Triple‐negative breast cancer: Present challenges and new perspectives

2.1. Incidence, recurrence and outcome

According to current estimates, TNBC accounts for 10–17% of all breast carcinomas, depending on thresholds used to define ER and PR positivity and HER2 overexpression (Reis‐Filho and Tutt, 2008). In different series and patient populations TNBC may range 6–28% of breast cancers (Haffty et al., 2006; Rakha et al., 2007; Dent et al., 2007; Kwan et al., 2009), but even higher incidence rates are reported for some ethnical groups such as African Americans and for younger patients (Stead et al., 2009; Trivers et al., 2009; Lund et al., 2009; Morris et al., 2007; Carey et al., 2006).

Despite its relatively small proportion among all breast cancers, TNBC is responsible for a relatively large proportion of breast cancer deaths, due to its generally aggressive clinical course. In a retrospective study on a cohort of 1601 patients with breast cancer (Dent et al., 2007), a subgroup of 180 women with TNBC had a significantly lower mean age at diagnosis (P < 0.0001), increased likelihood of distant recurrence (hazard ratio vs other breast cancer phenotypes equal to 2.6; 95% confidence interval (CI) 2.0–3.5; P < 0.0001) and death (hazard ratio 3.2; 95% CI 2.3–4.5; P < 0.001) within five years of diagnosis, but not thereafter. Also, patients with TNBC were more likely to present with larger mean tumor size (P < 0.0001) and histological tumor grade (P < 0.0001). Although data on differential lymph node spread in TNBC and other breast cancer subgroups are still conflicting (Dent et al., 2007; Reis‐Filho and Tutt, 2008; Rakha et al., 2008), an interesting feature was the substantial lack of correlation between nodal metastasis and tumor size among women with tumors smaller than 50 mm (Dent et al., 2007) (Fig. 1A). A similar trend was also found in BRCA1‐associated tumors (Fig. 1B), as discussed in Section 3.2 (Foulkes et al., 2003a). The patterns of recurrence in TNBC were qualitatively different from the non‐TNBC group. In the former, the risk of distant recurrence peaked at approximately three years and declined rapidly thereafter, whereas in the other breast cancer types the recurrence risk seemed to be constant over time (Dent et al., 2007). TNBC has a preference for visceral metastases (Liedtke et al., 2008).

It should be emphasized that TNBC currently includes a heterogeneous group of tumors. Already by simple morphology, a group of patients with TNBC can be identified who have a more favourable outcome, for example patients with invasive adenoid cystic, apocrine and typical medullary tumors (Japaze et al., 2005; Azoulay et al., 2005; Orlando et al., 2005). Even within the relatively homogeneous group of patients with triple‐negative invasive ductal carcinoma, patients with higher or lower risk may be identified, based on specific molecular markers (Viale et al., 2009).

2.2. Current cytotoxic treatment options

Several studies support the notion that primary TNBC is a chemosensitive disease. However, although TNBC is associated with enhanced pathological complete response to neoadjuvant chemotherapy (von Minckwitz et al., 2008), these cancers show worse survival due to higher relapse among those with residual disease after chemotherapy (Liedtke et al., 2008; Carey et al., 2007). In fact, despite the high sensitivity to chemotherapy, TNBC patients with truly chemosensitive disease still represent a minority among all TNBC patients. In the metastatic setting, patients progress quickly on first‐, second‐, and third‐line palliative treatment (Kassam et al., 2009).

As triple‐negative disease is often characterized by an impaired DNA repair process, cytotoxic agents inducing DNA damage may be of specific value. Recent small and non‐randomized studies have shown promising results with cisplatinum, both in the neoadjuvant and metastatic setting (Byrski et al., 2009; Sirohi et al., 2008), although the clinical response to paclitaxel and cyclophosphamide may also be high (Rouzier et al., 2005). A new approach under clinical trial includes the use of ixabepilone, a microtubule inhibitor (Thomas et al., 2007; Baselga et al., 2009b).

Larger prospective clinical trials are needed to determine the optimal cytotoxic treatment for TNBC.

