Construction of Prostate Tissue Microarray
A 255 core tissue microarray (TMA) was constructed from formalin-fixed and paraffin-embedded prostatectomy specimens obtained from patients with prostate cancer as described in the methods. The TMA was representative of the spectrum of prostate cancer Gleason grades, ranging from Gleason Score 5 to Gleason Score 9. A total of 50 tumor foci (one focus from 44 specimens and 2 foci from 3 specimens) including 35 of peripheral zone and 15 of transition zone origin were included in the array. A summary of the tissue core composition of the TMA is presented in Tables and .
Gleason Score Distribution of TMA
Distribution of Cores by Gleason Grade on TMA
Structuring of Immunohistochemical Data
The expression of selected biomarkers associated with cell cycle progression and/or cell death sensitivity was assessed using routine immunohistochemical methodologies. These markers included bax, bcl-2, bcl-xL, bin-1, CD95 (Fas/APO-1), MDM2, p21waf1/cip1, p53, and p65 (NFkB). In order to facilitate the statistical analysis of the immunohistochemical information of the various markers the data was structured in the following manner. The percentage of tumor exhibiting detectable levels (i.e., involvement) of a specific marker was scored 0 (no detectable staining above background), 1 (1-25% of the tumor cells exhibiting detectable staining), 2 (25-75% of the tumor cells exhibiting detectable staining), or 3 (> 75% of the tumor cells exhibiting detectable staining). The intensity of staining for the individual markers, if detectable, was scored as either low (1) or high (2). Staining intensity was assessed relative to internal positive controls, such as basal epithelial cells or lymphocytes present on the array, or tissue controls used to establish the appropriate antibody titrations. Data analysis revealed that there was no significant difference attributable to subcellular distribution of the biomarkers assessed in this study. In order to incorporate both the involvement and the intensity information for each biomarker a new variable, "expression" was generated, which was defined as the following:
Expression = 0 if involvement = 0 (intensity = 0 by definition)
= 1 if involvement = 1
= 2 if involvement = 2
= 3 if involvement = 3 and intensity = low
= 4 if involvement = 3 and intensity = high
After careful observation of the patterns shown in our original data, a new variable, expression, was defined (Table ). The data set had a special feature in that the intensity information only became relevant in defining ''expression'' when involvement = 3. When involvement = 0, 1 or 2, all cores had ''low'' intensity, except 3 cores when involvement = 2.
The number of cores by involvement and intensity category.
Statistical Analysis and Data Modeling of Biomarkers
The prostate cancer TMA was constructed to address several issues of both biologic and clinical interest. It was first of interest to determine the frequency of individual biomarker expression and to determine the extent to which individual biomarkers correlated with the prostate zone of tumor origin, i.e. peripheral versus transition zone within Gleason grade 3+3 tumors (Table ). Bcl-2 and p53 exhibited no detectable levels of expression in this subset of Gleason grade 3+ 3 tumors and were therefore not included in Table . For each of the other biomarkers, a linear mixed effect model was fit for the biomarker expression, using the zone of origin as a fixed effect and patient as a random effect. The mixed model also took into account the potential association among multiple cores within the same subject through modeling the variance-covariance matrix of the residuals. The model suggested that there was no significant difference in CD95 or bcl-xL expression between peripheral zone and transition zone tumors (p = 0.61 and 0.12, respectively). Among bax (p < 0.0001), bin-1 (p = 0.005), p21 (p = 0.0008), mdm2 (p = 0.02) and p65 NFkB (p = 0.02), the model suggested that the expressions level in transition zone was significantly lower than that of the peripheral zone. The Bonferroni method was used to adjust for multiple comparisons (i.e., using a cutoff value of 0.05/7 = 0.007), only bax and p21 remained to show significant differences between the peripheral and transition zone with regards to biomarker expression.
The fitted linear mixed effect models for each biomarker expression within Gleason grade 3+3 tumors, where zone of origin (TZ vs. PZ) was fitted as a fixed effect and patient as a random effect.
