Introduction

Over the last decades, advances in adjuvant chemotherapy have substantially improved the treatment of poor prognosis early breast cancer (EBC). Prognostic factors currently used by clinicians for decisions of adjuvant chemotherapy have been reported during the consensus conferences of National Institute Health (NIH) and St-Gallen (NIH, 2001; Goldhirsch et al., 2003). They include clinical (age<40 years) and pathological (tumor size >10–20 mm, lymph node invasion, Scarff–Bloom–Richardson (SBR) grade II–III, no hormonal receptivity) parameters. However, despite appropriate loco-regional treatment and adjuvant systemic anthracycline-based chemotherapy, 30–50% of patients will eventually develop metastatic relapse and die (EBCTCG, 1998). Thus, current clinical and pathological parameters are not able to identify, among this overall poor-prognosis population, those patients who will be actually cured by standard therapies and those who will ultimately relapse. Such a failure in prediction is thought to be due to a relative inability to fully address the molecular heterogeneity inherent to cancer process and represents a major obstacle to a more personalized management of cancer disease. The increasing development and the expected availability of alternative therapies in the adjuvant setting makes crucial the identification of parameters that might more accurately predict the clinical outcome in individual patients after standard adjuvant chemotherapy.

High-throughput RNA expression profiling of tumor was recently demonstrated as a powerful method to enlighten cancer complexity and heterogeneity as well as decipher numerous pathways and molecular networks that may simultaneously operate in cancer diseases (Golub et al., 1999; Clark et al., 2000; Perou et al., 2000). Notably, in EBC patients, DNA microarray studies have generated transcriptional signatures that better correlate with relapse-free or overall survival than conventional prognosis criteria (Bertucci et al., 2002, 2005; van de Vijver et al., 2002; van't Veer et al., 2003). Similarly, a RT-PCR-based multigene assay was recently shown to accurately predict the probability of recurrence in tamoxifen-treated node negative breast cancer (Paik et al., 2004).

An alternative and complementary approach is to perform proteome expression analysis. Clinical proteomic profiling studies were recently boosted by development of surface-enhanced laser desorption/ionization–time of flight (SELDI–TOF) mass spectrometry (MS) which allows relatively high-throughput protein analysis of highly complex biological samples, with limited preprocessing steps. This technology combines chromatographic fractionation of the proteome using protein biochips and TOF MS analysis that can be applied to various clinical samples, such as serum (Fung et al., 2001). As blood circulates through all areas of the body, it might be modified either qualitatively or quantitatively by virtually every tissue encountered. Consequently, serum may be an appropriate surrogate tissue to investigate since it may share the fingerprint of various physiologic or pathologic processes involving cancer tissue and/or host response (Fung et al., 2005). Coupled to appropriate bioinformatic tools, SELDI–TOF MS was recently shown as a very promising method for probing serum to identify protein patterns and/or biomarkers related to various stages and types of solid tumors (Adam et al., 2002a; Petricoin et al., 2002a, 2002b; Koopmann et al., 2004; Wadsworth et al., 2004; Zhang et al., 2004), which could serve as early diagnostic markers.

Herein, we have studied early postoperative serum proteomics in high-risk primary breast cancer patients and provided the first evidence that such an approach may have a significant prognostic value.

Results

Patient characteristics

Serum from 81 high-risk EBC patients receiving adjuvant chemotherapy was subjected to protein profiling using SELDI–TOF MS technology. Clinical and pathological characteristics of patient and samples are shown in Table 1. All patients had been treated by primary surgical resection and serum samples were collected postoperatively before starting any adjuvant treatment. All patients had received adjuvant chemotherapy, mostly anthracycline-based (97%), and subsequent locoregional radiotherapy. Hormonal therapy by antiestrogen (21 patients) or antiaromatase (one patient) was administered after chemotherapy and radiotherapy when appropriate. No patients received taxane-based adjuvant treatment.

Table 1 - Patient demographics.

Full table

After a median follow-up of 86 months (range 20–115), 48 patients displayed metastatic relapse (M+) and 33 patients were long-term metastasis-free survivors MFS (M-). The 5-year MFS and OS were 45.7% [95% CI 36–57.9] and 66.8% [95% CI 57.1–78.1], respectively.

Protein profiling of serum samples

Serum samples were first fractionated using anion exchange beads. Since preliminary experiments identified fractions one, four and six as the fractions generating the largest number of resolved peaks, only those fractions were bound to CM10 and IMAC-Cu ProteinChip arrays. These six conditions (F1 CM10, F1 IMAC, F4 CM10, F4 IMAC, F6 CM10 and F6 IMAC) generated 667 protein peaks in total, ranging from 96 to 129 peaks per condition. Excel files containing absolute linear and normalized log-transformed intensity values of all serum protein resolved across the sample population are available as supplementary material (Table S1).