3.1. The relationship between TNBC and basal‐like breast cancer

Molecular profiling of human breast cancers by gene expression assays provided in the last decade new grounds to a clearer understanding of the heterogeneous nature of these tumors with promising trends for outcome prediction and development of individualized therapies (van't Veer et al., 2002; van de Vijver et al., 2002; Perou et al., 2000; Sørlie et al., 2003; Chang et al., 2005). The gene expression based classification proposed by Perou and Sørlie ten years ago originally defined six subtypes (Perou et al., 2000; Sørlie et al., 2003). Three of these were characterized by expression of ER and luminal epithelial cell related genes (Luminal A, Luminal B and Luminal C) while the remaining three groups, basal‐like, ErbB2+ and normal‐like, showed an expression phenotype more similar to myoepithelial/basal epithelial cells. The ErbB2+ tumors were in particular characterized by high expression of ErbB2 and genes located adjacent to the ErbB2 locus and the normal‐like subgroup showed expression patterns similar to normal breast tissue samples. Although this seminal work was based on analyses on neoadjuvantly treated breast carcinomas, the main findings have since been validated in numerous independent cohorts and these intrinsic subtypes show different mutation patterns, prognosis and routes of progression (Sørlie et al., 2003; Calza et al., 2006; Hu et al., 2006; Langerod et al., 2007; Naume et al., 2007; Smid et al., 2008; Parker et al., 2009). Later, using array comparative genomic hybridization (aCGH), several investigators have found that genomic alterations seem to be more frequent in some of the intrinsic subclasses (Chin et al., 2006; Fridlyand et al., 2006; Bergamaschi et al., 2006). Of particular interest to this review, basal‐like tumors frequently have complex rearrangements with higher numbers of gains and losses compared to luminal subtypes (Chin et al., 2006; Fridlyand et al., 2006; Bergamaschi et al., 2006), although there is evidence of a subgroup of basal‐like tumors with low genomic instability (Chin et al., 2007). Such heterogeneity is recognized by gene expression classification as well (Kreike et al., 2007) and multiclonal basal‐like tumors have been described (Navin et al., 2010). Recently a distinct type of rearrangements dominated by genomewide duplications was identified in a selection of basal‐like tumors and cell‐lines (Stephens et al., 2009), suggesting that these types of tumors represent a distinct type of breast carcinomas with a unique path of progression (Navin et al., 2010; Dalgin et al., 2007). In addition, the emerging knowledge of a cellular hierarchy in the breast supports the theory that basal‐like tumors may have a distinct etiology (Villadsen et al., 2007; reviews in: Sims et al., 2007; Polyak, 2007).

Immunohistochemistry (IHC) is frequently used to explore the distribution of the molecular subtypes by using formalin‐fixed, paraffin‐embedded tissues from larger cohorts of breast cancer patients. The ultimate selection of surrogate markers is an ongoing debate and a consensus for an appropriate panel still has to be reached (Rakha et al., 2008). Triple negativity is often used to identify basal‐like tumors (Kreike et al., 2007) although a supplement of additional markers has superior prognostic value (Cheang et al., 2008; Nielsen et al., 2004). Triple negativity as a selection criterion is highly sensitive of basal‐like tumors, but not specific as both luminal, ErbB2+ and normal‐like tumors (as identified by gene expression) can be steroid receptor negative and HER2 non‐amplified (Naume et al., 2007). The heterogeneity of TNBC is acknowledged and includes both basal‐like and non‐basal‐like tumors (Rakha et al., 2009; reviewed in Hurvitz and Finn, 2009). IHC‐based studies use different markers to define their basal‐related tumors and the lack of a systematic classification scheme makes comparison of results difficult. Acknowledging the differences between the two terms is important in clinical studies aiming to identify prognostic markers and targets for therapy for these important subgroups of breast carcinomas.

Morphologically basal‐related tumors are typically of high histological grade with a high mitotic count and they frequently exhibit geographic tumor necrosis, central scar, pushing margings and/or stromal lymphocyte enrichment (Fulford et al., 2006; Livasy et al., 2006; Rakha et al., 2006). Although most TNBCs are classified as ductal carcinomas, tumors of ‘special types’ such as medullary and adenoid cystic carcinoma fall into this category as well (Bertucci et al., 2006; Jacquemier et al., 2005; Vincent‐Salomon et al., 2007; Weigelt et al., 2008). Many of the clinical features of the basal‐like phenotype are similar to those of TNBC (see Section 2.1), including shorter relapse‐free and overall survival times compared with other types of breast cancers, a tendency toward visceral versus bone metastasis (Rodriguez‐Pinilla et al., 2006; Rakha et al., 2008), and over‐representation in BRCA1 mutation carriers (Foulkes et al., 2003b; Haupt et al., 2010).

3.2. The BRCA‐associated triple‐negative breast cancers

Following identification, mapping and cloning of the two major breast cancer predisposing genes, BRCA1 (chromosome 7q21) and BRCA2 (chromosome 13q12) (Hall et al., 1990; Miki et al., 1994; Wooster et al., 1995), increasing attention has been focused on biological and molecular characteristics of breast cancers in BRCA‐ and non‐BRCA1/2 (BRCAX) mutation carriers, in relation to carcinogenesis, disease progression and outcome (Breast Cancer Linkage Consortium, 2007, 1997, 2006, 2007, 1998, 2002, 2008, 2002, 2007, 2004, 2004, 1998).

