RELATIONSHIP OF MAMMOGRAPHIC DENSITY AND GENE EXPRESSION: ANALYSIS OF NORMAL BREAST TISSUE SURROUNDING BREAST CANCER

Study population

The study population included 121 women from the PWBCS with available snap frozen extratumoral breast tissues and mammographic density. The PWBCS is a population-based case-control study conducted in two major cities in Poland (Warsaw and Łódź) during 2000-2003 (15). PWBCS cases were women aged 20-74 years with newly-diagnosed, pathologically-confirmed in situ or invasive breast carcinoma identified through a rapid identification system organized at five participating hospitals and via cancer registries. Fresh tissues from invasive tumors, non-neoplastic adjacent breast tissue and mammary fat tissue were collected at the time of breast surgery and snap frozen in liquid nitrogen. Tumor adjacent breast tissues used in this study were <2cm from the tumor margin. Based on in vitro evidence of their distinctive microenvironments (16), basal-like and luminal tumors were oversampled in this study. Information on clinicopathological, demographic, and anthropometric factors was collected from medical records and in-person interviews as described previously (15). All of the participants provided written informed consent under a protocol approved by the U.S. National Cancer Institute and local (Polish) institutional review boards.

Mammographic density measurement

Pre-treatment mammograms of the unaffected breast were sent to the Ontario Cancer Institute in Toronto, Canada where they were digitized using a Lumisys® 85 laser film scanner. Patient identifiers were permanently deleted from the electronic images. Craniocaudal views of digitized films were used to assess mammographic density with Cumulus®, an interactive computer-assisted thresholding program developed at the University of Toronto (17). One expert reader (NFB) measured absolute dense area (cm²) and total breast area (cm²) using the methods as described previously (17); percentage mammographic density was calculated by dividing the dense breast area by the total breast area and multiplying by 100. A repeat set of 49 images was assessed for reliability. The intra-class correlation coefficients for percent mammographic density, dense area, and total breast area were 0.95, 0.93, and 0.99, respectively, documenting excellent reproducibility.

Breast tissue composition measurement

Frozen non-neoplastic breast specimens of approximately 100 mg were cut over dry ice and then used to cut frozen sections. Sections were collected at both ends of the specimen and then constructed into 20 μm slides. The central portion was used for RNA extraction. After H&E staining, the slides were scanned into high-resolution digital images using the Aperio Scan-Scope XT Slide Scanner (Aperio Technologies, Vista, CA, USA) in the UNC Translational Pathology Laboratory. After excluding slides with poor resolution or having folded tissues, slides from 118 women (97.5%) were subjected to breast tissue composition analysis. To train the composition estimator in Aperio's Genie software, 15 representative digital slides were selected and manually annotated for epithelial area, stromal area, and total area (mm²) using Aperio ImageScope software. These digital area-based, quantitative estimates were used to train Aperio's Genie Classifier to partition epithelium, adipose tissue, non-fatty stroma, and glass into percentages. Examples of annotated digital images are presented in Supplementary Figure 1. The regular H&E counterparts of these 15 digital slides were also evaluated by eye by a pathologist who provided semi-quantitative estimates of the percentage of adipose tissue (10% bin width), epithelium (1% bin width), and non-fatty stroma (10% bin width). To assess the performance of Genie classifiers, we compared the results of three methods (by Genie, by pathologist digital slide-based, and by pathologist regular H&E slide-based). The trained classifier was positively and strongly correlated with manually scored area based on the digital images, for all three tissue compartments. The trained Genie classifier was strongly correlated with pathologist review based on regular H&E slides for stroma and adipose (Supplementary Table 1, Pearson correlation coefficient ranged 0.95-0.96), while relatively lower for epithelium (Pearson correlation coefficient=0.68). Compared with digital assessment, visual assessment (by human eye on regular H&E slides) of small percentage differences is weaker, such as epithelial tissue, which is sparse (<10%) in benign breast. Thus the digital image analysis data were used in analyses, and the trained Genie Classifier was then applied to the remaining slides.