The intra-assay variation of each SELDI ProteinChip assay was determined by SELDI profiling of a mix of pooled serums from the study population, spotted randomly onto 12 of the 96 wells of the ProteinChip arrays along with the 81 analytical samples. The pooled coefficient of variance (pCV) for peak intensity was calculated for each experimental condition and had a mean of 22% (12–35%), in agreement with previous reports (Petricoin et al., 2002a, 2002b; Koopmann et al., 2004; Zhang et al., 2004).

Identification of a prognostic multiprotein signature

Among the six experimental conditions, 40 protein peaks presented a statistically significant differential expression (P<0.05 using Mann–Whitney test) between patients with metastatic outcome (M+) and metastasis-free surviving patients (M-) (see supplementary material Table S2). Figure 1 provides scatter plot representation of normalized log-transformed expression of various differentially expressed proteins between M+ and M- patients.

Figure 1.

Differentially expressed serum proteins according to the clinical outcome. Serum proteomic markers with differential expression between patients with metastatic outcome (M+; reds) and metastasis-free surviving patients (M-; white) patients plotted as a function of their normalized log-transformed intensities. Mean levels (m) of and coefficient of variations (Cv) are indicated for each protein.

Full figure and legend (96K)

To identify a multiprotein signature that can predict metastatic relapse, we used a two-step biostatistic process. We first reduced the dimension of the data set using the partial least-squares (PLS) method. PLS allowed us to generate three linear combinations of the 40 protein peak intensities, creating three new variables C1, C2 and C3 that had been chosen in a supervised fashion to maximize the covariance of the data set with the phenotype to discriminate, that is, metastatic outcome. Figure 2a illustrates the projection of each sample according to its new coordinates in the C1–C2, C1–C3 and C2–C3 planes. As shown, patients with metastatic relapse were then easily separated from long-term MFS.

Figure 2.

Building a multiprotein prognostic index using serum protein pattern. (a) Partial least squares (PLS)-based projection of patients according to their new C1, C2 and C ordinates. Each dot is a patient and color relates to the actual outcome (M+ patients are in red and M- patients in white). (b) Probability graph. The probability of metastatic relapse was calculated for each patient using a logistic regression-based equation of C1, C2 and C3. Each dot is a patient and color relates to the actual outcome (M+ patients are in red and M- patients in white). A probability threshold of 0.5 was chosen as cutoff to distinguish between predicted good and poor prognosis patients.

Full figure and legend (77K)

Then, using a logistic regression model, we built an equation that gave for each sample the probability of metastatic relapse, knowing C1, C2 and C3. Figure 2b shows the probability of metastatic relapse according to the multiprotein index, along with the actual outcome of patients. Samples ordered using this probability were sorted in two classes: samples with a calculated probability greater than 0.5 were assigned to the 'poor prognosis' class, while those with a calculated probability less than 0.5 were assigned to the 'good prognosis' class. This classifier predicted rather successfully clinical outcome: 42/50 patients (84%) in the 'poor-prognosis' class displayed metastatic relapse whereas 6/31 (19%) in the 'good-prognosis' class did (OR=21.87 [95% CI 6.79–70.38], P<0.0001, Fisher's exact test). Thus, our multiprotein index was able to correctly predict outcome in 83% of patients with sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of 87, 76, 84 and 81%, respectively (Table 2). Consistency and robustness of the model were verified using leave-one-out cross validation. The observed correct prediction rate after crossvalidation was 72% and sensitivity, specificity, NPV and PPV were 73, 70, 78 and 63%, respectively.

Table 2 - Correlations between the molecular grouping based on the multiprotein index and the occurrence of metastatic relapse.

Full table

Multiprotein-based classification of breast cancer samples

We next searched for correlations between our multiprotein-based classification and histoclinical features of tumors. As mentioned above, there was a strong correlation with clinical outcome. As shown in Figure 3a, the 5-year MFS were very significantly different between the two classes of patients defined by the multiprotein index. The 5-year MFS in the 'good-prognosis' class was 84% compared to 22% in the 'poor-prognosis' class (P<0.0001, log-rank test). The 5-year OS was also very largely different between these two classes (94 vs 49%; P<0.0001, log-rank test) (Figure 3b). As shown in Table 3, tumor size, hormonal receptivity and age were not significantly different between the two prognostic classes. However, in the 'poor-prognosis' class, there were significantly more patients with 4 involved lymph node and more patients with grade III tumors. Of note, the multiprotein index retained prognostic significance regardless of lymph node invasion. For example, the multiprotein signature classified the 37 patients with 0 to less than four involved axillary lymph nodes in two classes that correlated with MFS. In the 'good-prognosis' class, 1/19 patients experienced metastatic relapse as compared with 14/18 in the 'poor-prognosis' class (OR=63 [95% CI 6.31–628.34], P<0.0001, Fisher's exact test). The same was true for the 43 patients with 4 or more involved lymph nodes: the OR for metastasis was 8.4 (95% CI 1.72–40.9, P=0.0095, Fisher's exact test) among the 32 women assigned to the 'poor-prognosis' class as compared to the 11 women assigned to the 'good-prognosis' class. Interestingly, the rate of metastasis in patients with 0 or less than four involved lymph nodes assigned to the 'poor-prognosis' class of patients (14/18) was higher than in patients with four or more involved axillary lymph node assigned to the 'good-prognosis' class of patients (5/11) (OR=4.2 [95% CI 0.81–21.32]). However, this difference did not reach statistical significance (P=0.11, Fisher's exact test).

