Protein expression based multimarker analysis of breast cancer samples

A high-density breast TMA was constructed using cores from formalin-fixed, paraffin embedded breast tissue donor blocks, consisting of 242 breast surgical cases of 210 patients who underwent surgery at the UCLA Medical Center between 1995 and 2000, as previously described [17, 18]. Archival samples were obtained from the UCLA Department of Pathology and Laboratory Medicine with oversight and approval from the UCLA institutional review board. Such samples were consented for use in biomedical research projects at the time of surgery. At least three cores of each available histologic type were arrayed from the donor blocks. Of the 242 surgical cases, 179 cases (from 157 patients) were of invasive breast cancers of various histologic types. For our multimarker analysis, we selected 82 primary surgical cases, each belonging to a unique patient with invasive cancer who did not receive neoadjuvant therapy, had disease-specific survival outcome, and were informative for expression in most of our protein markers.

Immunohistochemical staining of the breast TMA was performed using a standard two-step indirect avidin-biotin complex method (Vector Laboratories, Burlingame, CA) or a two-step polymer detection method (DakoCytomation, Inc., Carpinteria, CA) as previously described [19‐22]. The following primary antibodies were used: BS106, BU101, Mammaglobin (Abbot Laboratories, Abbott Park, IL), prolactin-inducible protein (Signet Laboratories, Inc., Dedham, MA), S100A7 (Imgenex Corp., San Diego, CA), 14-3-3 σ (Research Diagnostics, Inc., Concord, MA), Her-2/neu (Zymed Laboratories, Inc., South San Francisco, CA), progesterone receptor, estrogen receptor alpha, p53 (DakoCytomation, Inc., Carpinteria, CA), RIN1, annexin A1, beta-catenin (BD Biosciences Transduction Laboratories, Lexington, KY), Na-KATPase-β1, Na-KATPase-α, GATA3, Smad2 (Santa Cruz Biotechnology, Inc., Santa Cruz, CA), Smad4 (Millipore, Billerica, MA), YY1 (Geneka Biotechnology, Inc., Montreal, Quebec, Canada), TGF β receptor II (Abcam, Inc., Cambridge, MA), H3K4 and H3K18 (Upstate, Lake Placid, NY), and MED28 (gift from Dr. Mai Brooks). Briefly, 4 μm sections were deparaffinized, treated with 0.3% hydrogen peroxide in methanol, blocked with 5% serum, and incubated with primary and secondary antibodies. Diaminobenzidine was used for color detection. A concentration-matched isotype control IgG was used for negative controls. Note that the 26 markers analyzed here were originally chosen for other oncogenic studies conducted in our laboratory.

The level of protein expression in glandular epithelial cells was quantitatively assessed by a pathologist blinded to all clinico-pathological variables. We used the percentage of cells staining, referred to as "pos", as the quantitative measure of protein expression. To arrive at a single staining measure per patient (referred to as "pos.mean"), we averaged the pos measures of multiple cancer spots per patient as described in [23].

To validate our WGCNA* and COX mortality group definitions, we selected all Affymetrix HG-U133A gene expression data sets from the Gene Expression Omnibus (GEO) that were published in 2005 or later. This resulted in three independent data sets published in 2005-2006, that had the following GEO identifiers: Miller 2005 - GSE3494 (251 arrays), Pawitan 2005 -GSE1456 (159 arrays), Sotiriou 2006 - GSE2990 (189 arrays) [24‐26]. Data sets were pre-processed as described in [27]. Briefly, within each data set we evaluated array quality by comparing inter-array correlations. Arrays with low inter-array correlation were removed according to default recommendations [27]. When expression analysis was distributed across multiple centers, we checked for center-related batch effects. If batch effects were present we removed them using the combat function [28]. The pre-processing steps removed 3-12% of arrays in an unbiased fashion resulting in 222, 146 and 183 arrays for the Miller 2005, Pawitan 2005 and Sotiriou 2006 data sets, respectively. Finally, we removed all samples with missing survival data, resulting in a total of 207, 146 and 173 patients for the Miller, Pawitan and Sotiriou data sets, respectively.

