Case definition, ascertainment, and selection.

Cases were women with a biopsy for BBD and with a subsequent first diagnosis of invasive breast cancer (at least one year after the index BBD biopsy). The occurrence of breast cancer in the cohort was ascertained by linking records from the BBD cohort to the KPNW Tumor Registry. The KPNW Tumor Registry has operated continuously since 1960 and is approved by the Commission on Cancer of the American College of Surgeons. It was computerized in 1978 and includes all cases diagnosed since 1960. Cases of cancer are identified primarily from pathology reports forwarded directly from the KPNW Department of Pathology to the Tumor Registry at the time of diagnosis. These cases are supplemented by examination of the inpatient discharge logs, referrals to radiation oncologists, and surveillance of radiology reports marked as particularly suspicious for cancer. The terms of insurance coverage by Kaiser Permanente serve to concentrate members’ medical care within the walls of KPNW, particularly for chronic diseases such as cancer. (In a review of 1,846 incident breast cancers diagnosed during the period 1960-1985, it was found that 99.1% were diagnosed and treated at KPNW, 0.6% were diagnosed at KPNW and treated elsewhere, 0.1% were diagnosed at autopsy, and only 0.2% were diagnosed elsewhere but treated at KPNW (A.Glass, personal communication).) The KPNW Tumor Registry has an excellent follow-up rate, even for women who are no longer health plan members, and it has maintained a follow-up rate of 98% of patients (living and dead) enrolled since its reference date of 1960.

Our goal was to include in the discovery case-control study approximately 2/3 of the total number of cases expected to be ascertained in the cohort through the end of the study period. S1 Fig shows that, after exclusions for various reasons, the discovery stage included 316 cases (and matched controls) and the test stage included 165 cases (and matched controls).

Control definition and selection.

Controls were women with a biopsy for BBD who were alive but had not developed breast cancer during the same follow-up period as that for the corresponding cases. For each of the cases, we randomly selected one control using risk-set sampling. Each control was individually matched to the corresponding case on age (+/- 1 year) and age at diagnosis of BBD (+/- 1 year)(and implicitly, given the risk-set sampling, on duration of membership of KPNW), and was sampled randomly from the risk-set with replacement [11]. In addition to being alive and free of invasive breast cancer, the controls were required not to have undergone a mastectomy before the date of diagnosis of breast cancer for its matched case.

RNA extraction.

For each study subject, a minimum of 6 morphologically guided 1mm tissue cores was obtained using a chemicon microarrayer from the identified benign lesions and collected in RNase-free eppendorf tubes. For RNA extraction, we used the method that we have described elsewhere, and that has been shown to yield very high correlations (r>0.94) between global miRNA expression profiles obtained using the Illumina miRNA expression profiling platform on RNA extracted from matched fresh and FFPE tissue [16]. In brief, the FFPE tissue cores were washed, rehydrated and homogenized in the presence of proteinase K digesting buffer (50mM Tris-HCl pH7.5, 75mM NaCl, 5mM Cacl2 and 0.1% SDS) and 30mg/ml proteinase K (Roche Diagnostics, Inc.). After Butanol-1 extractions, the 150μl tissue digests were subjected to TRIzol (Invitrogen) extractions, following the manufacturer’s instructions. The RNA obtained from the upper layer was precipitated at -20°C O/N and centrifuged to obtain RNA pellets. The RNA pellets were washed, resuspended in 20μl in 1x Tris HCl pH7.5, EDTA and incubated at 70°C for 30min for removal of cross-linkages. The RNA was then quantified on a Nanodrop ND-1000 and on an Agilent Bioanalyzer Nano chip before being stored at -80°C.

Discovery stage—MiRNA expression profiling.

For the first case-control study, miRNA expression profiling of the FFPE specimens was performed using the Illumina MiRNA Expression Profiling Assay with 200ng of total FFPE RNA. This assay, now discontinued, was an adaptation of the DASL^® (cDNA-mediated Annealing, Selection, Extension, and Ligation) Assay. With this assay, we simultaneously analyzed the expression of 1,145 miRNAs from FFPE RNA specimens in batches of 48 individual samples (24 matched case-control pairs), in duplicate (96 samples per experiment). For the expression profiling of sample batches we followed the manufacturer’s instructions, using the 96 samples assay kit containing 8 beadchip microarrays, where each beadchip had 12 individual microarrays. In brief, 200ng of total FFPE RNA in 5μl from 48 samples in duplicate in a 96 well-plate was poly-adenylated for 1h at 37°C. The poly-adenylated miRNAs were then reverse transcribed using a biotin-labeled oligo-dT primer and the supplied reagents at 42°C for 1h. Then, a mix of miRNA-specific oligonucleotides was added and subjected to hybridization of the reverse-transcribed material for 4h on a heat-block (annealing based on temperature decrease from 70°C to 40°C). The duplex cDNA-oligonucleotides were purified by binding to magnetic streptavidin beads, and the oligonucleotides were extended to create a double-stranded DNA template and eluted from the magnetic beads. The eluted DNA products were subjected to PCR amplification (34 cycles at 95°C for 35 seconds, 56°C for 35 seconds and 72°C for 2 minutes) and purified using streptavidin magnetic beads. The individually purified PCR products (96 samples) were then loaded onto the opening of the individual microarrays (12 per beadchip) and incubated at 58°C for 16h under agitation, using the appropriate hybridization reagents. The microarrays were washed the next day using Illumina’s reagents and scanned on an Illumina beadarray reader. The raw data were extracted using GenomeStudio and exported to Excel sheets for statistical analysis.