Figure 3.

Kaplan–Meier analysis of the metastasis-free survival (a) and overall survival (b) according to the serum multiprotein-based classification.

Full figure and legend (98K)

Table 3 - Clinical and pathological parameters within newly generated prognosis groups.

Full table

Uni- and multivariate analysis

We then estimated the prognostic value of conventional clinical and pathological factors in our population. As expected, in univariate analysis 5-year MFS correlated with pathological tumor size (pT1/pT2 vs pT3, P<0.0001), grade (grade I/II vs III tumors; P<0.01) and lymph node invasion (less than four positive lymph nodes vs four or more positive lymph nodes; P<0.001), while hormonal receptivity did not reach statistical significance (P=0.09). However, in a multivariate Cox regression model that included the multiprotein index, grade (I/II vs III), pathological tumor size (pT1/pT2 vs PT3), lymph node invasion (less than four vs four or more), and hormonal receptivity (yes vs no) (Table 4) the multiprotein index retained the strongest statistically significant association with MFS (HR=4.81 [95% CI 1.96–11.86], P=0.00063).

Table 4 - Cox proportional-hazards multivariate analyses in metastasis-free survival.

Full table

Serum protein identities

According to their m/z ratio, the fraction from which they were derived, the ProteinChip surface of capture, as well as data from previously performed serum profiling studies (ET Fung, unpublished data), it was possible to suggest identities for several potential protein biomarkers participating to the multiprotein index (see Table S3). These identities were further confirmed by serum immunodepletion experiments using specific antibodies.

Thus, 9192 and 81763 Da proteins, which were upregulated in serum patients with subsequent metastatic relapse were identified as Haptoglobin alpha 1 chain and transferrin, respectively, while the 8936 Da serum protein which was positively correlated to MFS was shown to be C3a complement fraction. Similarly, we identified other proteins of molecular weights 28284 and 6647 Da as apolipoprotein A1 and apoliprotein C1, respectively, with low expression was associated to metastatic relapse (data not shown).

Discussion

Although conventional anthracycline-based chemotherapy improves the outcome of primary breast cancer (EBCTCG, 1998), a large number of patients with high-risk EBC still relapse and die of the disease. Among this unfavorable group classically defined by lymph node invasion and/or large tumor size, young age, high SBR grade, negative hormonal receptivity (NIH, 2001; Goldhirsch et al., 2003), a more accurate identification of patients that will be ultimately cured with conventional adjuvant chemotherapy or that will experience metastatic relapse is critical. Such an a priori knowledge could confer a higher probability of cure with currently available therapeutic strategies to certain patients, while redirecting others toward more innovative and/or aggressive strategies.

Using SELDI–TOF MS, we have retrospectively investigated postoperative early serum proteomics from a population of 81 high-risk EBC patients receiving adjuvant chemotherapy. Serum samples collected after surgery and before any specific adjuvant treatment were subfractionated by combining anion exchange beads and retention on chromatographic ProteinChip arrays and analyzed by TOF-based MS. Proteins differentially expressed according to the metastatic outcome were selected and subjected to biostatistical analysis combining supervised PLS projection and logistic regression modeling. A 40-protein index was generated that correctly predicted the clinical outcome in 83% of patients and identified in this population 2 classes of patients ('good prognosis' and 'poor prognosis') with highly significant difference in 5-year MFS and OS. Importantly, the multiprotein index had the strongest independent prognostic value in this population when compared to conventional clinical and pathological factors with known prognostic significance, such as lymph node invasion, pathological tumor size, grade and hormonal receptivity. Some components of this multiprotein index were identified and included haptoglobin alpha 1 chain, transferrin, C3a complement fraction, apolipoprotein C1 and apolipoprotein A1.

SELDI–TOF MS profiling of serum samples has recently gained popularity as a new promising tool that can generate diagnostic biomarkers in a broad range of cancer diseases, including ovarian cancer (Petricoin et al., 2002a; Kozak et al., 2003; Zhang et al., 2004), prostate cancer (Adam et al., 2002b; Petricoin et al., 2002b), breast cancer (Vlahou et al., 2003). However, application of this technique to addressing clinical questions relating to prognosis and/or therapeutic response prediction has been limited.