Univariate Cox proportional-hazards models were constructed for the WGCNA* patient groups and COX rule patient groups for each of the three data sets. We used a moderate significance level of 0.1 to allow for expected expression differences between genes and proteins.

In the following, we outline the analysis steps for conducting a WGCNA of the TMA patient data. An overview diagram is provided in Figure 1. R software for WGCNA and accompanying software tutorials are freely available at: http://​www.​genetics.​ucla.​edu/​labs/​horvath/​CoexpressionNetw​ork/​.

We applied the WGCNA methodology outlined above to a high-density TMA platform consisting of 26 putative tumor biomarkers measured on 82 breast cancer patients (Tables 1, 2). The patients were clustered by their expression profiles, which were transformed to adjacencies using the soft threshold (power) β = 10 and the signed network option for calculating adjacency in the WGCNA R package. The cutreeDynamic function was used (with options minClusterSize = 2 and deepSplit = 3) to generate 16 modules (not shown). Since we did not expect that these modules were robustly defined given the relatively small data set, we merged similar modules using the mergeCloseModules R function with the "cutHeight" parameter set to 0.15, which was optimal for obtaining fewer but more robustly defined modules that were significantly associated with survival (p-value = 3.9 × 10^-4). Since the groups were defined with respect to the survival outcome, the p-value is overfit and should be interpreted as a descriptive (not inferential) measure. To arrive at an unbiased evaluation of the patient groups, we used independent gene expression data sets as described below. The WGCNA groups corresponded to mortality rates of 5.4%, 22%, and 50% (colored white, grey and black, respectively in Figure 2A, 2B, 2C).

To test whether the median values of ordinal variables differed between WGCNA patient groups, we used the Kruskal Wallis test, which is a non-parametric multi-group comparison test. Boxplots were used to visualize the distribution for each group. Lymph node involvement, stage, metastasis and estrogen receptor positivity were significant at the 0.05 level, but none of these variables could completely define one or more of our patient groups (data not shown). Furthermore, an ordinal multivariate regression model predicting our patient groups from these four variables resulted in a McFadden pseudo R-square of only 0.068 (SPSS v16.0). After verifying that our WGCNA groups were distinct from our clinical variables, we evaluated its utility for survival prediction in the presence of other predictive variables. Variables that were significant at the 0.05 level in univariate Cox proportional-hazards models included lymph node involvement, stage, metastasis and Her2 positivity. A multivariate Cox proportional-hazards model that included these variables and our WGCNA variable found all predictors to be non-significant (p > 0.05) except the WGCNA mortality groups, where the high mortality group had a p-value of 0.037. These results suggest that the WGCNA mortality groups have distinct molecular characteristics that predict breast cancer survival independently of prognostic clinical variables. However, the WGCNA cluster variable was defined with respect to the 26 markers, which is cumbersome to validate. Therefore, we aimed to develop a simple classification rule (referred to as WGCNA*) which assigns each patient to its respective WGCNA cluster. Toward this end, we used classification trees implemented in the rpart R package, which automatically selects significant markers and corresponding thresholds. The classification tree led to a WGCNA* rule based on three markers p53, Na-KATPase-β1, and TGF β receptor II; with optimal thresholds corresponding to the 75^th, 33^rd and 66^th percentiles, respectively (Figure 3A). Due to missing data in the p53, Na-KATPase-β1, and TGF β receptor II markers (2, 4 and 8 missing values, respectively), the WGCNA* rule initially resulted in only 66 patients with mortality group assignments. As a result, we confirmed that the missingness pattern of each marker was unrelated to survival and then replaced missing values by the median (rather than the average due to skewed distributions), so that all 82 patients were assigned to a WGCNA* mortality group. By construction, the WGCNA* mortality groups closely matched the original WGCNA groups, differing by only 10 patients (Table 3). As a result, the WGCNA* patient groups were highly predictive of patient survival (p-value = 9.1 × 10^-5, Table 4), with mortality rates of 5.4%, 24%, and 67% (Figure 2D, 3A). The WGCNA* rule suggests that breast cancer patients with high p53 and low Na-KATPase-β1 have a high risk of death in comparison to other molecular profiles. Furthermore, patients with low p53 and high TGF β receptor II have a moderate mortality risk.

