Integrative comparison of mRNA expression patterns in breast cancers from Caucasian and Asian Americans with implications for precision medicine

Abstract

Asian Americans (AS) have significantly lower incidence and mortality rates of breast cancer (BRCA) than Caucasian Americans (CA). While this racial disparity has been documented the underlying pathogenetic factors explaining it are obscure. We addressed this issue by an integrative genomics approach to compare mRNA expression between AS and CA cases of BRCA. RNA-seq data from the Cancer Genome Atlas showed that mRNA expression revealed significant differences at gene and pathway levels. Increased susceptibility and severity in CA patients were likely the result of synergistic environmental and genetic risk factors, with arachidonic acid metabolism and PPAR signaling pathways implicated in linking environmental and genetic factors. An analysis that also added eQTL data from the Genotype-Tissue Expression Project and single nucleotide polymorphism (SNP) data from the 1000 Genomes Project identified several SNPs associated with differentially expressed genes. Overall, the associations we identified may enable a more focused study of genotypic differences that may help explain the disparity in BRCA incidence and mortality rates in CA and AS populations and inform precision medicine.

Introduction

In 2016, an estimated 246,660 new breast cancer (BRCA) cases will be diagnosed in the United States (1). Despite advances in treatment and earlier detection, 40,880 will die from this disease (1). Breast cancer is a very heterogeneous disease. Understanding its heterogeneity is the key for developing personalized treatment strategies and an inevitable step towards the complete cure of the disease. In clinical practice, BRCA is divided into subgroups based on the expression of the estrogen receptor (ER), progesterone receptor (PR), and the status of gene amplification of human epidermal growth factor 2 receptor (HER2). The four predominant subtypes reported extensively in literature are categorized based on these receptors: luminal A (ER+ and/or PR+, HER2−), luminal B (ER+ and/or PR+, HER2+), HER2 (ER−, PR−, HER2+), and triple negative (ER−, PR−, HER2−). A subtype of triple negative breast cancer (TNBC) is the basal or basal-like phenotype that is negative for all three receptors but expresses specific basal markers (cytokeratin 5/6, cytokeratin 14, cytokeratin 17, and epidermal growth factor receptor) (2). This phenotype is more likely to undergo metastasis and is associated with poorer prognosis (3). Recent large-scale genomic studies have suggested that the whole spectrum of breast cancer subtypes can be much more complex than the four commonly known subtypes (4,5).

The incidence and death rates of breast cancer differ among different race groups. For example, age-standardized breast cancer incidence and mortality rates for Asian American (AS) are 30% and 50% lower, respectively, than those in non-Hispanic whites (1). Various aspects of this health disparity have been studied in the past with qualitatively similar observations (6-8). Asian American are the fastest-growing racial/ethnic group in the US, representing 6.3% of the population (20.0 million/318.7 million) in 2014 (9). Breast cancer is the most commonly diagnosed cancer and the second leading cause of cancer death among Asian American women, with a total of 11,090 new invasive cases and 1,180 deaths expected to occur in 2016 (1). There is a substantial variation in breast cancer occurrence within the Asian population, with lower rates among groups that have immigrated more recently (1,8). These differences are thought to be related to the extent of adoption of western behaviors that increase breast cancer risk, such as a later age at childbirth, fewer births, and higher body weight (10). Despite the differences in socioeconomic status and life styles, genetic factors likely play roles in the health disparity between the two racial groups (1,11). The identities of these genetic factors, however, are poorly understood. Consequently, the molecular basis, if there is any, for the change in incidence rate after immigration has been unknown. Such molecular level understanding may shed light on better understanding of tumorigenesis of BRCA and help develop personalized treatments.

A recent study has compared the gene and miRNA expression between a Chinese breast cancer population and an Italian cohort (12). The study did not detect significant differences between the two patient groups. The study then concluded that “Transcriptional similarity across transethnic cohorts may simplify translational medicine approaches and clinical management of breast cancer patients worldwide.” In a recent study, we compared the expression of genes and transcripts between Caucasian American and African American breast cancer patients and identified hundreds of differentially expressed genes and several interesting differentially regulated pathways between the two patient populations (13). The study identified a novel non-coding RNA, which could be useful for early detection of BRCA (13). We believe a similar study comparing CA and AS BRCA patients may also offer some clues on the biological factors causing the health disparity and better understanding of the biology of BRCA.

We used next generation sequencing (NGS) data from the Cancer Genome Atlas (TCGA) to determine differentially expressed genes and other transcripts between a large number of age- and stage-matched AS and CA patient primary breast tumor samples. Using NGS data we can determine differential expression between non-coding RNAs in addition to coding RNAs. This study represented, to our knowledge, the first time that NGS data was used in combination with a large patient cohort to investigate expression differences of genes and other transcripts between AS and CA BRCA patients. And this work represented the largest and most comprehensive genomic study on health disparity between CA and AS BRCA to date, where differentially expressed genes by stage and subtype between AS and CA tumors were identified for the first time, in addition to differentially expressed subnetworks and pathways.