In our study, the patient population was retrospectively retrieved based on the availability of appropriately stored serum samples and on a sufficient follow-up (at least 6 years for disease-free surviving patients). Patient characteristics clearly displayed high-risk features with 91% of patients having lymph node invasion with a median number of four lymph nodes involved, 45% having grade three tumors, and with a median tumor size of 25 mm. Long-term MFS and OS in our population were consistent with previously reported results in this subgroup of poor-prognosis EBC (Bonadonna et al., 1995). Although data were generated from a retrospective unicentric and small-sized sample population, the multiprotein index validity and robustness were tested using the standard leave-one-out crossvalidation method. Consequently, these results may serve as a 'proof-of-concept', supporting serum profiling studies with prognosis objectives. However and importantly, they do require further validation in an independent and larger sample set – the accrual of which is now ongoing at our institution – before entering prospective evaluation.

How can early postoperative serum be informative for long-term clinical outcome in breast cancer patients? Metastatic relapse after primary surgery of EBC can be thought as arising from clinically undetectable minimal residual disease after disruption of the complex interactions between cancer cells and tightly-regulated physiological tumor-control processes. These processes include stroma-generated growth factors, angiogenesis system and immune survey (Demicheli et al., 1997; Heimann and Hellman, 2000; Pupa et al., 2002), all being potentially affected by the surgical procedure itself (Fisher et al., 1989; Tagliabue et al., 2003). Consequently, protein peaks participating to our multiprotein index could theoretically stem from tumor cells and/or the host response, which both have the potential to modify qualitatively or quantitatively the serum proteome. For example, the significant excess of high grade tumor and patients with 4 involved lymph node in the predicted 'poor prognosis' class may indicate that the multiprotein index contains proteins related directly or indirectly to proliferation and invasion processes. In addition, since adjuvant chemotherapy was administered to all patients, this serum protein signature may also relate to the chemosensitivity along with the intrinsic tumor phenotype.

SELDI–TOF-based serum profiling studies reported to date have only identified as relevant biomarkers nonspecific host response-generated proteins, present at rather high levels, around g/ml (Diamandis, 2004). Although SELDI–TOF MS has greater analytical sensitivity than this, the presence of the abundant proteins obscures the less abundant proteins. We attempted to address this problem by prefractionating the samples using anion exchange chromatography. It is likely that other approaches to prefractionation will reveal other protein peaks. However, considering the postoperative time as a period when the tumor burden is the lowest, it is unlikely that proteins differentially expressed according to the clinical outcome may arise directly from the persistent minimal residual disease. Rather, serum multiprotein signature could correlate with postoperative host response, including growth factor-mediated cell activation, negative immune regulation and early enhancement of the angiogenic process. As suggested by Demicheli et al. (1997), those early postsurgical events may play a critical role in subsequent metastatic process, while determining the patients who can benefit from cytotoxic adjuvant treatment.

We identified some of the components of this index and observed that these proteins are well known relatively abundant host-response proteins. However, some of them may yet potentially directly impact on the metastatic process. For example, haptoglobin, an acute-phase protein mainly produced in the liver has been shown to be upregulated in the serum of patients with various solid tumors (Ahmed et al., 2004; Bharti et al., 2004; Kwak et al., 2004; Tolson et al., 2004), and has also been demonstrated to participate in angiogenesis, tissue remodeling and cell migration (Cid et al., 1993; De Kleijn et al., 2002). Similarly, a proposed role for transferrin signalling has been suggested in regulating the metastatic capacity of various solid tumors including breast cancer (Nicolson et al., 1992; Inoue et al., 1993; Cavanaugh and Nicolson, 1998; Cavanaugh et al., 1999). Transferrin was also demonstrated as promoting the angiogenic phenotype (Carlevaro et al., 1997). Additionally, early impediment of immune surveillance, as potentially reflected in our study by a decrease in activated complement components such as C3a, may favor subsequent tumor relapse. Interestingly, it was recently shown that specific matrix metalloproteinases involved in the metastatic process were able to cleave C3b complement component thereby inhibiting subsequent complement cascade activation and protecting breast cancer cells from complement-mediated injury (Rozanov et al., 2004).

Finally, posttranslational alterations of relatively abundant host-response serum proteins without clear relevance to the metastatic process could be generated by a given tumor-specific enzymatic machinery. Rather than diluted, low level tumor-specific proteins, SELDI-analysis could detect more abundant but specifically tumor-processed host-response proteins (Paradis et al., 2004; Zhang et al., 2004). These issues will benefit from purification and identification of the other contributors to the multiprotein index that are currently ongoing. Interestingly, some of the identified biomarkers in our study were found to be modulated by various cancer diseases in other serum-based proteomic studies. Thus, haptoglobin was found upregulated in serum of ovarian, oesophageal and small-cell lung cancer patients (Steel et al., 2003; Ahmed et al., 2004; Bharti et al., 2004), while complement C3 and apolipoprotein A1 (Steel et al., 2003; Zhang et al., 2004) were found downmodulated in liver or ovarian cancer. However, these studies were dedicated to the identification of serum biomarkers associated to cancer diagnosis, rather than prediction of clinical outcome.