To elucidate the clinical meaning of the WGCNA* groups, we related them to clinical variables. We found that the WGCNA* groups are significantly related to stage, metastasis and estrogen receptor positivity (p < 0.05, Figure 4A). A multivariate Cox proportional-hazards model that included the prognostic clinical variables (lymph node involvement, stage, metastasis and Her2+) had a slightly lower R² (0.306 versus 0.326) and hazard ratio (3.8 versus 5.9) for the high mortality group when WGCNA* was used as a predictor rather than the original WGCNA grouping variable based on all 26 markers (Table 5). However, these differences are negligible given that WGCNA* substantially reduced the number of markers needed for validation.

We verified that no single marker could define the WGCNA groups (Figure 4B). Cox proportional-hazards models that included WGCNA* and each marker individually (dichotomized at its optimal survival prediction threshold) found WGCNA* to be the top predictor in all cases (data not shown).

While WGCNA defines patient mortality groups that predict survival independently of other clinical variables, it is interesting to know how this approach compares with a more conventional stepwise variable selection approach [37‐39]. We analyzed the same TMA data by first dichotomizing each marker at an optimal threshold for survival prediction. Ten markers could be dichotomized at a level that achieved a minimum of five patients per group and a univariate Cox proportional-hazards model p < 0.05. We included these 10 markers in a multivariate Cox proportional-hazards model and removed the least significant predictor in a step-wise manner, until all remaining variables achieved significance at the 0.05 level. Four markers: MED28 expressed in the cytoplasm, p53, Smad4 expressed in the cytoplasm and Her2 were retained. The resulting model had an R² of 0.321. We then used classification trees (rpart with complexity parameter 0.1) to identify a subset of markers and their thresholds for defining low, moderate and high mortality patient groups. This was achieved with two markers MED28 and Smad4, dichotomized at their 75^th percentiles, where high MED28 resulted in high mortality and low MED28 in conjunction with high Smad4 resulted in moderate mortality (Figure 3B). The resulting "COX" patient groups were significantly related to survival (p = 1.6 × 10^-4) and had mortality rates of 4.1%, 33% and 50%. While the COX variable had similar mortality rates to WGCNA and WGCNA*, there was a substantial 35% difference in the assignment of patients to mortality groups (Table 3, Additional File 1). A multivariate Cox proportional-hazards model that included the COX variable and the prognostic clinical variables, found the COX variable to be the best predictor with p-values of 0.016 and 0.036 for the moderate and high mortality groups, respectively (Table 5). Finally, to directly compare the WGCNA* and COX variables, we included both in a Cox proportional-hazards model with and without the prognostic clinical variables. The WGCNA* high mortality group was the most significant predictor (p = 0.017) in the absence of clinical variable data, but when prognostic clinical variables were included, the COX moderate and high mortality groups were the only significant predictors in the model.

In summary, the WGCNA* and COX mortality groups are distinct from each other and are both important predictors of breast cancer survival in our TMA data. While COX outperforms WGCNA* in the presence of prognostic clinical variables, it was also created by optimizing the significance of its underlying markers in a Cox proportional-hazards model. Thus, the COX variable's superior performance could possibly be explained by over-fitting in our TMA data set and may not validate in other data sets. To test this, we attempted to validate the WGCNA* and COX mortality group rules in independent gene expression data sets.

We applied the WGCNA* and COX rules to three independent Affymetrix HG-U133A data sets (GSE3494, GSE1456, GSE2990) [24‐26]. After data cleaning (as described in the Methods section) the data sets consisted of 207, 146 and 173 breast cancer patients for the Miller 2005, Pawitan 2005 and Sotiriou 2006 data sets, respectively (Additional File 2A). The data sets had mortality rates ranging from 17%-27%, with the Pawitan 2005 and Sotiriou 2006 data sets being most similar to the 16% rate in our TMA data. Other than survival, grade was the only clinical variable common to all three data sets, and it appeared fairly consistent across the gene expression and TMA data sets. The probe set distributions were also consistent, where the interquartile ranges overlapped for all Miller 2005 and Pawitan 2005 probe sets; which overlapped with Sotiriou 2006 in most cases (Additional File 2B). There were no anomalies to disqualify a data set or probe set for validation analysis.