Major clinical characteristics of breast cancer in BRCA1 mutation carriers, compared with age‐matched patients unselected for family history were envisaged to be younger age at onset (Robson et al., 2004; Cornelis et al., 1995), frequent bilateral occurrence (Hall et al., 1990; Ford et al., 1994; Easton et al., 1995), high frequency of ductal histotype cancer, although with a relative excess of medullary and atypical medullary histotypes, higher overall grade and worse histoprognostic features (Breast Cancer Linkage Consortium, 1997, 1996, 1995, 1998, 2002, 1996, 1998). Histopathological characteristics frequently detected in BRCA1‐associated breast cancers are higher mitotic counts, greater degree of nuclear polymorphism and less tubule formation. However, multifactorial analysis of histopathological differences between sporadic breast cancers and cancers involving BRCA1 and BRCA2 mutations (Lakhani et al., 1998) demonstrated that many of these features were linked with each other, the only factors independently associated with BRCA1 being high mitotic count, presence of lymphocytic infiltrate and presence of smooth noninfiltrative pushig border. A similar multifactorial analysis showed that reduction in tubule formation (together with higher overall grade) and presence of continuous pushing margins were also significantly associated with BRCA2‐related breast cancer.

Combined immunohistochemical and molecular analyses of cancer‐associated genes and encoded proteins carried out within a large collaborative study of the Breast Cancer Linkage Consortium (Breast Cancer Linkage Consortium, 1997, 1998, 2002) showed that breast cancers in patients with BRCA1 germline mutation are more often negative for ER, PR and HER2 and are more likely to be positive for p53 protein compared with controls. BRCA2 tumors did not show, instead, any significant difference in the expression of these proteins. In fact, only 10% of BRCA1‐mutated cancers showed positive staining for ER compared with 66% of BRCA2 and control tumors; positivity to PR was 21% for BRCA1‐cancers compared with 55–61% for BRCA2 and control cancers; and a low positivity (3%) was found for HER2 in both BRCA1 and BRCA2 tumors. Multiple logistic regression analysis of the immunohistochemical factors (ER‐negative, PR‐negative and HER2‐negative) and morphological features that were significant predictors of BRCA1 status indicated that the ER status was the most significant risk factor (Lakhani et al., 2002). On the other hand, the expression of ER is known to inversely correlate with tumor grade (Henderson and Patek, 1998) and is reported as one of the most important prognostic and predictive markers for breast cancer (Osborne, 1998).

Unsupervised cluster analysis of non‐BRCA1/2 breast cancer families (50 probands) demonstrated heterogeneity of BRCAX families (Honrado et al., 2007). In fact two main groups were identified, one of high grade and ER‐negative (50%) and one of low‐grade and ER‐positive tumors (50%). These two groups were in turn subdivided into five subgroups: three among the high‐grade and two among the low‐grade groups; one overexpressing HER2 (18%); one with a basal‐like phenotype (14%); one with a normal breast‐like phenotype (18%); a luminal A subgroup (36%) and a luminal B subgroup (14%).

At the molecular level, BRCA1 and BRCA2 proteins are known to be involved in DNA repair (Scully, 2000), cell cycle checkpoint control through regulation of p53 activity (Liu and Kulesz‐Martin, 2001) and maintenance of global chromosome stability (Venkitaraman, 2002).

A more complete molecular portrait has been recently consolidated for the majority of breast tumors in BRCA1‐mutated patients. This general genomic and proteomic profile typically includes lack of (or low) expression of hormone receptors, HER2 and BCL2 and overexpression of p53, EGFR and basal cytokeratines (CK 5/6) (Lacroix and Leclercq, 2005; Honrado et al., 2006), associated at clinical level with high frequency of ductal type, high proliferation rate, high histological grade and manifested lymphocyte infiltration. The reported molecular features are also characteristic to basal‐like carcinomas (Perou et al., 2000; Gorski et al., 2009), although overlap of tumor biological features between BRCA1‐associated and basal‐like cancer subtype is not complete (Dawson et al., 2009). Depletion of BRCA1 affects differentiation and enhances proliferation of mammary epithelial cells, as also confirmed by studies on animal models (Kubista et al., 2002; Furuta et al., 2005; Bradley and Medina, 1998).

Over‐representation of the TNBC phenotype in BRCA1‐associated tumors allows a rational retrospective interpretation of a series of similarities separately reported for the clinico‐pathological features of these two breast cancer subgroups. Of particular interest in this respect are the younger age of the first breast cancer event; high histopathological grade; some common morphological features (e.g. smooth noninfiltrative pushing border); and the above mentioned lack of correlation between tumor size and nodal status (Fig. 1).