Generally speaking, three types of genetic factors can be produced in a genome wide comparison: (1) genetic factors that are truly associated with cancer health disparity, which we call category one factors; (2) genetic factors with significant differences between two different racial populations in general, unrelated to the cancer health disparity (category two factors); and (3) genetic factors that appear simply by chance due to multiple comparisons (category three factors). Gene level comparison will generally produce differentially expressed genes/transcripts in all these three categories. Category three factors can be removed by correcting for multiple comparisons using more stringent criteria, with the cost of also removing some of the category one factors (an issue of balancing type I and type II errors). To discern category two factors, we performed various analyses beyond individual gene level, such as gene set enrichment analysis, pathway enrichment analysis and network-based analysis. The differentially regulated gene sets/pathways/subnetworks were then examined in the context of literature to identify the category one factors.

Among the category one factors, some are “drivers” that are upstream regulators causing health disparity, and some are “passengers” that are regulated by those drivers. Pathway and network-based analyses also help pinpoint the driver genetic factors and understand their mechanisms of actions.

We found from this study that there are widespread differences in the expression of genes and other transcripts between AS and CA primary breast tumor tissues. The health disparity is likely caused by an aggregated effect of all the associated biological factors (together with environmental factors), with some favoring one race and others favoring the other. The simple picture, where only one or a few genetic factors can explain the biological causes of the health disparity, may not exist. Both environmental and genetic factors work cooperatively to make CA population more breast cancer prone and also make the cancer of CA patients more severe. Specifically, arachidonic acid metabolism and peroxisome proliferator-activated receptor (PPAR) signaling pathways may play central roles of linking environmental and genetic factors. Incorporation of eQTL and SNP data may help find SNPs associated with the differentially expressed genes and shed light on the genotype differences that drive the observed health disparity. Comparisons of mRNA expression of cancer between two races (or among multiple races) may offer a unique angle and some unique advantages for studying cancer biology and cancer heterogeneity compared to existing methods.

Materials and Methods

RNA-Seq Data from TCGA

Data used in this study were downloaded from Insilicom's BioKDE platform (insilicom.com), where processed, de-identified genomic and clinical data were downloaded from TCGA and integrated into a relational database for convenient queries. The de-identified patient data included up to 20483 RNA sequence-derived gene expression values and clinical characteristics (e.g., age, cancer stage, receptor status). Fifty-seven AS tumor samples were matched by age and stage with 574 CA tumor samples, and a summary of the clinical characteristics of these patients is presented in Table 1. Subtypes were assigned to patient samples following Ref (13). The raw gene expression counts were normalized with RPM (reads per million), unless otherwise noted.

Table 1.

Breast cancer patient clinical data from TCGA.

Characteristics	Caucasian American (n=728)		Asian American (n=57)		Fisher's Exact Test
	Number	Pct.	Number	Pct.	P-value
Age					0.013
< 50	197	27.1%	23	40.4%
50-64	311	42.7%	26	45.6%
65+	220	30.2%	8	14.0%
Tumor Stage					0.17
0	1	0.137%	0	0
1	134	18.4%	4	7.02%
2	409	56.2%	41	71.93%
3	165	22.7%	12	21.05%
4	11	1.51%	0	0
Unknown	8	1.10%	0	0
Tumor Type					0.0084
Luminal A	298	40.9%	21	36.84%
Luminal B	77	10.6%	5	8.77%
Triple Negative	67	9.20%	8	14.04%
HER2 Type	19	2.61%	7	12.28%
Unknown	267	36.7%	16	28.07%
Vital Status					0.022
Living	648	89.7%	56	98.2%
Deceased	80	10.3%	1	1.75%

Differential Gene Expression

There are many methods for differential gene expression analysis based on different distribution assumptions (14-17). We chose to use DESeq2, an R Bioconductor package, for our analyses (18,19). Patient gene counts input into DESeq2 were rounded to the nearest whole number but not normalized. Jobs were run on the Insilicom BioKDE server (insilicom.com). All the comparisons are AS vs CA, where increased expression (up-regulation) means the expression in AS BRCA is higher than the expression in CA BRCA, and decreased expression (down-regulation) means the expression in AS patients is lower than the expression in CA BRCA.