On the Affymetrix HG-U133A array, the protein Na-KATPase-β1 was represented by gene ATP1B1 which had two probe sets 201242_s_at and 201243_s_at. Similarly, p53 corresponded to the TP53 gene (201746_at and 211300_s_at); and TGF β receptor II corresponded to TGFBR2 (207334_s_at and 208944_at). Thus there were eight versions of the WGCNA* rule corresponding to each probe set combination (2 × 2 × 2). Similarly for the COX rule, the proteins MED28 cyt and Smad4 cyt were represented by genes MED28 (214831_at, 218438_s_at) and SMAD4 (202526_at, 202527_s_at) resulting in four versions of the COX rule per validation data set. The strongest WGCNA* validation was achieved in the Pawitan 2005 data set, where all eight probe set combinations validated for the high mortality group (p < 0.1, Table 6, Figure 5). In the Miller 2005 data set, two probe set combinations validated. No probe sets combinations validated in the Sotiriou 2006 data set. Applying the same validation criteria (p < 0.1), the COX rule did not validate in any data set for any probe set combination (data not shown). In summary, the WGCNA* high mortality group validated in two of three Affymetrix breast cancer data sets, suggesting that patients with high TP53 and low ATP1B1 may have a worse prognosis than patients with other profiles. Since high p53/TP53 is a well known indicator of poor prognosis, the following section verifies that the combination of Na-KATPase-β1/ATP1B1 and p53/TP53 in both the protein and gene expression data sets is a stronger survival predictor than p53/TP53 alone.

Since the WGCNA* high mortality group was defined by low Na-KATPase-β1 and high p53, we checked whether p53 alone would be a sufficient or possibly superior survival predictor. In our TMA data, the optimal dichotomized threshold for p53 was the 75^th percentile. In a Cox proportional-hazards model that included both the WGCNA* high mortality group (coded as high versus moderate and low combined) and the dichotomized p53 marker, the hazards ratio for the WGCNA* high mortality group was more than two-fold higher at 4.5 (p = 0.07) versus a hazards ratio of 2.1 (p = 0.38) for the dichotomized p53 marker. In the gene expression data, the continuous form of the TP53 variable was not significant while the WGCNA* high mortality group maintained significance at the 0.05 level for all 8 of the Pawitan 2005 models. The dichotomized TP53 marker did achieve significance at the 0.05 level in two of the Miller 2005 models, but high TP53 indicated a protective effect, which is inconsistent with current (protein-level) findings (HR = 0.35 and p = 0.04 for both models). In summary, low Na-KATPase-β1 (< 33^rd percentile) in combination with high p53 (>75^th percentile) is a stronger predictor of mortality than p53 alone in both our TMA data and the Pawitan 2005 gene expression data set.

Since the WGCNA* high mortality group consistently validated in the Pawitan 2005 data set, we explored the relationship between this group and the available Pawitan 2005 variables: subtype (Basal, ERBB2, luminal A, luminal B and normal like) and grade (I-III). In this data set, the high mortality group consisted of 11 patients, 7 of which were luminal B, one luminal A, one basal, and two were missing subtypes. Thus, subtype was significantly related to WGCNA* high mortality (Fisher's exact test p = 1.5 × 10^-4). Similarly, seven of the high mortality group patients were grade 3, two were grade 2, and two were grade 1, although this relationship did not achieve significance (Fisher's exact test p = 0.194). In a multivariate Cox proportional-hazards model with subtype coded as luminal B versus other types and grade coded as an ordinal variable, the WGCNA* high mortality group was the strongest predictor with a hazards ratio of 4.22 (95% CI: 1.3, 14.1, p = 0.019). See Additional File 3 for characteristics of the WGCNA* mortality groups.