Tumors arising from BRCA1 and BRCA2 mutation carriers appear to have specific pathological and gene expression profiles (Honrado et al., 2006), while BRCAX tumors are increasingly believed to originate from multiple distinct genetic events (Lacroix and Leclercq, 2005). The use of microarray technology in the evaluation of the immunophenotypic features of hereditary breast cancer (Palacios et al., 2003) revealed distinct characteristics in BRCA1 and in non‐BRCA1/2 tumors, whereas BRCA2 tumors presented intermediate patterns. No cases of HER2 amplification and/or overexpression were found by the authors, except in sporadic breast cancers.

The BRCA1 tumor suppressor gene and the HER2 oncogene are located in close proximity on the long arm of chromosome 17 (17q11–21), but the aggressive pathological features of BRCA1‐associated tumors appear unrelated to amplification of the adjacent HER2 oncogene (Grushko et al., 2002).

The molecular mechanisms responsible for tissue–specific carcinogenesis in BRCA1‐associated breast cancer are still to be clarified. The existence of common genomic features in the basal‐like phenotype of BRCA1‐associated breast cancers and that of normal breast stem cells recently suggested that BRCA1 may act as a human mammary stem cell fate regulator (Foulkes, 2004; Liu et al., 2008). In particular, the BRCA1 expression was found to be required for the differentiation of ER‐negative stem/progenitor cells to ER‐positive tumoral cells and was proposed to result in the accumulation of genetically unstable breast stem cells, providing prime targets for further carcinogenic events (Liu et al., 2008). On the other hand, the basal‐like molecular subtype can also be found in sporadic, BRCA2‐ and BRCAX‐associated cancers, although with a much lower frequency (10–20% compared with up to 90% in BRCA1‐classified cancers). A critical role of the BRCA1 protein has also been shown in the development of these basal‐like carcinomas (Melchor and Benitez, 2008 and ref. therein).

Attention has been focused on the possible molecular mechanisms underlying the very low or inexistent overexpression (and gene amplification) of HER2 (0–3%) in BRCA1/2‐associated breast cancers (Lakhani et al., 2002; Palacios et al., 2003; Grushko et al., 2002) compared with that (15–20%) in BRCAX‐associated tumors (Honrado et al., 2007). A possible co‐deletion of HER2 and BRCA1 loci (Johannsson et al., 1997) would not explain the lack of HER2 overexpression in BRCA2‐associated cancers. It has been postulated (Melchor and Benitez, 2008) that the defects in the DNA repair system associated with deleterious BRCA1 mutations may not be compatible with the proliferation stress of the aberrant signaling cascade triggered by HER2 tyrosine phosphorylation and therefore, under these conditions, HER2‐overexpressing cancer cells are unlikely to survive.

5.1. Concepts and tools

Classification of tumors by molecular profiles and increasing knowledge on altered regulation of gene expression at the translational, transcriptional and epigenetic levels (Fig. 2) allowed in the last decade the identification of possible novel markers and targets for pathway‐driven therapeutics in tumors (Cleator et al., 2007; Tan and Swain, 2008; Dowsett and Dunbier, 2008; Swanton and Caldas, 2009; Bouchalova et al., 2009; Goldhirsch et al., 2009; Bosch et al., 2010).

Accumulation of defects in cellular regulation mechanisms is the underlying cause for disease heterogeneity and differential response to treatment, also within a subgroup like TNBC. This can be at the level of gene amplification, methylation, expression, mutations, or other as yet unknown regulation mechanisms. Activation of an oncogene can increase activity in downstream pathways without necessarily requiring overexpression of proteins in these pathways. Beyond the wide diversity in the combination of overexpressed pathways, the basal‐like subtype is reported to be associated with increased activity of the HER1, RAS, CTNNB1, TP53 and E2F3 pathways (Bild et al., 2009). Other notable characteristics of TNBC include a high level of proliferative genes, including Ki‐67, basal cytokeratins such as CK5 and CK17, caveolin‐1, alpha B‐crystallin, frequent p53 mutations, Retinoblastoma (Rb) pathway inactivation, high c‐kit expression, reduced DNA repair capability, increased angiogenesis and TGF‐regulated genes (Schneider et al., 2008).

The above mentioned TNBC biomarkers and characteristics are known to have relevant implications on cell signaling pathways, tumor metabolism, evasion of cell‐cycle control, invasion and metastasis, as summarized below (Table 1).

Understanding the relations between these potential markers may help to appreciate the heterogeneity of these tumors and improve the response prediction to single agents and combination treatments.