Sample Matching

All the comparisons done in this study were matched by age and stage using R package, Matching (20). The matching was done differently than in our earlier study (13). The stage was matched exactly and age was matched in a way that any patients are considered matched if the difference of their ages is smaller or equal to 10. This is more flexible than the three categories used in the earlier study. The Matching package also performs optimal matching once the ratio of the two types of subject are determined. To find the optimal ratio, we maximize the quantity, nm/(n+m), where n is the number of patients of one race and m is the number of patients of the other race. In this setup, we assume the power calculation of testing differential gene expression can be approximated with a two sample test of two normal samples with different sample sizes. We found that the new matching method produced notable differences, but its exact effect was not evaluated systematically.

Pathway and Network Analysis

Differentially expressed genes identified by DESeq2 were input into the server at humanmine.org (21). A separate analysis was also performed using Insilicom's Integrative Genomic Analysis pipeline. All the p-values were corrected for multiple comparisons.

Gene Expression Network Analysis (GXNA) was used to identify subnetworks where differentially expressed genes are enriched compared to other parts of the whole gene/protein interaction network (22). Gene expression data input into GXNA was first log₂-transformed so that differences in means corresponded to fold-change, and default GXNA parameters were used. STRING, a database of protein-protein-interactions was used to visualize GXNA results (23).

Pathway visualization

The visualizations of differentially regulated pathways were made possible using the Pathview package in R (24). It allows users to easily spot up or down regulated genes in the pathway to understand how it is differentially regulated. The workflow of our differential gene expression analysis is given in Supplementary materials (Figure S1).

Association of SNPs to gene expression in breast cancer

Combining data obtained from GTEx (Genotype-Tissue Expression), eQTL (Expression Quantitative Trait Loci) studies, and the 1000 Genomes project with the gene expression data from TCGA, we infer potential associations of certain SNPs to gene expressions in breast cancer. Firstly, given a gene of interest, SNPs that are associated with expression of this gene are obtained from GTEx data portal for eQTL (http://www.gtexportal.org/home/eqtl), regardless of tissue types; secondly, the SNPs are used to compute the allele frequencies for races of interest using data from 1000 Genomes project (http://www.1000genomes.org/); thirdly, the effect sizes from eQTL and the allele frequencies are used to compute relative expressions of the gene (marginal expression without conditioning on other factors); finally, these relative expressions computed from allele frequencies and eQTL effect sizes are sorted to compare with the actual expressions observed in TCGA data.

Correlation analysis

For continuous numerical clinical features, Spearman's rank correlation coefficients and two-tailed P values were estimated using 'cor.test' function in R. For categories of continuous type of clinical data, Wilcoxon rank sum test was applied to compare their mean difference using 'wilcox.test (continuous.clinical ~ as.factor(group), exact=FALSE)' function in R. This test is equivalent to the Mann-Whitney test.

Results

Table 1 shows the breakdown of the demographics of all the patients obtained from TCGA. CA women have a mortality rate of 10.3%, while AS counterpart has a mortality rate of only 1.75% (fisher's exact test p-value is 0.022). When matching age and stage, all the AS patients were kept and 399 CA patients were selected. The mortality rate of CA reduced to 8.77%. The fisher's exact test (p-value = 0.044) is still significant. As mentioned before, this disparity is likely caused by both socioeconomic and biological factors. Information regarding socioeconomic status or access to care is not available at TCGA to assess the effects of these factors to the health disparity.

The outline of the analyses we performed in this study is shown in Figure 1. Firstly, differential expression analysis at the individual gene level was conducted to identify genes and other transcripts that differ between CA and AS BRCA patients using DESeq2; secondly, pathway analysis and gene set enrichment analysis were performed using the server at humanmine.org and using R Pathview package; thirdly, Gene Expression Network Analysis (GXNA) was used to identify differentially expressed subnetworks; finally, patients were stratified to study differences in stage- or subtype-specific gene expression again with DESeq2 (Figure 1).

Figure 1.

Differential gene expression in PPAR signaling pathway. Green color indicates log2 fold changes of down-regulated genes in AS BRCA patients and red color indicates log2 fold changes of up-regulated genes in AS BRCA patients.

In total, 927 unique genes and other transcripts were shown to be differentially expressed (adjusted p-value <= 0.05 and fold change >= 2) between the various comparisons performed (Supplementary Material Table S1-S8). To identify overall trends in gene expression, primary breast tumor expression data between AS and CA patients irrespective of subtype or stage (i.e., the “overall” comparison) was first compared. Two hundred eighty four genes and other transcripts, 1.4% of the 20,483 genes and other transcripts assessed, were identified by DESeq2 as differentially expressed between AS and CA primary breast tumors (adjusted p-value <= 0.05 and fold change >= 2, Supplementary material Table S1). Forty seven of these genes and other transcripts showed increased expression in AS tumors while 236 were decreased. Selected genes and other transcripts with high fold-change and relevance to cancer are shown in Table 2.

Table 2.

Differentially expressed genes and transcripts between CA and AS breast cancer from overall compassion, with age and stage matched.

