Breast cancer DNA methylation profiles are associated with tumor size and alcohol and folate intake

Abstract

Although tumor size and lymph node involvement are the current cornerstones of breast cancer prognosis, they have not been extensively explored in relation to tumor methylation attributes in conjunction with other tumor and patient dietary and hormonal characteristics. Using primary breast tumors from 162 (AJCC stage I-IV) women from the Kaiser Division of Research Pathways Study and the Illumina GoldenGate methylation bead-array platform, we measured 1,413 autosomal CpG loci associated with 773 cancer-related genes and validated select CpG loci with Sequenom EpiTYPER. Tumor grade, size, estrogen and progesterone receptor status, and triple negative status were significantly (Q-values <0.05) associated with altered methylation of 209, 74, 183, 69, and 130 loci, respectively. Unsupervised clustering, using a recursively partitioned mixture model (RPMM), of all autosomal CpG loci revealed eight distinct methylation classes. Methylation class membership was significantly associated with patient race (P<0.02) and tumor size (P<0.001) in univariate tests. Using multinomial logistic regression to adjust for potential confounders, patient age and tumor size, as well as known disease risk factors of alcohol intake and total dietary folate, were all significantly (P<0.0001) associated with methylation class membership. Breast cancer prognostic characteristics and risk-related exposures appear to be associated with gene-specific tumor methylation, as well as overall methylation patterns.

Figure 1. Unsupervised clustering heatmap of CpG methylation in breast carcinomas.

Unsupervised hierarchical clustering heat map based on Manhattan distance and average linkage of the 750 autosomal CpG loci with the highest variance. Samples are in rows (n = 162), and CpG loci are in columns. Blue indicates methylated and yellow indicates unmethylated.

Figure 2. Array methylation is validated by Sequenom EpiTYPER.

Results from GoldenGate array methylation values are plotted versus Sequenom EpiTYPER quantitative methylation values. (A) Sequenom FES methylation is significantly correlated with GoldenGate methylation average β at the coordinate array CpG (Spearman correlation rho = 0.68, P = 1.1E-12, n = 85). (B) Sequenom P2RX7 methylation is significantly correlated with GoldenGate methylation average β at the coordinate array CpG (rho = 0.65, P = 8.6E-12, n = 88). (C) Sequenom HSD17B12 methylation is significantly correlated with GoldenGate methylation average β at the coordinate array CpG (rho = 0.34, P = 5.4E-05, n = 137). (D) Sequenom GSTM2 methylation is significantly correlated with GoldenGate methylation average β at the coordinate array CpG (rho = 0.83, P<2.2E-16, n = 140).

Figure 3. Recursively partitioned mixture model of CpG methylation in breast carcinomas.

The figure depicts the results of RPMM. Columns represent CpG sites and rows represent methylation classes. The height of each row is proportional to the number of observations residing in the class, total n = 162. Blue indicates methylated and yellow indicates unmethylated. Methylation classes are numbered one through eight on the left. The color of the columns within each class represents the average methylation of the CpG for that class.

Figure 4. There is an opposite trend for direction of association between breast carcinoma CpG methylation and alcohol intake compared to folate intake.

P-values for alcohol intake (g/day) and total dietary folate (µg/day) are plotted versus regression coefficients from locus-by-locus analysis of CpG methylation. Horizontal blue dotted line intercepts the y-axis at 0.05 to illustrate significance (before correction for multiple comparisons). The vertical solid black like intercepts the x-axis at zero to illustrate the contrasting trends. (A) There is a trend toward decreased methylation with increasing alcohol intake. (B) There is a trend toward increased methylation with increasing dietary folate.

Figure 5. Probability of methylation class membership is significantly associated with tumor size, patient age, alcohol intake, and dietary folate when controlling for potential confounders in a multinomial logistic regression model.

Results from a multinomial logistic regression plot the probability of methylation class membership versus covariates controlled for age, race, alcohol consumption, total dietary folate, tumor stage (low vs. high), tumor grade, tumor size, estrogen receptor status, and histology. The referent class (methylation class 3) is on the bottom of the plot in blue-green, remaining classes are plotted in numeric order from bottom to top as shown in the legend. (A) Patient age is significantly associated with methylation class membership (Overall Wald P<0.0001), and all methylation classes except class 4 are individually significantly associated with patient age. (B) Alcohol consumption is significantly associated with methylation class membership (Wald P<0.0001), and methylation classes 2, 4, 5, and 8 are individually significantly associated with alcohol intake. (C) √Total dietary folate intake is significantly associated with methylation class membership (Wald P<0.0001), and all methylation classes are individually significantly associated with total dietary folate. (D) Tumor size is significantly associated with methylation class membership (Wald P<0.0001), and all methylation classes except class 4 are individually significantly associated with tumor size.