Expression patterns and prognostic value of Bag-1 and Bcl-2 in breast cancer

The MDA-MB-468, MCF-7, T47D, BT-474 and SKBR3 (human breast cancer) cell lines were purchased from ATCC (Manassas, VA, USA). Western blotting of protein extracts was performed using standard methods. Bag-1 and Bcl-2 expression were detected by overnight incubation with mouse anti-Bag-1 IgG (Chemicon, Temecula, CA, USA) at 1:400 and with mouse anti-Bcl-2 IgG (Dako Corp., Carpinteria, CA, USA) at 1:6,000. Protein loading was assessed using rabbit anti-β-actin (Sigma-Aldrich, St Louis, MO, USA) at 1:5,000.

The breast cancer tissue microarrays were constructed as previously described [67]. A total of 319 node-negative and 319 node-positive breast cancer cores, each measuring 0.6 mm in diameter, were spaced 0.8 mm apart on two glass slides. The cohort was constructed from paraffin-embedded, formalin-fixed tissue blocks obtained from the Yale University Department of Pathology Archives. Specimens and clinical information were collected under the guidelines and approval of a Yale University Institutional Review Board.

By standard immunohistochemistry, ER staining was positive in 52%, PR in 46% and HER2/neu in 14% of specimens. Nuclear grade 3 (on a 1 to 3 scale) was seen in 28% of the specimens, and 59% were larger than 2 cm. The histological subtypes included 72% invasive ductal carcinoma and 1% lobular carcinoma, and 14% had mixed or other histology. The specimens were resected between 1962 and 1983, with a follow-up range between 4 months and 53 years, and a mean follow-up time of 12.6 years. Patient age at diagnosis ranged from 24 to 88 years (mean age, 58 years).

A complete treatment history was not available for the entire cohort. Most patients were treated with local irradiation. None of the node-negative patients were given adjuvant systemic therapy. A minority of the node-positive patients (approximately 15%) received chemotherapy, and approximately 5.6% of patients received tamoxifen (postmenopausal patients with ER-positive tumors, treated after 1978). The time between tumor resection and tissue fixation was not available.

A pathologist reviewed slides from all of the blocks to select representative areas of invasive tumor to be cored. The cores were placed on the tissue microarray using a Tissue Microarrayer (Beecher Instruments, Silver Spring, MD, USA). The tissue microarrays were then cut into 0.5 μm sections and were placed on glass slides using an adhesive tape-transfer system (Instrumedics, Inc., Hackensack, NJ, USA) with UV cross-linking.

Staining was performed for automated analysis of breast cancer specimens as previously described [68]. Briefly, slides were deparaffinized in xylene, and were transferred through two changes of 100% ethanol. For antigen retrieval, the slides were pressure-cooked in 6.5 mM sodium citrate (pH 6.0). Endogenous peroxidase activity was blocked in a mixture of methanol and 2.5% hydrogen peroxide for 30 minutes. To reduce nonspecific background staining, slides were incubated for 30 minutes in 0.3% bovine serum albumin/1 × Tris-buffered saline.

Slides were then incubated at 4°C overnight with the following primary antibodies: mouse monoclonal anti-Bcl-2 (clone 124; Dako) at 1:40; mouse monoclonal anti-Bag-1 (MAB4611; Chemicon) at 1:150; mouse monoclonal anti-ER (Dako) at 1:50; mouse monoclonal anti-PR at 1:50 (Dako); and rabbit anti-Her2/neu at 1:8,000 (Dako). All antibodies were diluted in Tris-buffered saline containing 0.3% bovine serum albumin. Goat anti-mouse (or anti-rabbit for Her2/neu) horseradish peroxidase-decorated polymer backbone (Envision; Dako) was used as a secondary reagent, and Cy5-tyramide (Perkin Elmer Life Science, Waltham, MA, USA) was used to visualize the target.

To create a tumor mask, primary slides were simultaneously incubated with rabbit anti-human cytokeratin antibodies diluted at 1:200. For Bcl-2, Bag-1, ER and PR, rabbit anti-cytokeratin was used. Mouse anti-cytokeratin was used for Her2/neu. The anti-cytokeratin antibodies were visualized with secondary Alexa 488-conjugated goat anti-rabbit or goat anti-mouse antibodies (Molecular Probes, Inc., Eugene, OR, USA). Coverslips were mounted with ProLong Gold antifade reagent with 4',6-diamidino-2-phenylindole (Invitrogen Corp, Grand Island, NY, USA).