Unfortunately, due to the retrospective nature of the study, only limited data were available regarding to the ERBB2 status of the tumor samples, preventing us from any relevant analysis of the correlation between the multiprotein index and ERBB2 overexpression. Future prospective studies addressing the prognosis value of the serum protein profile will have to include this important parameter.

Importantly, one must note that the serum multiprotein index generated in our study only addresses the outcome of very high-risk EBC patients, allowing to distinguish among an overall poor prognosis population, those patients who could be cured with conventional treatment and those who should be offered more aggressive or innovative treatments. It remains to be investigated whether or not the serum protein signature retains prognostic significance in good prognosis patients, where the main challenge is to deliver or not adjuvant chemotherapy. Furthermore, our results need to be discussed in front of the recent introduction of taxanes in the adjuvant treatment of EBC. Indeed, recent studies have clearly demonstrated a benefit in term of MFS and OS when paclitaxel or docetaxel are associated with anthracylines in chemotherapy regimen administered after primary surgery of EBC patients with lymph node invasion (Henderson et al., 2003; Nabholtz et al., 2002). However, the taxane–anthracycline combination, although effective is probably more toxic and certainly more expensive, justifying efforts to identify patients likely to be cured by conventional anthracycline-based regimen alone. In addition, the benefit of taxanes seems to be restricted to a particular subgroup with less than four lymph nodes involved, whereas no clear advantage can be shown in higher lymph node involvement. Therefore, it should remain of interest to identify among these patients those with a high probability of relapse, making them candidate to alternative strategies, potentially more effective but also more aggressive such as dose-dense-based approaches (Citron et al., 2003).

What could be the practical interests of postoperative serum proteomic profiling in high-risk EBC? First, serum may be easily collected after surgical procedure, without major technique limitations. Second, the putative prognostic information that serum proteomic patterns may share, could help to select the nature and the intensity of the postoperative systemic treatment to deliver. Finally, serum protein patterns could combine to tumor tissue-based profiling technologies (such as microarrays and/or multigene assay) in order to achieve an optimal outcome prediction, ultimately leading to highly individualized postoperative systemic treatments.

In conclusion, using a SELDI–TOF MS-based profiling approach on a high-risk EBC population receiving standard adjuvant chemotherapy, we have identified a postoperative serum proteomic profile that may predict metastatic relapse. Validation on a larger and independent population as well as full identification of the components of this multiprotein index are currently in progress. Notwithstanding the final results of this process, these data suggest a prognostic value for serum in cancer patients receiving specific treatment and should support the prospective collection of serum samples in translational studies directed to predict clinical outcome and/or therapeutic response.

Materials and methods

Samples

The study involved a retrospective series of 81 postoperative serum samples from high-risk EBC patients, collected between 1994 and 2000 at the Institut Paoli-Calmettes (Marseille, France) with available clinical, pathological and follow-up annotations. All samples were collected with institutional approval. Selection of samples was based on the following criteria: (i) EBC with delivery of adjuvant anthracycline-based chemotherapy because of a high risk of metastatic relapse defined according to the following parameters: pathological lymph node involvement or no pathological lymph node involvement but negative hormonal receptor status or pathological tumor size >20 mm or age <40 years or grade III; (ii) availability of postoperative and pre-chemotherapy serum; (iii) no relapse after a minimal follow-up of 6 years after diagnosis or metastatic relapse within 6 years. Importantly, patients with second primary cancer or local/contralateral relapse before metastatic recurrence were excluded from the sample population. Serum was obtained within 21 days after surgery and before initiation of any other anticancer treatment. All samples were processed within one hour after collection and rapidly frozen at -80°C.

Protein expression profiling

Samples were subjected to SELDI–TOF MS profiling using the ProteinChip Biomarker System as recommended by Ciphergen Biosystems (Fremont, USA). Briefly, serum samples (20 l) were first denatured and fractionated using anion exchange chromatographic beads and pH gradient elution (pH 9/flow through, 7, 5, 4, 3, organic solvent, referred as to F1, F2, F3, F4, F5 and F6). Aliquot of fractions (10 l) were bound with a randomized chip/spot allocation scheme to IMAC-Cu and CM10 ProteinChip arrays. The energy absorbing molecule (crystallization matrix) sinapinic acid was dissolved in 50% acetonitrile/0.5% trifluoroacetic acid and was promptly applied. Spotted arrays were read using the PBS IIC ProteinChip reader. All samples to be compared in a given experimental condition were processed together. For each experimental condition, arrays were read at two setting either optimized for low molecular weight (2–30) or high molecular weight (20–200) ranges.

