diseases, 38% to 75% of patients fail to respond to a treatment. The average response rate of a cancer drug is the lowest at 25%.

Adverse drug reactions as a consequence of treatment are more of a problem. Among drugs approved in the U.S., 16% have shown adverse drug reactions [1]. A frequently cited meta-analysis revealed that 6.7% of all hospitalized patients are associated with adverse drug reactions in the U.S. and that the number of deaths exceeds 100,000 cases annually [2]. A study conducted in a major hospital identified 2,227 cases of adverse drug effects among hospitalized patients and reported that 50% of these cases are likely to be related to genetic factors [3].

Personalized medicine is the ability to segment heterogeneous subsets of patients whose response to a therapeutic intervention within each subset is homogeneous [4]. Under this new healthcare paradigm, physicians can make optimal choices to maximize the likelihood of effective treatment and simultaneously avoid the risks of adverse drug reactions; scientists can improve the drug discovery process, and pharmaceutical companies can manufacture medical devices to forecast patient prognosis, facilitating early disease detection.

The ultimate goal of personalized medicine is to furnish the proper treatment to the right person at the right time [5]. The potential impact of personalized medicine is contingent upon a systematic discovery of a novel biomarker from genome-wide candidates that account for variations across individuals. This review begins with an overview of personalized medicine and illustrates the most encountered statistical approaches for uncovering biomarkers utilized in the recent literature.

Personalized medicine has been defined in many ways. According to the U.S. National Institutes of Health (NIH), personalized medicine is "an emerging practice of medicine that uses an individual's genetic profile to guide decisions made in regard to the prevention, diagnosis, and treatment of disease" [6]. The U.S. Food and Drug Administration defined personalized medicine as "the best medical outcomes by choosing treatments that work well with a person's genomic profile or with certain characteristics in the person's blood proteins or cell surface proteins" [7]. The President's Council of Advisors on Science and Technology (PCAST) described personalized medicine as "tailoring of medical treatment to the individual characteristics of each patient" [4].

It is important to recognize that personalized medicine does not literally mean individuality. The idea of personalized medicine has often been exaggerated, as suggested in a headline in Newsweek (June 10, 2005) "Medicine Tailored Just for You." In fact, a new treatment regimen is assessed on a group of carefully selected patients but not individuals [5]. As such, PCAST reports that personalized medicine is "the ability to classify individuals into subpopulations that differ in their susceptibility to a particular disease or their response to a specific treatment" [4]. If a new treatment works effectively on a sub-patient group, a preventive intervention can then be furnished to those who will benefit, avoiding adverse drug effects and sparing expense for those who will not.

A biomarker is a reliable and accurate measurement that indicates a normal biological process, a pathogenic process, or a pharmacological response to a therapeutic intervention [8]. With this broad and general definition, biomarkers include physiological measurements such as lung function, blood pressure or electroencephalography, molecular (DNA, protein, metabolite) or cellular measures from biofluids (blood, plasma, serum, and urine), molecular, cellular or histopathological measures from solid tissue samples, and measurements from magnetic resonance imaging or computed tomography images [9].

In this review, we will concentrate on "prognostic" and "predictive" biomarkers that forecast patient outcomes. A prognostic biomarker is related with a patient's clinical outcome and can be used to select patients for an adjuvant systemic treatment irrespective of the patient response to treatment, whereas a predictive biomarker is related to the patient's response to a particular intervention.

According to a U.S. NIH Consensus Conference, "a clinical useful prognostic biomarker must be a proven independent, significant factor that is easy to determine and interpret and that has therapeutic consequences" [10]. A prognostic biomarker provides information about the patients overall cancer outcome irrespective of the therapeutic response [11]. Therefore, a prognostic biomarker can be exploited to select patients for an adjuvant systemic treatment but does not forecast the treatment response [6].

Decision making about adjuvant systemic treatment for breast cancer is usually based on nodal status [12-14], tumor size [15,16], tumor type/grade [17-20], lymphatic and vascular invasion [21,22], tumor hormone receptor and human epidermal growth factor receptor 2 (HER2)/neu status [23-26], age [27,28], and ethnicity [29-31]. Prognostic biomarkers that provide better information on relapse risk could prevent many patients from chemotherapy toxicity without compromising survival [32]. Significant prognostication of a biomarker needs to be demonstrated in prospective randomized clinical trials.

In contrast, a predictive biomarker provides information about the effect of a therapeutic intervention [32]. In other words, a predictive biomarker enables screening of a subset of patients that are responsive to a specific therapy where response is defined by any of the clinical endpoints commonly measured in clinical trials [33]. As a predictive biomarker indicates heterogeneous benefits contingent upon sub-patient risk groups classified by the status of the biomarker, a significant interaction between treatment effects and patient categories needs to be statistically validated, ideally in a randomized clinical trial [34].

Predictive biomarkers can help physicians to forecast the effects of a particular treatment. Numerous proteins and genes exist that are specifically associated with breast cancer growth, proliferation, and metastasis. The deeper understanding of their roles regarding the responses of various therapies may empower physicians to determine optimal treatments for patients with breast cancer [35].

Some biomarkers are both prognostic and predictive (Table 1) [36,37]. For example, patients with estrogen receptor (ER) and/or progesterone receptor (PR)-positive tumors have longer survival than those with hormone receptor-negative tumors [15,38]. Additionally, a recent randomized trial reported that high cellular ER and PR expression predicts the benefit from adjuvant tamoxifen [39].