RNA isolation and microarrays

All microarrays on non-neoplastic breast tissue were performed at the University of North Carolina at Chapel Hill. The central section of fresh frozen non-neoplastic tissue (as described above) was homogenized using a MagnaLyser homogenizer (Roche), and RNA was isolated by Qiazol extraction followed by purification on an RNeasy column as described in Troester et al. (5). RNA quality and quantity were analyzed on an Agilent 2100 Bioanalyzer and a ND-1000 Nanodrop spectrophotometer, respectively, before running two-color 4X44K Agilent whole genome arrays. Cy3-labeled reference was produced from total RNA from Stratagene Universal Human Reference (spiked 1:1,000 with MCF-7 RNA and 1:1,000 with ME16C RNA to increase expression of breast cancer genes) following amplification with Agilent low RNA input amplification kit. The same protocol was applied to total RNA from breast tissues, with all patient samples labeled with Cy5. Data were lowess-normalized, and probes that had a signal of <10 dpi in either channel were excluded as missing. Probes that had more than 20% missing data across all samples were excluded from further analysis. In expression data preprocessing, we 1) eliminated the probes without corresponding ENTREZ ID, 2) collapsed the duplicate probes by averaging, 3) imputed missing data using k-nearest neighbors (KNN) method with k=10, and 4) median-centered each gene. Microarray data for 121 specimens used in this analysis are publicly available through the Gene Expression Omnibus (GSE49175).

Statistical analysis

Samples were classified as having Active or Inactive extratumoral subtype using unsupervised hierarchical clustering (average linkage) on the genes in the published Active/Inactive extratumoral signature (6). 324 of 3518 genes in the original Active signature were excluded in the clustering analysis as they did not pass filtering criteria described above. Consensus clustering was used to evaluate whether two clusters resulted in optimal segregation of the sample classes (18). The Active or Inactive extratumoral subtype obtained by clustering was validated using the Creighton correlation method (19), described in Supplementary Figure 2. Briefly, a standard vector corresponding to all genes in the Active/Inactive signature was constructed, with 1 assigned to up-regulated genes and -1 assigned to down-regulated genes. A Pearson correlation coefficient was calculated for this standard vector vs. the vector of median centered gene expression for each patient. The identities of the 3194 genes and their corresponding standard vector for Creighton correlation in the present study were shown in Supplementary Table 2. Patients were classified as Active if the Pearson correlation coefficient was greater than zero, and Inactive if the coefficient was less than zero.

The distributions of breast cancer risk factors by the Active/Inactive subtype were assessed, including age at diagnosis (continuously and categorically as <50 y and ≥ 50 y), BMI (continuously and categorically as <30 kg/m² and ≥ 30 kg/m²), age at menarche (<12 y and ≥ 12 y), parity (nulliparous and parous), age at 1^st full term birth in parous women (<25 y and ≥ 25 y), use of oral menopausal hormone therapy (MHT, never and ever), menopausal status (premenopausal and postmenopausal), family history of breast cancer in the first degree relative (yes and no), previous benign breast disease history (defined as if a subject ever had any biopsy or partial removal procedure, but no cancer detected; yes and no), percentage mammographic density (continuously and categorically as <25% and ≥ 25%), and dense area (continuously and categorically as <median and ≥ median; median=36.52 cm²). Moreover, the distributions of clinicopathologic characteristics by the Active/Inactive subtype were evaluated, including tumor size (≤2 cm and >2 cm), histological type (ductal and others), differentiation stage (well/moderate and poor), number of positive axillary lymph nodes (0 and ≥1), ER status and intrinsic subtype. To identify potential confounding factors, the distributions of clinicopathologic characteristics and breast cancer risk factors by percent mammographic density and dense area were also examined. The corresponding associations were evaluated using Chi-square tests for categorical variables and Student's t-tests for continuous variables. Moreover, to assess the magnitude of association between risk factor exposure (e.g. obese vs. nonobese) and extratumoral microenvironment subtype, we computed odds ratios (OR) for the Inactive subtype (vs. Active, referent) and corresponding 95% confidence intervals (CI). Analysis of covariance (ANCOVA) was used to examine the relationship between Active/Inactive subtype and mammographic density, adjusting for potential confounders, age and BMI, continuously.

To explore the biological meaning of microenvironment subtypes, the gene expression profile in each sample was compared to five previously reported gene expression signatures of related phenotypes, including young age (20), obesity (21), desmoid-type fibrosis (DTF) (22), TGF-β (6) and estrogen (E2) response (23). To ensure that each of these signatures was independent, overlap between gene sets was assessed and found to be small (Supplementary Table 3). The median centered expression profile of each patient was evaluated by calculating Pearson correlation coefficients, using the method of Creighton et al. described above (19). More information regarding the gene ID and direction of each signature is listed in Supplementary Table 4. The associations of existing signatures with the extratumoral subtype (Active and Inactive), percentage mammographic density (<25% and ≥25%), and breast tissue composition (percentage adipose, epithelium and stroma) were assessed by Chi-square tests and Student's t-test, respectively. Non-parametric exact methods were used when expected cell count was less than 5. Logistic regression and generalized linear regression were used to further evaluate the associations after adjusting for age and/or BMI. Probability values of less than 0.05 were considered statistically significant. All analyses were performed using R (version 2.9.2).