Fractionation steps, ProteinChip array binding and matrix applications were performed using a Biomek 2000 Robot (Beckman Coulter, Fullerton CA, USA) equipped with the ProteinChip Biomarker Integration Package. Only F1, F4 and F6 fractions that had been shown in preliminary experiments as the fractions containing the largest number of peaks were subjected to analysis. A pool of randomly spotted human serum specimens was used for monitoring the intra-assay reproducibility. External mass calibration was performed daily and instrument performances were monitored weekly, using appropriate purified protein mix.

Spectra were externally calibrated, baseline subtracted, and normalized to total ion current within m/z (mass/charge) range of 1.5–150 kDa. Qualified mass peaks (signal/noise >5; cluster mass window at 0.3%) within the m/z range of 2–20 kDa (LMW) and 20–200 kDa (HMW) were selected automatically using integrated Biomarker Wizard software. The resulting Excel files containing absolute intensity and m/z ratio of protein peaks resolved were obtained and subjected to data analysis.

Analysis of proteomic data

Logarithmic transformation was applied to the peak intensity before analysis for biomarker discovery. Protein peaks resolved from each experimental condition were tested for differential expression between patients with metastatic relapse (M+) and long-term MFS (M-) using Mann–Whitney test. Differentially expressed proteins were defined as those with a P-value <0.05.

Selected proteins were then subjected to partial least squares (PLS) projection (Nguyen and Rocke, 2002a, 2002b; Antoniadis et al., 2003). PLS is a method that reduces high dimensional multivariate data (x₁, x₂, ..., x_n) by creating several linear combinations (herein C1, C2, and C3) that substitute for the original variables: (C1 (x₁, x₂, ..., x_n)=₁,0+₁,1 x₁+₁,2 x₂+..., ₁,n x_n); (C2 (x₁, x₂, ..., x_n)=₂, 0+₂,1 x₁+₂,2 x₂+..., ₂,n x_n); (C3 (x₁, x₂, ..., x_n)=₃,0+₃,1 x₁+₃, 2 x₂ +... ₃, n x_n). Coefficients were chosen in a supervised fashion to maximize covariance of C1, C2 and C3 with the phenotype to discriminate (M+ or M-). Logistic regression was applied to the new variables to build a model that calculates the probability (between 0 and 1) of displaying metastatic relapse, knowing C1, C2 and C3. A probability threshold of 0.5 was chosen as cutoff to distinguish between predicted good and poor prognosis patients. Biostatistics was performed using R software version 2.0.0.

We tested our model for consistency, robustness and validity by using the leave-one-out crossvalidation class prediction method. Briefly, one withholds a sample, builds a predictor based only on the remaining samples, and predicts the class of the withheld sample. The process is repeated for each sample and the cumulative error rate is calculated.

Statistic analysis

Distributions of molecular, pathological and clinical factors were compared using either the ² or Fisher's exact tests for categorical variables and the Wilcoxon and Mann–Whitney tests for continuous variables. MFS was calculated from the date of diagnosis to the time of metastasis as first event or time of last follow-up for censored patients. Overall survival (OS) was calculated from the date of diagnosis to the time of death as first event or time of last follow-up for censored patients. No patient died from cause other than breast cancer. Survival estimates were derived from Kaplan–Meier method (Kaplan and Meier, 1958) and compared by log-rank test. The influence of molecular grouping, adjusted for other factors, was assessed in multivariate analysis by a Cox proportional hazard model (Cox, 1972). Survival rates, odds ratio (OR) and relative risks (RR) are presented with their 95% confidence intervals (95% CI). Statistical tests were two-sided at the 5% level of significance. All statistical tests were performed using R software version 2.0.0.

Identification of protein markers by immunodepletion

Immunodepletion experiments using commercially available antibodies against selected proteins were performed. Antibodies included mouse anti-human apolipoprotein A1 (Calbiochem 178474); sheep anti-human haptoglobin (Cortex, CR2114SP); rabbit anti-human apolipoprotein C1 (Academy Bio-Medical Company, 31A-R1a); rabbit anti-human complement C3a (Cortex CR6032RP); control mouse IgG (Sigma, I5381)); rabbit anti-human transferrin (Dako 0061), control sheep IgG (Sigma, I5131); control rabbit IgG (Sigma, I5006).

Briefly, antibody was coupled to 100 l Aminolink Plus coupling gel (Pierce). In all, 20 l serum was fractionated using the EDM serum fractionation kit (Ciphergen, K100-0007). Relevant fractions, or crude serum, were diluted in binding buffer (phosphate-buffered saline (PBS) containing 0.1% triton) and incubated with antibody-coupled beads at 4°C. The beads were then washed with PBS or PBS with Triton three times, and then briefly with water. Bound material was eluted with 10 l elution buffer (33.3% acetonitrile, 16.7% isopropanol, 0.1% trifluoroacetic acid). The elutions were pooled and applied to NP20 ProteinChip arrays (Ciphergen) with sinapinic acid as matrix.