As another example, HER2/neu gene amplification, which leads to overexpression of its receptor on the cell membrane in approximately 30% of human breast tumors, is related with a worse prognosis in patients with node-positive breast cancer due to increased proliferation and angiogenesis and inhibition of apoptosis [23-26]. HER2/neu is also the target for the monoclonal antibody trastuzumab from which patients with HER2/neu overexpressing tumors benefit in a metastatic and adjuvant setting [40-42].

The wide-ranging impacts and myriad opportunities provided by personalized medicine can be summarized in reference to its four major attributes [5].

The critical component to success in personalized medicine is to uncover gene signatures that drive individual variability in clinical outcomes or drug responses. A number of systematic approaches have been proposed to identify molecular fingerprints that are predictive of patient prognosis and response to cancer treatments. In this review, we focused on the most encountered methods for biomarker discovery: data-driven and knowledge-driven approaches.

In the data-driven approach, biomarkers associated with tumor traits are objectively searched in genome-wide analysis using data-mining tools. Unbiased biomarker discovery is the merit of this approach. A downside is that gene signatures identified by the data-driven approach are often difficult to interpret due to limited knowledge about their biological functions. In contrast, the knowledge-driven approach attempts to select candidate genes using prior knowledge or surveying the literature for evidence of linkage to either cancer pathological processes or pathways important in drug responses. As such, genes that are unknown to be involved in a process cannot be included.

The combination of the data-driven and knowledge-driven approach has been used to develop gene signatures [48]. Biomarker discovery in genome-wide analysis is subject to the curse of dimensionality, i.e., the situation in which there are far more genomic variables than the number of samples [52]. One way to deal with this issue is to use the knowledge-driven approach to reduce the number of candidate genes detected by an objective genome-wide search.

As an illustration of the data-driven approach, recently proposed systematic data-driven approaches based on in vitro-generated predictive profiles using cell-line models entail five key technical steps: 1) data collection, 2) quality control, 3) identification of candidate gene biomarkers, 4) construction of a multivariate prediction model, and 5) independent validation of the prediction model (Figure 2) [53-57].

Biomarker discovery begins by collecting molecular data in a drug response experiment. A large amount of genomic or genetic characteristics on cell-lines are experimentally determined using high-throughput technologies. The drug's patterns of activity in cells are measured on a continuous (percent of cell survival or death) or discrete scale (responsive or resistant).

The immediate procedure following acquisition of a large amount of molecular data is quality control or pre-processing. Due to the nature of high-throughput technologies that introduce inevitable non-biologic noises and biases during data collection, appropriate normalization according to specific array technologies is performed before further analysis. It is important to note that quality control can affect downstream data analysis.

The subsequent step after assuring an adequate level of normalization is to identify the subset of genes that are candidate predictors highly associated with drug activities. This step reduces the parameter space of gene variables in a very high dimension [41]. In the previous studies, various practical approaches have been used, including classical two-sample t-tests, variant t-tests [58-61], empirical Bayes methods [62-64], a linear mixed-effect model [65], the generalized likelihood ratio test [66] and the local-pooled-error test [67]. Note that these statistical approaches rely on underlying assumptions such as distributional specifications, exchangeability for a random-effect distribution, constant coefficients of variation, a mean-variance relationship, and others.

Upon narrowing down candidate genes to a few hundred, a statistical classification modeling technique is then used to construct a multivariate prediction model. Single biomarkers are less likely to furnish sufficient sensitivity and specificity for most applications [35]. Several classification methods have been utilized, including a variant of linear discriminant analysis [68], support vector machines [69-71], Bayesian regression [72], partial least squares [73], principal component regression [74], and between-group analysis [75]. The performance of a statistical prediction model should be tested and assessed by various statistical measures such as classification error rate and area under the receiver operating characteristic curve, the product of posterior classification probabilities [76-78], and an index so-called the misclassification-penalized posterior [79]. The leave-one-out approach, random splitting, and bootstrapping are often employed for an internal cross validation. Additionally, multicenter validation is also performed for an external cross validation. It has been implied from previous studies that no one dominating classifier outperforms all other methods.

Finally, the ultimate evidence of the usefulness of a prediction model in a clinical setting is randomized, prospective validation in a clinical trial [80]. After refinement and validation in independent cohorts, the covariates in the prediction model can be used to develop assays that accurately predict prognosis and responses to chemotherapeutic agents, contributing to the development of "personalized medicine" for patients with cancer.

Personalized medicine is receiving a large amount of growing attention for its tremendous potential with myriad new opportunities. The ultimate promise of personalized medicine depends on the discovery of the personal genetic causes of disease. The remarkable advent of current high-throughput technologies in combination with improved knowledge of the molecular basis of malignancy provides a solid base for identifying novel molecular targets. This revolutionized paradigm in healthcare is already beginning to affect both research and clinical practice.

The use of high-throughput technologies is expected to greatly increase in the next few years as the cost of technologies will continue to drop (Figure 3) [81]. Genomic sequencing and its interpretation will have to be further developed and standardized for routine clinical practice to develop efficient and effective methods for discovering and verifying new biomarkers and enabling personalized medicine technologies. In particular, efforts to standardize existing technologies will lead to more reproducible and robust identification of biomarkers.

Several challenges must be overcome before this flood of profile data is successfully translated into clinical utilities for patients with breast cancer. Improved knowledge obtained using advanced profile technologies will not be sufficient for this purpose, but all stakeholders involved in personalized medicine should work together to take responsibility. Regulatory authorities should provide clear guidelines for evaluating and approving newly developed personalized drugs and should validate the capabilities of the diagnostic devices that predict patient prognoses or drug responses. Medical educational institutions should prepare the next generation of physicians to use and interpret personal genetic information appropriately and responsibly. Finally, public and private insurers need to evaluate the clinical and economic utility of personalized drugs and devices to facilitate reimbursement.