Associations between Active/Inactive subtype and breast cancer risk factors

Unsupervised hierarchical clustering (average linkage) on thê3,000 genes from the published Active/Inactive signature was used to classify extratumoral breast tissues from 121 cases into two homogeneous groups. Consensus clustering documented that two groups provided an optimal segregation of the data (Supplementary Figure 3). Using the Creighton correlation method to categorize the microenvironment subtype as described in Methods also confirmed that the direction of association for Active and Inactive groups by clustering was concordant with the previous publication of the signature, showing high agreement of the two classification methods with Kappa statistic of 0.92 (p<0.0001). Approximately 49% of patients (n=59) were Active in the current study, compared to 38% of patients (n=27) in Roman-Perez et al. (6).

Consistent with our previous findings (6), we observed no association between extratumoral subtype and ER status and other tumor characteristics (Supplemental Table 5). Table 1 shows the distribution of participant characteristics, overall and categorized according to extratumoral microenvironment subtype. Of all the breast cancer risk factors evaluated, mammographic density measures showed the strongest differences by Active/Inactive subtype. Compared with the Active group, mean percentage density was 9% higher and mean dense area was 8.8 cm² more in the Inactive group (percentage density: p=0.0013; dense area: p=0.049). Substantial (but statistically insignificant) associations were noted for age [age <50y vs. ≥50y: OR = 2.00, 95% CI=(0.88, 4.57)] and BMI [BMI <30kg/m² vs. ≥30kg/m²: OR = 1.76, 95% CI=(0.79, 3.93)], but no other breast cancer risk factors showed substantial associations with Active/Inactive subtype.

We also evaluated potential confounders of the association between mammographic density and extratumoral microenvironment subtype. Both elevated percentage density and dense area were significantly associated with younger age, lower BMI, ever use of oral MHT and premenopausal status (Supplementary Table 6). Of these factors, only age and BMI showed substantial associations with extratumoral subtype. Thus, we performed multivariable analysis to study mean differences in mammographic density by extratumoral microenvironment subtype, controlling for age and BMI (Table 2). In multivariable models, percent density and dense area remained higher among the Inactive group (percentage density: p=0.0038; dense area: p=0.049).

Biological relevance of Active/Inactive subtype

To better understand the molecular characteristics of extratumoral microenvironment subtype and mammographic density-associated breast cancer risk, associations were evaluated between Active/Inactive subtype, mammographic density, and biologically-relevant, published gene expression signatures. As shown in Table 3, both the Inactive subtype and higher percentage density (defined as ≥ 25%) were significantly positively associated with a young-like age signature and an increased estrogen response expression signature (E2), while they were significantly inversely associated with an obesity signature and a TGF-β signature. After adjusting for patient age and/or BMI, most of these associations remained statistically significant, but the association of high percentage density with the young-like age signature and low TGF-β signature was no longer significant (Supplementary Table 7). The previously reported signature derived from desmoid-type fibroid tumors and indicative of fibroblast response (DTF) was only associated with lower percentage density, but not after adjusting for BMI and age.

Associations between subtype/mammographic density and breast tissue composition

Given variation in stroma-derived signatures, we expected that stromal composition may vary by extratumoral microenvironment subtype. Indeed, as shown in Figure 1, there were substantial differences in breast tissue composition by Active/Inactive subtype. These differences paralleled the differences in tissue composition by percentage mammographic density and dense area. The percentage of adipose tissue was significantly lower and the percentage of non-fatty stromal tissue was significantly higher among Inactive/mammographically dense patients (Supplementary Table 8). No differences were detected in the percentage of epithelial tissue by extratumoral microenvironment subtype or measures of mammographic density.

Finally, because Active/Inactive subtype and mammographic density appeared to be tracking cellular composition, the association between published gene expression signatures and cellular composition was also considered (Table 3). Again mirroring microenvironment subtype and mammographic density, the percentage of non-fatty stroma was associated with young-like gene expression, decreased expression of obesity-associated genes, decreased expression of a TGF-β signature, and increased expression of estrogen response signatures. Epithelial composition was modestly associated with gene expression signatures for age and obesity (higher proportion in young-like and non-obese groups) after adjusting for BMI and age respectively (Supplementary Table 7).