Images were acquired using automated quantitative analysis (AQUA), as described previously [67]. Briefly, areas of tumor were distinguished from stroma by creating a mask with the cytokeratin signal tagged with Alexa 488. Coalescence of cytokeratin at the cell surface was used to identify the membrane/cytoplasm compartment within the tumor mask, while 4',6-diamidino-2-phenylindole (DAPI) was used to identify the nuclear compartment within the tumor mask. The target markers, Bag-1, Bcl-2, ER, PR or Her2/neu, were visualized with Cy5 (red). Multiple monochromatic, high-resolution (1,024 × 1,024 pixels, 0.5 μm) grayscale images were obtained for each histospot, using the 10 × objective of an Olympus AX-51 epifluorescence microscope (Olympus, Melville, NY, USA) with an automated microscope stage and digital image-acquisition driven by custom program and macro-based interfaces with IPLabs software (Scanalytics Inc., Fairfax, VA, USA).

Images were analyzed using algorithms that have been previously extensively described [68]. Two images (one in-focus and one out-of-focus) were taken of the compartment-specific tags and the target marker. A percentage of the out-of-focus image was subtracted from the in-focus image for each pixel, representing the signal-to-noise ratio of the image. An algorithm called the Rapid Exponential Subtraction Algorithm was used to subtract the out-of-focus information in a uniform fashion for the entire microarray. Subsequently, a second algorithm called the Pixel Locale Assignment for Compartmentalization of Expression algorithm was used to assign each pixel in the image to a specific subcellular compartment, and the signal in each location was calculated. The data were expressed as the average signal intensity per unit of compartment area on a scale of 0 to 255, and were expressed as target signal intensity relative to the compartment area.

To generate the principal component analysis biplot (presented in Figure 1), we tabulated our dataset in a matrix consisting of 439 patients (rows) and five biomarkers (columns). We only included the 439 patients (out of 638 patients) who had values for all five biomarkers. The first step in this analysis involves data preprocessing obtained by first normalizing each marker 439-dimensional profile by its median value, followed by a two-way data-centering procedure of this normalized data matrix. The centering procedure involves transforming each entry of this matrix by subtraction of the row and column means of this entry and addition of the overall matrix mean, leading to a transformed matrix having all row sums and all column sums equal to zero. Reduction in the number of variables is useful for visualizing how the patient population is distributed in the five-dimensional variable space. Principal component analysis is an unsupervised dimension reduction method that generates a new set of decorrelated variables (principal components) as linear combinations of the original variables (biomarkers). The majority of the variation associated with the Her2/neu, ER, PR, Bcl-2 and Bag-1 variables can be captured by the most dominant principal components. An additional advantage of expressing the data in terms of the leading principal components is their robustness to noise.

The projections of the samples onto the leading principal components were computed by applying the singular value decomposition to the data matrix (after preprocessing as described above and in previous works [69]). We used principal component biplots to display the biomarkers (columns of the data matrix) and the patients (rows of the data matrix) simultaneously as points in a two-dimensional space [70]. The biplot provides an optimal approximation of the data matrix by such a two-dimensional structure, in that it displays the singular value decomposition – which gives the rank-two approximation to the data matrix having the smallest mean-squared error. The expression of a given biomarker in a given patient sample is approximated by the projection of the biomarker vector onto the direction of the patient sample vector, multiplied by the length of the patient sample vector. In the present rank-two approximation, therefore, for a given biomarker and for patient vectors of a given length, the biomarker is expressed at a higher (or lower) level in patients whose vector points in nearly the same (or opposite) direction as the biomarker. A biomarker is not overexpressed or underexpressed for patients whose vector is orthogonal to the biomarker vector, and is underexpressed for patients whose vector form an obtuse angle with the biomarker vector.

Western blot analysis for Bag-1 showed bands at 36 kDa, 46 kDa and 50 kDa (Figure 2). These represent the three isoforms reported in the literature [49], and all three isoforms are recognized by the antibody. The strongest expression of Bag-1 was observed in BT-474, consistent with findings by other researchers [42]. For Bcl-2, a single 28 kDa band was seen; expression was high in the MCF-7 cell line and low in the SKBR3 cell line, as reported in the literature [26, 42, 52].