The multiplicity of ongoing efforts to identify, characterize, and validate new biomarkers and therapy targets for TNBC (Bosch et al., 2010) shows the complexity of such a goal and the need for well designed and rigorous studies, at all levels of investigation, to resolve conflicting results, which halt or slow down the advancement of this field. Systems biology can be an important tool for elucidating crosstalk between pathways and understanding temporal effects of tumor behaviour.

5.2. Targetable pathways and cellular processes

Proliferation and death are uncommon features for non‐tumorigenic, healthy cells and therefore these events are under tight control at many levels. Fluxes through pathways are intricately balanced by stimuli of different nature and checked by multiple feedback control mechanisms at cellular, tissue, organ and organism level. Backup mechanisms to assure proper functioning in case the first level of control goes awry are present in normal cells. Tumor cells have accumulated defects that allow them to bypass the normal control mechanisms that check proliferation. For tumor cells to grow and proliferate, they must evade checks on several cellular controls. The increased energy requirement of tumor cells causes metabolic stress at the level of nutrient and oxygen supply. The accumulation of gene amplifications, translocations and mutations must evade cellular control on DNA damage during cell cycle checkpoints. Whereas most cells can only survive with close contacts to their neighbours, metastatic tumor cells must overcome the apoptosis directed controls associated with anoikis. Within one tumor, large differences exist in local environment that may cause cells at the edge to proliferate, while cells that are deprived of nutrients and oxygen may undergo necrosis, therewith exposing their neighbours to stressful conditions. Histopathology can make this cellular diversity visible, but molecular techniques like gene expression profiling and proteomics determine the average of all these microenvironments in a selected tissue sample.

5.3. TNBC biomarkers for assessment of prognosis and therapy targeting

Parameters like the proliferating index, ploidity, presence of p53, cytokeratins, HER1, and numerous other molecular alterations, may also be useful for prognostic evaluation, for predicting therapeutic response and for guiding patient management.

Given the poor outcome of TNBC despite cytotoxic treatment, alternative approaches are currently explored for possible application at the clinical level (Table 1).

6.1. Multivariate analysis of multi‐modal, multi‐dimensional ‘omics’ data

In the search for biomarkers for breast cancer diagnosis and recognition of subtypes, such as TNBC, for predictive markers of response to treatment, and in the attempt to understand the biological processes/pathways involved in breast cancer (which could also lead to new biomarkers and therapy targets), high dimensional data are frequently used. These data result from the different modern ‘omics’ related analyses (e.g. microarrays in genomics, mass spectrometry combined with some kind of separation, such as liquid chromatography (LC), proteomics or NMR‐based metabolomics). Multivariate data analysis techniques are especially suited for this kind of high and multi‐dimensional data sets, in order to improve medical knowledge discovery and integrate various sources of biomedical information (Fig. 3). However, analysis and integration of (bio)medical data remain a challenging task because of the highly complex nature and the large diversity of the data. Therefore, these approaches, which include data modeling, data visualization and data mining, require a multidisciplinary research environment, unifying expert knowledge from the field of medicine, biochemistry, bioinformatics, biology, chemistry, machine learning and statistics. Data analysis processes and, in particular, the integration of heterogeneous biomedical information provide improved prevention, diagnosis and treatment of disease (e.g. cancer), since the technology enables to better understand the underlying biological relations. Discovery of these relations results in the development of new screening techniques and strategies in the early diagnosis of cancer, including tools for the monitoring and interpretation of disease progression to expand the possibilities and effectiveness of already existing therapies. There is an increasing interest in using biological characteristics to subcategorize cancer within a histological class to provide an improved diagnostic system. For instance, breast tumors could be classified into subtypes, having distinct differences in gene expression patterns, by means of multivariate analysis techniques (Perou et al., 1999, 2000). In addition, evidence has been reported that the incorporation of gene expression signatures into clinical risk stratification can refine prognosis, resulting in optimized therapeutic strategies for breast cancer (Acharya et al., 2008).