The Spearman's correlation coefficient was used to assess potential significant associations between Gleason grades and biomarker expression for each of the nine biomarkers assessed. The array was comprised of 30 cores of grade 2 tumor, 111 cores of grade 3 tumor, 96 cores of grade 4 tumor, and 18 cores of grade 5 tumor. Using the median biomarker expression within each subject, there was no significant association between the level of mdm2, bcl-xL or p53 expression and Gleason grade (p = 0.19, p = 0.27 and p = 0.75, respectively). However, all of the six remaining biomarkers exhibited a significant association with Gleason grade. This association was strongest for bax (r = 0.62, p < 0.0001) and p21 (r = 0.57, p = 0.0002) with each marker exhibiting increasing expression with corresponding increases in Gleason grade. The majority of the cores for Fas, and p65 exhibited no detectable expression, however, when expressed these markers were associated with higher Gleason grade (p = 0.01 and p = 0.002, respectively). The "heat map" enabled a visual representation of the variation in individual biomarker expression associated with Gleason grade (Figure ).
Plot of biomarker expression vs. Gleason grade. Bright red corresponds to the highest expression level, black corresponds to moderate expression and light green is the lowest expression.
The Spearman correlation coefficient was computed to identify significant associations between pairs of biomarkers (Table ). The median biomarker expression level within each subject was used for this calculation. This analysis demonstrated that the expression of bax and bin1, bax and MDM2, bax and p21, and bax and p65 NFkB was highly associated. Other significant associations were identified between bin1 and p21, bin1 and MDM2, bin1 and p65 NFkB and between p21 and p65 NFkB. Similar associations were identified when using the mean, instead of the median biomarker expression level within each subject to calculate the Spearman correlation coefficient.
Spearman correlation among biomarkers with regards to the median expression levels within subject.
Univariate and multivariable logistic regression models were used to predict the biologic potential of Gleason score 7 (either 3 + 4 or 4 + 3) tumors using a profile of the biomarkers included in this study. For the purposes of this study Gleason score 5 and 6 tumors were considered clinically indolent (good prognosis) and Gleason score 8 and 9 tumors were considered clinically aggressive (poor prognosis). Our goal was to assess whether Gleason score 7 tumors are more similar to Gleason score < 7 (good prognosis) tumors or closer to Gleason score > 7 (poor prognosis) tumors, based on a fitted model using biomarker expression profiles. First, a generalized estimation equation (GEE) model was fit for the patients with Gleason score not equal to 7. The response variable was equal to 1 if the patient had poor prognosis (i.e., Gleason score > 7), and 0 otherwise. The GEE method takes into account the association among the multiple cores obtained from the same patient through modeling a working correlation matrix. In our case, the working correlation matrix was specified to be "unstructured" (UN). After a stepwise model selection procedure, a final fitted model was determined (Table ). Based on this model, Bax involvement = 3, FAS involvement > 0, p21 involvement = 3 was associated with an increased odds of having poor prognosis, and Bcl-XL intensity = high was associated with a reduced odds of having poor prognosis.
Fitted GEE model for predicting poor prognosis (GS > 7) in patients with Gleason score < 7 or > 7.
The model shown in Table was used to predict the probability of having poor prognosis for the patients with Gleason score 7 tumors (n = 10 patients). The results indicate that the predicted probability of poor prognosis in those Gleason score 7 patients (i.e., the within-patient mean predicted probability) ranges from 0.340 to 0.467. In contrast, in patients with Gleason score greater than 7, the predicted probability for poor prognosis ranges from 0.386 to 0.417, while in patients with Gleason score less than 7, the predicted probability ranges from 0.312 to 0.376, except for one patient whose predictive probability is 0.426. If 0.386 is used as the cutoff value for classifying poor prognosis, then 6 out of the 10 patients with Gleason score 7 would be classified as having poor prognosis.
Subsequently, it was of interest to determine whether all Gleason grade 3 tumor areas were equivalent with respect to the biomarker expression profile, or whether subsets could be defined which exhibited differences in biomarker expression. The Gleason scores containing '3' include 2+3, 3+2, 3+3, 3+4, 4+3, 3+5 and 5+3. The association between Gleason score and biomarker expression was assessed using the Spearman correlation coefficient. Again, the median biomarker expression was used in each patient. This analysis revealed significant differences between Gleason grade 3 biomarker signatures that varied within the context of the Gleason score (Table ). This finding implies that, with respect to the expression of the biomarkers assessed in this study, not all Gleason grade 3 tumor areas are equivalent. Further, the results indicate that the Gleason grade 3 signatures obtained correlated with the corresponding clinically indolent or clinically aggressive categories of the overall Gleason score.
The association between the median biomarker expression within subject and the Gleason scores in patients having Gleason Grade 3.