Intermediate and validation stage—MiRNA reverse-transcription and quantitative PCR (qPCR).

In an intermediate stage, we selected 81 miRNAs (plus 4 controls) to be analyzed on 15 case-control pairs from the discovery stage of the study, to perform cross-platform validation between the Illumina MicroRNA profiling assay and Taqman^® miRNA assays and reagents (S1 Table). The 15 pairs selected for these analyses were randomly selected from the FFPE RNA specimens available, but given that a total of 81 miRNAs were analyzed at this stage and that each miRNA assay required 10 ng FFPE RNA, selection was restricted to those samples that yielded at least 1μg of RNA (75% of all samples produced at least 1μg of RNA). The correlation between data obtained from these two platforms was examined and the 14 miRNAs that were most differentially expressed between cases and controls and that also showed the strongest correlations between the two platforms were selected for the validation stage.

In the validation stage, the 14 miRNAs were each quantified in triplicate in total RNA extracted from FFPE BBD lesions of matched cases and controls using Taqman^® miRNA qPCR assays and reagents (Applied Biosystems, CA, USA) as described elsewhere [17]. MiR-191 and RNU6b were used as endogenous controls for data normalization [17]. Quantitative PCR measurements were performed on a Step-One-Plus qPCR instrument using the manufacturer’s recommended 96 well-plates and covers, and using the qPCR cycles recommended for Taqman^® miRNA quantifications. Quantitative data were then transferred to Excel sheets for statistical analysis.

Discovery stage.

Unconditional logistic regression was used to generate odds ratios and 95% confidence intervals (CI) for the association between lifestyle factors and breast cancer risk. Quantitative factors, including body mass index (BMI (weight (kg)/height (m²)), age at menarche, age at first live birth and number of pregnancies, were categorized according to their distribution in the controls. A matched pair analysis was not used because a substantial proportion of subjects had missing data on lifestyle factors. Instead, we included the matching variables, age, age at diagnosis of BBD, and duration in the cohort, as covariates in the regression model. The association between BBD pathology and breast cancer risk was analyzed using logistic regression in the same way. The categories of “No breast lesion” and “Non-proliferative lesion” were combined and treated as the reference category.

For analysis of the miRNA data from the discovery stage, obtained using the Illumina MiRNA Expression Profiling Assay, we log-transformed and quantile normalized the miRNA intensity values obtained from the Illumina bead arrays, and used them in all further analysis of miRNA expression levels. For each sample, correlation coefficients were computed between the duplicated arrays to assess repeatability. Batch effects associated with use of the Illumina arrays were discovered by visualization and by examining the statistical significance of a variable representing batch, as assessed by ANOVA. In analysis of the association between miRNA levels and breast cancer risk, a total of 246 miRNAs were excluded, as in more than 95% of the 1248 arrays they were below the detection level (P-values > 0.05 compared to background noise, as reported by Genome Studio.)

Linear regression was used to evaluate the association between the expression level of each miRNA (response) and the risk factors (explanatory variables) such as cigarette smoking, BMI, etc., using the control samples only, and adjusted for batch effect by including a factor representing batch as a covariate. In the regression, the miRNA expression level for each sample was the average of two replicates, weighted by the reciprocal of the sample variance between the two replicates.

Linear regression was also used to evaluate the association between the expression level of each miRNA and breast cancer risk, in which the average of two replicated miRNA measures was used as the response and case-control status was the explanatory variable, and the batch factor was entered as a covariate. Subjects (n = 14) with atypical hyperplasia were excluded from this analysis, as atypical hyperplasia itself was strongly associated with breast cancer risk. We computed variances between the two replicates for each sample. Those measurements with larger variances are less reliable; therefore, in the regression analysis for each miRNA, the samples were weighted based on the between-duplicate variance for that miRNA. To avoid certain samples exerting excessively high influence because of very small between-duplicate variance that might have occurred by chance, we assigned the weights in the following way: those samples whose between-duplicate variances were below the median between-duplicate variance received equal weight, at the value of the reciprocal of the median between-duplicate variance; the other half, whose between-duplicate variances were at or above the median, were weighted lower, using the reciprocal of their between-duplicate variance. From this analysis, the top 71 miRNAs with lowest P-values were selected for qPCR validation in the intermediate stage.

A similar analysis was also used to identify the miRNAs associated with atypical hyperplasia, using only the samples from cases. From this analysis, a total of 10 miRNAs with the lowest P-values were added to the intermediate stage. In further analysis, restricted to the 156 case-control pairs that showed good repeatability (r>0.95), paired t-tests were used for each miRNA to test if they were differentially expressed between two groups. We repeated the same analysis after excluding the atypical hyperplasia samples, with 149 matched case-control pairs.