To account for intratumor heterogeneity, two separate sets of slides – each containing a core from a different area of the tumor for each patient – were used to evaluate the expression of each marker. Bcl-2 and Her2/neu did not have significant amounts of nuclear staining, and only the membranous/cytoplasmic compartments were analyzed, and vice versa, for ER and PR staining. Bag-1 staining was either nuclear or cytoplasmic, and many specimens had both nuclear and cytoplasmic staining. We performed log-regression analyses to assess the correlation between the two slides for Bag-1 and Bcl-2, as demonstrated in Figure 3. The matching spots on each array were highly correlated for all four markers (R = 0.6 for Bag-1, R = 0.7 for Bcl-2, R = 0.74 for the ER, R = 0.79 for the PR and R = 0.92 for Her2/neu). AQUA scores ranged from 8.3 to 146.02 for the total Bag-1 score (median, 24.07), from 8.37 to 259.61 for Bcl-2 (median, 32.04), from 2.34 to 94.097 for the PR (median, 6.58), from 1.085 to 61.164 for Her2/neu (median, 2.03), and from 2.29 to 105.39 for the ER (median, 13.46). Examples of Bag-1 and Bcl-2 staining are shown in Figure 4.

For each of the markers, the AQUA scores from both sets of slides were combined to give a single dataset. Tumor spots were deemed uninterpretable if they had insufficient tumor cells, loss of tissue in the spot or an abundance of necrotic tissue. For patients who had two interpretable histospots, a composite score was formed by taking the average of the two scores. For patients with only one interpretable core, the single score was used. The combined dataset for Bag-1 had scores for 574 patients. We obtained scores for Bcl-2, Her2/Neu, ER and PR for 528 patients, 608 patients, 601 patients and 595 patients, respectively.

Given that the nodal status often determines the standard clinical approach to patients, we assessed the prognostic value of Bag-1 and Bcl-2 in the entire cohort, as well as within the node-positive and node-negative subsets of patients. Using the Cox univariate survival analysis of raw AQUA scores, we found that Bag-1 expression (nuclear, cytoplasmic and total) was associated with breast cancer-specific survival in the node-positive subset only (P = 0.006 for the total Bag-1 score), whereas Bcl-2 expression was associated with survival in the entire cohort and in the node-positive subset (P = 0.008 and P = 0.002, respectively). The association with survival for Bag-1 and Bcl-2 within the node-positive subset (but not the node-negative subset) might be due to the larger number of events (deaths) within the node-positive subset.

Table 1 presents the associations between continuous scores of Bag-1, Bcl-2, ER, PR and Her2/neu and survival in the entire cohort and in the node-negative and node-positive subsets. There were no remarkable differences between nuclear Bag-1 scores, cytoplasmic Bag-1 scores and total Bag-1 scores as predictors of survival; the remainder of these analyses will therefore focus on the total Bag-1 scores.

Continuous AQUA scores were then dichotomized arbitrarily by the median score, reflecting the use of routine statistical divisions in the absence of an underlying justification for division of expression levels. Kaplan–Meier survival curves were generated for the total Bag-1 score and Bcl-2, as shown in Figure 5. The log-rank analysis revealed a statistically significant association with survival in the node-positive subset for Bcl-2 and Bag-1 (P = 0.016 for both markers), and for Bcl-2 in the entire cohort (P = 0.0042).

Using the Cox proportional hazards model, we performed multivariable analyses to assess the independent prognostic value of Bag-1 and Bcl-2 expression. Neither of these markers retained their independent prognostic value in this model, and the only markers to be independent predictors of survival were Her2/neu, tumor size and nodal status.

To assess the association between Bag-1 and Bcl-2 expression and other commonly used clinical and pathological parameters, we performed analyses of variance. No associations were found between expression of Bag-1 or Bcl-2 and nodal status, tumor size, age or nuclear grade.

For patients who had AQUA scores for all five biomarkers, we assessed the associations between biomarkers by Spearman's rho test, as demonstrated in Table 2. Bag-1 and Bcl-2 were strongly correlated with the ER, and with the PR but to a lesser degree.

Figure 1 shows a principal component biplot, in which both patient profiles and biomarker profiles are projected onto the leading principal components. As can be seen, the samples are distributed in a tree-like branch structure. The samples on the branch in the positive direction of the first principal component are associated with high Her2/neu expression levels and low to medium levels of ER, PR, Bag-1 and Bcl-2. Similarly, the samples in the upper left branch have elevated PR expression levels, low to medium ER, Bcl-2 and Bag-1 expression levels, and low Her2/neu expression levels. The samples in the lower left branch have elevated ER, Bcl-2 and Bag-1 levels, low to medium PR levels, and low Her2/neu levels. The bulk of the samples are localized next to the origin, and therefore their inner products with any of the five biomarker vectors are not high – thus indicating that these patients have median or below median expression values across all five biomarkers. Censored patients with less than 10-year follow-up time were omitted from the biplot. The figure further demonstrates the strong association between ER, Bag-1 and Bcl-2.