References

1. Adam BL, Qu Y, Davis JW, Ward MD, Clements MA, Cazares LH et al. (2002a). Cancer Res 62: 3609–3614. | PubMed | ISI | ChemPort |
2. Adam B-L, Qu Y, Davis JW, Ward MD, Clements MA, Cazares LH et al. (2002b). Cancer Res 62: 3609–3614. | PubMed | ISI | ChemPort |
3. Ahmed N, Barker G, Oliva KT, Hoffmann P, Riley C, Reeve S et al. (2004). Br J Cancer 91: 129–140. | Article | PubMed | ISI | ChemPort |
4. Antoniadis A, Lambert-Lacroix S, Leblanc F. (2003). Bioinformatics 19: 563–570. | Article | PubMed | ISI | ChemPort |
5. Bertucci F, Finetti P, Rougemont J, Charafe-Jauffret E, Cervera N, Tarpin C et al. (2005). Cancer Res 65: 2170–2178. | Article | PubMed | ISI | ChemPort |
6. Bertucci F, Nasser V, Granjeaud S, Eisinger F, Adelaide J, Tagett R et al. (2002). Hum Mol Genet 11: 863–872. | Article | PubMed | ISI | ChemPort |
7. Bharti A, Ma PC, Maulik G, Singh R, Khan E, Skarin AT et al. (2004). Anticancer Res 24: 1031–1038. | PubMed | ISI | ChemPort |
8. Bonadonna G, Zambetti M, Valagussa P. (1995). JAMA 273: 542–547. | Article | PubMed | ISI | ChemPort |
9. Carlevaro MF, Albini A, Ribatti D, Gentili C, Benelli R, Cermelli S et al. (1997). J Cell Biol 136: 1375–1384. | Article | PubMed | ISI | ChemPort |
10. Cavanaugh PG, Jia L, Zou Y, Nicolson GL. (1999). Breast Cancer Res Treat 56: 203–217. | Article | PubMed | ISI | ChemPort |
11. Cavanaugh PG, Nicolson GL. (1998). J Cell Physiol 174: 48–57. | Article | PubMed | ISI | ChemPort |
12. Cid MC, Grant DS, Hoffman GS, Auerbach R, Fauci AS, Kleinman HK. (1993). J Clin Invest 91: 977–985. | PubMed | ISI | ChemPort |
13. Citron ML, Berry DA, Cirrincione C, Hudis C, Winer EP, Gradishar WJ et al. (2003). J Clin Oncol 21: 1431–1439. | Article | PubMed | ISI | ChemPort |
14. Clark EA, Golub TR, Lander ES, Hynes RO. (2000). Nature 406: 532–535. | Article | PubMed | ISI | ChemPort |
15. Cox DR. (1972). J R Stat Soc [B] 34: 187–220.
16. De Kleijn DPV, Smeets MB, Kemmeren PPCW, Lim SK, van Middelaar BJ, Velema E et al. (2002). FASEB J 16: 1123–1125. | PubMed | ChemPort |
17. Demicheli R, Retsky MW, Swartzendruber DE, Bonadonna G. (1997). Ann Oncol 8: 1075–1080. | Article | PubMed | ISI | ChemPort |
18. Diamandis EP. (2004). J Natl Cancer Inst 96: 353–356. | PubMed |
19. EBCTCG (1998). Lancet 352: 930–942. | Article |
20. Fisher B, Gunduz N, Coyle J, Rudock C, Saffer E. (1989). Cancer Res 49: 1996–2001. | PubMed | ISI | ChemPort |
21. Fung ET, Thulasiraman V, Weinberger SR, Dalmasso EA. (2001). Curr Opin Biotechnol 12: 65–69. | Article | PubMed | ISI | ChemPort |
22. Fung ET, Yip TT, Lomas L, Wang Z, Yip C, Meng XY et al. (2005). Int J Cancer 115: 783–789. | Article | PubMed | ISI | ChemPort |
23. Goldhirsch A, Wood WC, Gelber RD, Coates AS, Thurlimann B, Senn H-J. (2003). J Clin Oncol 21: 3357–3365. | Article | PubMed | ISI |
24. Golub TR, Slonim DK, Tamayo P, Huard C, Gaasenbeek M, Mesirov JP et al. (1999). Science 286: 531–537. | Article | PubMed | ISI | ChemPort |
25. Heimann R, Hellman S. (2000). Eur J Cancer 36: 1631–1639. | Article | PubMed | ISI | ChemPort |
26. Henderson IC, Berry DA, Demetri GD, Cirrincione CT, Goldstein LJ, Martino S et al. (2003). J Clin Oncol 21: 976–983. | Article | PubMed | ISI | ChemPort |
27. Inoue T, Cavanaugh PG, Steck PA, Brunner N, Nicolson GL. (1993). J Cell Physiol 156: 212–217. | Article | PubMed | ISI | ChemPort |
28. Kaplan EL, Meier P. (1958). J Am Stat Assoc 53: 457–481. | ISI |