Intermediate stage.

In the intermediate stage, we used linear regression to evaluate whether the miRNA expression levels measured by the Illumina array were consistent with the expression level measured by qPCR on 15 matched case-control samples randomly selected from the 149 pairs with good repeatability. We first regressed the normalized log-intensities on the batch factor first, and took the residual. We then computed the correlation coefficient between the residuals and the qPCR measures, and obtained a P-value for statistical significance by regressing the residuals on the qPCR measure. Based on this analysis, we selected 14 miRNAs for the test stage in the following way. First, we identified 22 miRNAs for which the correlation coefficients were less than -0.25 and the P-values for correlation were below 0.05. From these, we selected the 14 miRNAs that had the greatest fold-changes between cases and controls (after adjusting for batch effect) in the discovery stage. For these 14 miRNAs, the fold-changes in the discovery stage were between 1.12 and 1.20 (in either direction).

Validation stage.

For the analysis of miRNA data from the validation stage, obtained using qPCR, we first used linear regression to determine whether any of the breast cancer risk factors were associated with miRNA expression (average of triplicate measures for each miRNA). We then computed the average miRNA expression fold-changes between the cases and controls and performed paired t-tests on the 165 matched case-control pairs included in the validation stage. In addition, we also tested for case-control differences by using linear regression models on the 318 samples in which miRNA expression (the average of the triplicate measures) was the response and the final disease status of the patient and the assay plates were the independent variables. Other independent variables in the models were the three variables used to match the samples, namely age at BBD diagnosis, age, and duration in the cohort.

All statistical tests were two-sided and all data analyses were performed using R (www.r-project.org).

Association between lifestyle factors, benign breast disease histology, and breast cancer risk.

In the discovery stage of the study, chart data were obtained for 353 cases and 339 controls (constituting 353 case-control pairs because some controls were matched to more than one case). Table 1 shows the association between various factors and risk of invasive breast cancer subsequent to BBD (the associations for the entire cohort have been reported elsewhere [18]). Most of the factors examined were not associated with altered risk. However, there was some suggestion of a reduction in risk of breast cancer in association with a bilateral oophorectomy, and there was a statistically significant 4-fold increase in risk in women who had ever used hormone therapy. Additionally, for the 689 participants that had pathology data, there was a strong positive association between atypical hyperplasia and risk of subsequent breast cancer (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Association between lifestyle factors and breast cancer risk based on subjects included in the discovery stage.

https://doi.org/10.1371/journal.pone.0191814.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Association between benign breast disease histology and breast cancer risk.

https://doi.org/10.1371/journal.pone.0191814.t002

miRNA assays.

All miRNA assays in the discovery stage were done in duplicate. We evaluated the 618 samples (316 case-control pairs) that were processed, and computed the correlation coefficients between the two replicates. The median was 0.973, and 388 (about 63%) of the samples had correlations greater than 0.95. We evaluated the p-values reported by Genome Studio for detection of each miRNA on 1248 arrays. A total of 246 miRNAs were below the detection level (P-value < 0.05) in more than 60 arrays. We excluded these 246 miRNAs in assessing the false discovery rate. We also evaluated the relation between cigarette smoking, BMI, and miRNA expression among the control samples. There were no statistically significant associations.

There was a very strong batch effect in the miRNA assays. S2 Fig illustrates the batch effects by plotting the log-intensity of miRNA HS_10 and miRNA HS_104 for all 618 samples against the order in which they were processed. Each batch had 48 samples. Thirteen batches are separated by the vertical lines in the plot. The top plot shows the result of the first replicate of each sample, and the bottom one shows the results of the second replicate. Given this observation, we corrected for the batch effect of miRNA expression by including batch as a factor in the regression models relating miRNA to breast cancer risk.

miRNA expression and breast cancer risk.

First, for each miRNA, we weighted each sample based on the sample variance between the two replicate measurements for that miRNA. While the samples showing low variance (below the median for all samples) received the same weight, the samples with high variance (at or above the median for all samples) received lower weight (proportional to the reciprocal of the variance). The results showed no signal, and excluding the atypical hyperplasia samples did not improve the signal. Second, we weighted each sample by their repeatability measure (r²)—the samples with r>0.973 received equal weights. This weighting is very light, as the lowest weight is about ⅓ of the highest weight. Among 899 miRNAs that were above the detection level, a total of 114 miRNA were associated with risk of breast cancer at P-values < 0.05 (FDR = 0.40), 79 were associated with risk at P-values < 0.025 (FDR = 0.28), and 47 were associated with risk at P-values < 0.01 (FDR = 0.20). We selected the 71 miRNAs with the smallest P-values for PCR validation.

miRNA association with atypical hyperplasia.

We investigated if any miRNAs showed differences in expression levels between atypical hyperplasia and other samples. We limited our analysis to the cases, and compared the 12 atypical hyperplasia samples with the 304 in the other pathology categories. We selected 10 additional miRNAs with the lowest P-values that were not already in the list of 71 miRNA selected from the case-control analysis (see above) for PCR validation.