In the context of breast cancer all kinds of studies are performed using ‘Omic’ data. Most studies involved one type of data, mostly targeted at classification: healthy versus tumor, benign versus tumor, good prognosis versus bad prognosis, etc. Regarding for instance transcriptomic data, Chen et al. (2009) recently combined this data with prior knowledge from already known pathways in order to improve prediction (they applied supervised principal component analysis (PCA) to aggregate genes in so‐called supergenes which are mostly related to outcome). Proteomics data is widely used in breast cancer recognition. An overview has been recently presented by Gast et al. (2009), based on measurements in various matrices like serum, plasma, nipple aspirate fluid, saliva, etc., although the data are multivariate and high dimensional advanced multivariate techniques are rarely used (see e.g. Belluco et al., 2007). An additional problem was the relative low number of samples involved. Furthermore, most of the emphasis was on classification and less on identification of the peaks involved. Metabolites are, in contrast to proteins, much easier to identify. Altered metabolic activities of cancer cells have been observed in many studies. Data in these studies result from mass spectrometry, and in vivo and ex‐vivo MRS (e.g. HR‐MAS MRS of breast cancer tissues). An overview of ex‐vivo MRS studies is presented by Sitter et al. (2009). Again, also in these studies relative simple univariate statistical tests are used to differentiate benign from cancer tissues concentrating on choline related peaks. A broader overview of the applications and perspectives of metabolomics is presented by Claudino et al. (2007). Multivariate techniques (Partial Least Squares, Probabilistic Neural Networks, Support Vector Machines) are applied in e.g. Bathen et al. (2007) on HR‐MAS spectra of breast cancer tissue biopsies, in Woo et al. (2009) using endogenous steroids and nucleosides profiling from GC–MS and LC‐MS measurements of urine, and in Henneges et al. (2009) using a limited set of metabolites from LC‐MS spectra of urine samples. Combining the data and information resulting from the various ‘omics’ techniques is expected to improve prediction and classification. An example of this kind of studies is presented in Chuang et al. (2007). The authors applied a protein‐network‐based approach to analyze the expression profiles of the two cohorts of breast cancer patients previously reported by two other groups. The protein–protein interaction network information was extracted from yeast data from literature. Still mostly statistical techniques were applied to combine the information, after which logistic regression was used to classify the patient data. An improved accuracy was achieved compared to the original articles.

Although having their value in the diagnosis, sub‐classification and prognosis prediction of breast cancer, most of these studies provided just limited or partial additional insight on the complex biological processes involved in breast tumor formation, the biological differences underlying the various types of breast tumor and prognosis, and the processes influencing the response differences of the various therapies. Recent functional genomic and proteomic approaches include protein–protein, protein–DNA or other ‘component–component’ interaction mapping (interactome mapping), systematic phenotypic analyses (phenome mapping) and transcript or protein localization mapping (localizome mapping). ‘Omic’ approaches have already been applied to many biological processes, leading to large lists of genes, proteins and metabolites potentially involved in the corresponding biological processes. As stated earlier, an improvement is expected in combining the data and the results of the various single ‘omics’ approaches leading to an approach within Systems Biology called “top‐down” modeling of the biological processes. Examples of studies more or less aiming at this goal are in Pujana et al. (2007), related to breast cancer and in Ergűn et al. (2007), related to prostate cancer, whereas Nam et al. (2009) is on the border between the application of various data sources to improve classification and the use of these various sources to obtain better insight. In Nam et al. (2009), gene expression data sets of breast cancer and normal subjects, transcriptional regulation information and metabolic pathway information from appropriate databases were combined. Using statistics the authors tried to find altered metabolism (altered pathways) which then lead to a set of potential biomarkers. The latter were then tested in urine samples of various types of breast cancer patients and healthy individuals. Six biomarkers appeared to be statistically different, but the area under the receiver operating characteristic (ROC) curve improved when the biomarkers were combined using multivariate classifiers. Beside this, the candidate biomarkers were tested for potential use as indicators of cancer progression. Pujana et al. (2007) started from four known breast cancer‐associated genes and their products, BRCA1 and BRCA2 and ATM and CHEK2, called ‘reference genes/proteins’, and used data from gene transcript abundance measurements from samples of healthy human tissues or organs and three cell‐lines, to find genes which potentially are functionally associated (using correlation). Next, these potential functional associations were investigated and ranked by integration with functional associations found in largely non‐overlapping ‘Omic’ data sets by using interologous functional relationships from four species. This ultimately has led to genes/proteins which later were proven to have a functional relationship with the reference genes/proteins. Ergűn et al. (2007) used an approach called mode‐of‐action by network identification (MNI). The goal is to find transcript (genes) which are mostly inconsistent with ‘normal’ regulatory gene networks and which could be mediators of a disease. These ‘normal’ regulatory gene networks are first trained using a large set of microarray expression data from 13 projects spanning 7 cancer types by training a relative simple mathematical model. PCA was applied to reduce the massive dimensions of the data. Then using test data from three types of prostate cancer, those genes were identified which could potentially be mediators for each type. Shen et al. (2009) developed a technique which could be used in this context. They developed a latent variable model to allow multiple data types for the purpose of integrative clustering. Additional progress in Systems Biology is expected from bioinformatic and chemometric techniques, which can handle and intelligently explore and apply the complex multi‐modal multi‐dimensional ‘omics’ data.