29. Koopmann J, Zhang Z, White N, Rosenzweig J, Fedarko N, Jagannath S et al. (2004). Clin Cancer Res 10: 860–868. | PubMed | ISI | ChemPort |
30. Kozak KR, Amneus MW, Pusey SM, Su F, Luong MN, Luong SA et al. (2003). Proc Natl Assoc Sci 100: 12343–12348. | Article | ChemPort |
31. Kwak J-Y, Ma T-Z, Yoo M-J, Hee Choi B, Kim H-G, Kim S-R et al. (2004). Exp Hematol 32: 836–842. | Article | PubMed | ISI | ChemPort |
32. Nabholtz JM, Pienkowski T, Mackey J, Pawlicki M, Guastalla JP, Vogel C et al. (2002). Proc Am Soc Clin Oncol 21: 36a (Abstract 141).
33. Nguyen DV, Rocke DM. (2002a). Bioinformatics 18: 1216–1226. | Article | ISI | ChemPort |
34. Nguyen DV, Rocke DM. (2002b). Bioinformatics 18: 39–50. | Article | PubMed | ISI | ChemPort |
35. Nicolson GL, Cavanaugh PG, Inoue T. (1992). J Natl Cancer Inst Monogr 13: 153–161. | PubMed |
36. NIH (2001). J Natl Cancer Inst 93: 979–989. | Article |
37. Paik S, Shak S, Tang G, Kim C, Baker J, Cronin M et al. (2004). N Engl J Med 351: 2817–2826. | Article | PubMed | ISI | ChemPort |
38. Paradis V, Degos F, Dargere D, Pham N, Belghiti J, Degott C et al. (2004). Hepatology 41: 40–47. | Article | ISI | ChemPort |
39. Perou CM, Sorlie T, Eisen MB, van de Rijn M, Jeffrey SS, Rees CA et al. (2000). Nature 406: 747–752. | Article | PubMed | ISI | ChemPort |
40. Petricoin EF, Ardekani AM, Hitt BA, Levine PJ, Fusaro VA, Steinberg SM et al. (2002a). Lancet 359: 572–577. | Article | PubMed | ISI | ChemPort |
41. Petricoin III EF, Ornstein DK, Paweletz CP, Ardekani A, Hackett PS, Hitt BA et al. (2002b). J Natl Cancer Inst 94: 1576–1578. | Article | PubMed | ChemPort |
42. Pupa SM, Menard S, Forti S, Tagliabue E. (2002). J Cell Physiol 192: 259–267. | Article | PubMed | ISI | ChemPort |
43. Rozanov DV, Savinov AY, Golubkov VS, Postnova TI, Remacle A, Tomlinson S et al. (2004). J Biol Chem 279: 46551–46557. | Article | PubMed | ISI | ChemPort |
44. Steel LF, Shumpert D, Trotter M, Seeholzer SH, Evans AA, London WT et al. (2003). Proteomics 3: 601–609. | Article | PubMed | ISI | ChemPort |
45. Tagliabue E, Agresti R, Carcangiu ML, Ghirelli C, Morelli D, Campiglio M et al. (2003). The Lancet 362: 527–533. | Article | ChemPort |
46. Tolson J, Bogumil R, Brunst E, Beck H, Elsner R, Humeny A et al. (2004). Lab Invest 84: 845–856. | Article | PubMed | ISI | ChemPort |
47. van de Vijver MJ, He YD, van't Veer LJ, Dai H, Hart AA, Voskuil DW et al. (2002). N Engl J Med 347: 1999–2009. | Article | PubMed | ISI | ChemPort |
48. van't Veer LJ, Dai H, van de Vijver MJ, He YD, Hart AA, Bernards R et al. (2003). Breast Cancer Res 5: 57–58.
49. Vlahou A, Laronga C, Wilson L, Gregory B, Fournier K, McGaughey D et al. (2003). Clin Breast Cancer 4: 203–209. | PubMed | ChemPort |
50. Wadsworth JT, Somers KD, Cazares LH, Malik G, Adam BL, Stack Jr BC et al. (2004). Clin Cancer Res 10: 1625–1632. | PubMed | ISI | ChemPort |
51. Zhang Z, Bast Jr RC, Yu Y, Li J, Sokoll LJ, Rai AJ et al. (2004). Cancer Res 64: 5882–5890. | Article | PubMed | ISI | ChemPort |

Acknowledgements

We thank Dr Françoise Birg and Dr Claude Mawas for their support. This project is funded by Institut Paoli-Calmettes and INSERM. This work was supported by la Ligue Nationale Contre le Cancer (Labels to DB and JPB) and Ministries of Health and Research (Cancéropôle PACA).

Supplementary Information accompanies the paper on Oncogene website (http://www.nature.com/onc).