An essential aspect of the use of multivariate analysis tools in clinical applications is setting up research studies at more than one medical center. A multi‐center approach allows researchers to address a larger population, which is decisive for a better understanding of the disease and the underlying biological relations (Garcia‐Gomez et al., 2009). Depending on the disease, the number of data is often limited. Large scale multi‐center data collection overcomes the problem of sparse data, scattered over local centers' databases and potential bias due to single‐center effects. Therefore, intelligent multivariate analysis should be applied to a centralized collection of multi‐center data or to a distributed database, where each node represents a single center. Furthermore, analysis of multi‐center data, including different types of data, requires the introduction of a common terminology, which is related to the design of ontologies. In addition, the data of the different centers have to be compatible, demanding a fixed protocol for data acquisition and acquisition conditions. In conclusion, results from the analysis of multi‐center studies can potentially unravel the biological dynamics of disease and improve healthcare.

An illustration of the use of multivariate analysis in multi‐center studies is the development of decision support tools within the context of the European projects IOTA, eTUMOUR and HealthAgents (http://www.etumour.net; http://www.healthagents.net; Van Holsbeke et al., 2007; Timmerman et al., 2005). The multi‐center project IOTA focuses on the characterization of ovarian tumors. Ultrasound images and clinical data from 1066 patients were acquired during the period 1999–2002 (Van Holsbeke et al., 2007). Mathematical models to predict the risk of malignancy were developed and were internally and externally validated on nearly 2000 patients with adnexal tumors in 20 centers throughout the world during the periods 2002–2005 and 2005–2007, respectively (Timmerman et al., 2005). Currently, the IOTA project investigates the incorporation of proteomics data in the process of pattern analysis. The multidisciplinary eTUMOUR and HealthAgents projects aimed to develop and validate decision support tools for brain cancer diagnostics by combining multi‐modal multi‐dimensional clinical data, ‘Omic’ data (genomics, metabolomics, transcriptomics) and bioimaging data (in vivo magnetic resonance spectroscopy (MRS), magnetic resonance spectroscopic imaging (MRSI), magnetic resonance imaging (MRI), which were acquired over multiple centers from all over the world and stored in a database system (Lluch‐Ariet et al., 2008). Models for tumor diagnosis were developed from these data and included artificial neural networks, support vector machines, discriminant analysis and nearest neighbour classifiers (Bishop, 1995; Duda et al., 2001; Vapnik, 2002). Furthermore, these models were included in a graphical user interface to make them available for clinicians. The existing breast cancer consortia either focus on ‘Omic’ data only (such as METAcancer (http://www.metacancer‐fp7.eu), MetaBre (http://www.metabre.org), TRANS‐BIG (http://www.breastinternationalgroup.org/Research/TRANSBIG.aspx), or on imaging data only (such as MammoGrid (http://www.cems.uwe.ac.uk/cccs/project.php?name=mammogrid) using mammographic data). This overview enlightens the need for setting up similar European‐wide or worldwide consortia for breast cancer diagnostics by combining all data, from the nano‐level (‘Omic’ data) up to the macro‐level (in vitro, in vivo) so as to further improve breast cancer diagnostics and prognosis.

6.2. The power of systems biology approaches in cancer research

Systems biology is an area that has both provided many promises to support understanding of cellular regulation and that requires new ways of producing and analyzing data. Experimental research and computational modeling of biological processes ranges from small regulatory networks to full‐scale genomic models. Depending on the specific question, they are of different value. Clinically relevant studies need models that provide testable predictions.

Network models comprise interaction information of all compounds of a type such as proteins or genes. Model analysis may reveal insight into hidden relations. A respective study has, for example, linked breast cancer susceptibility with centrosome dysfunction (Pujana et al., 2007): by combining expression profiles with functional genomic and proteomic data, 118 genes were aligned with potential functions. Among those genes is the one coding for hyaluronan‐mediated motility receptor, HMMR, and a functional association with the breast cancer‐associated gene BRCA1 was predicted. Another large‐scale study has systematically linked disorders and genes associated with the diseases (Goh et al., 2007).

Signaling pathways by which cells receive and process external information are a subject of intensive study in experimental and computational systems biology. Amongst the most carefully studied pathways are the EGFR (Samaga et al., 2009; Schoeberl et al., 2002), Wnt (Lee et al., 2003), Jak/STAT (Swameye et al., 2003), TGFβ (Zi and Klipp, 2007; Schmierer et al., 2008), and NFkB pathways. As example for the relevance of signaling pathways for breast cancer, it has been demonstrated recently that TGFβ signaling switches breast cancer cell motility (Giampieri et al., 2009). The models are frequently formulated as sets of ordinary differential equations (ODEs) describing the temporal evolution of the involved components and their activity. They serve various purposes, among them most notably (i) just to understand their architecture and the observed dynamics and (ii) to rationalize the interdependence of independently measured data such as ligand supply, phosphorylation states, protein interactions, and gene expression effects (e.g. Chen et al., 2009). Dynamic features such as the effect of positive or negative feedback (e.g., Bluthgen et al., 2009; Kholodenko, 2000), or crosstalk between different pathways (Borisov et al., 2009) have been studied extensively.

Cell cycle progression is highly regulated and failure to respond to checkpoints or external signals may be a first step to cancer. Mathematical modeling of cell cycle has initially focused on model organisms such as frog eggs or yeast (Chen et al., 2000, 2004), but also on mammalian cells (Alfieri et al., 2009). Again, these models first of all serve to test our understanding of the structure or wiring of cell cycle machinery, but they also analyze detailed questions such as the effect of mutants or of unreplicated DNA on cell cycle progression (Zwolak et al., 2009). While including more and more details of the protein machinery and other components into models, the regulation of cell cycle by nutrition, signaling, or checkpoints becomes more and more predictable based on model simulations.

Development, stem cell differentiation and cell reprogramming have been described with mathematical models on different levels. The Davidson lab has delivered gene regulation networks of Xenopus eggs (Davidson et al., 2003) with ever increasing detailedness over last years, which contain activating and inhibiting influences of developmental genes on each other and also provide a link to regulatory signaling pathways such as Wnt pathway. Translating this static representation into a dynamic model is a challenge on its own, but appears feasible with upcoming data. A computational ODE model describes the dynamics of the core network governing stemcellness and cellular differentiation (Chickarmane and Peterson, 2008).

Target prediction is an important goal of systems biology approaches. Based on sufficiently well described networks and properties of the compounds' interaction, mathematical models can be useful to predict targets for treatment and test the outcome of different target positions, treatment strengths, target combinations or temporal combination scenarios (Fitzgerald et al., 2006; Schulz et al., 2009). This field is only in its infancy, since there is still a lack of sufficiently well understood networks, however there are already very promising examples among signaling pathway models, such as the study of the ErbB network with sensitivity analysis (Schoeberl et al., 2009), which identified ErbB3 as a key node in response to ligands. Boolean modeling (assigning activity values of 0 and 1 to nodes which are updated in time based on their inputs from other nodes) of the ErbB receptor regulated G1/S transition has revealed new potential targets in case of de novo trastuzumab resistance in breast cancer.

Chronotherapy is a field that is based on the fact that living beings exhibit various rhythms, such as cell cycle, circadian or annual rhythms. Circadian rhythms have been investigated, their components identified (Ueda et al., 2005) and their dynamics described with mathematical models (Brown et al., 2008). It has turned out that cancer treatment has different effects if supplied at different times of the day (Levi and Schibler, 2007).

Taken together, different directions of systems biology combining experimental and computational analysis have already made first promising contributions to understand aspects of cellular regulation and dysfunction. Further investigations need precise and reproducible data and sensible mathematical descriptions to produce predictive and helpful models in cancer research (Aebersold et al., 2009).

7.1. Challenges and limitations

A better understanding of pathological mechanisms of TNBC onset and progression, including the still unclear association with BRCA1 mutations, and the causes of TNBC phenotypic heterogeneity is expected to improve diagnosis, prognosis, treatment and prevention of these cancers. A major limitation to a robust delineation of differences between basal‐like and triple‐negative breast cancer subtypes is the present lack of consensus on immunohistochemical assays to be used in the clinical setting, and the still unclear relationships between traditional diagnostic classifiers and genomic aberrations.

7.2. Limitations of the present therapeutic options

Larger prospective clinical trials are needed to optimize chemotherapy treatments and predict the response of different TNBC subgroups to cytotoxic agents. Although promising agents are presently being developed and tested, no targeted treatment algorithms are as yet available for TNBC at the clinical level.

7.3. Perspectives

Improved awareness of the significance of gene expression signatures combined with a holistic characterization of cell signaling pathways' aberrations, may lead to more suitable prognostic and predictive models, to more adequate algorithms of therapy targeting, and improved molecular imaging approaches for earlier and non invasive tumor response monitoring. Well designed and rigorous studies are needed to resolve still conflicting or partial results.

A possible breakthrough in the field may derive from application of robust multivariate data analyses on well programmed collections of multi‐modal, multi‐dimensional data sets provided by the simultaneous use of different ‘omics’ at pre‐clinical and clinical level, and processing of the results in the frame of a suitable Systems Biology approach.