Gene	Log2 Fold-Change	P-adj	Relevance
PCSK1	−1.96	1.01e-10	Associated with Obesity a known risk factor for breast cancer.
CIDEC	−1.84	1.83e-9	Expression correlates with BMI.
GLYAT	−1.8	2.89e-8	Associated with lean body mass.
SHC4	−1.77	1.74e-9	SHC4 signaling adaptor facilitates ligand-independent phosphorylation of the EGF receptor.
OR3A2	−1.75	4.5e-9	Upregulated in tobacco smokers. A breast cancer risk factor.
LEP	−1.72	6.75e-8	Associated with breast cancer risk, particularly in obese women. Plays a major role in the regulation of body weight.
PIWIL2	−1.71	2.71e-12	The combination of piR-932 and PIWIL2 may be a positive regulator in the process of breast cancer stem cells, and they both could be the potential targets for blocking the metastasis of breast cancer.
MGAM	−1.69	4.13e-9	Over-expressed in oral squamous cell carcinoma.
SLC19A3	−1.63	4.13e-9	Upregulation occurs to maintain nutrient uptake in hypoxic regions of tumors.
TDRD12	−1.63	5.07e-7	Hypomethylated in adenoid cystic carcinoma.
LY6D	−1.63	9.51e-7	Increased expression associated with poor outcomes in various cancers, including breast cancer.
SLC4A4	−1.59	5.16e-9	Regulates PH in hypoxic and acidic tumor environments, aiding in tumor growth and metastasis. A potential drug target.
PSG4	−1.59	2.4e-6	Pregnancy specific glycoprotein. Expressed in pregnant women. Hormonal changes in pregnancy could be associated with breast cancer.
C8A	1.62	2.35e-7	Potential circulating biomarker for hepatocellular carcinoma.
RETN	2.05	1.62e-11	Differentially expressed in several studies investigating cancer health disparity
UTS2	2.23	2.76e-14	There is strong evidence that suggest urotensin-II as the significant contributor of angiogenesis as well as cell proliferation and tumor biology.

The 284 significantly differentially expressed genes and other transcripts from the overall comparison were queried at humanmine.org. Humanmine.org has several widgets analyzing the given gene sets in terms of their enrichment in gene ontology (GO) term, pathways, protein domains, interactions, publications, and interaction networks. In this study, we looked at pathway, GO terms and differentially regulated network modules. In pathway/GO enrichment analysis, genes were placed into known pathways/GO terms and ranked by the probability that the associated genes are enriched in the input gene list, and a low P-value indicates a higher enrichment and greater chance of differential regulation. The significantly differentially regulated pathways (Table 3) are discussed in more details below.

Table 3.

Differentially regulated pathways between AS and CA BRCA.

Pathway Name	Representative Genes and Transcripts	P-adj
PPAR signaling pathway	UCP1, SLC27A6, PCK1, CD36, AQP7, PLIN1, FABP4, CYP8B1, ADIPOQ, PLIN4	3.88e-4
Neuroactive ligand-receptor interaction	GLP2R, CCKBR, GLP1R, CRHR1, CHRM3, LHCGR, AGTR1, DRD2, GRIA1, CHRNB2, NPY5R, GRIK1, LEP, GPR156, NPY4R	7.0e-03
Arachidonic acid metabolism	PTGS2, GGT1, PTGDS, PLA2G4A, PTGIS, GPX3	1.07e-04
Long-term potentiation	PRKACB, ITPR1, PRKCA, GRIA1, GRIN2A	6.20e-03
Pathways in cancer	PTGS2, NFKB2, RAC1, IL6, TP53, ERBB2, PRKCA, VEGFD, CTNNB1	6.59e-03
VEGF signaling pathway	PTGS2, PLA2G4A, RAC1, PRKCA, SRC	9.17e-03

Peroxisome proliferator-activated receptors (PPAR) signaling pathway

The top significantly differentially regulated KEGG pathway is PPAR signaling pathway (adjusted p-value = 3.88e-4) with 10 genes significantly differentially expressed. PPARs are a group of nuclear receptor proteins that function as transcription factors that play essential roles in the regulation of cellular differentiation, development, and metabolism (carbohydrate, lipid, protein), and tumorigenesis of higher organisms (25,26). Three types of PPARs have been identified: alpha, gamma, and delta (beta), where PPARgamma is expressed in virtually all tissues. Endogenous ligands for the PPARs include free fatty acids and eicosanoids. The 10 genes in the pathway identified by the analysis are all significantly down-regulated (fold change >= 2 and adjusted p-value <=0.05) in AS BRCA tissues, indicating PPAR pathway is down-regulated in AS BRCA patients compared to CA patients. The PPAR pathway and the differential expression of associated genes drawn by Pathview can be visualized in Figure 1.

The other significant KEGG pathway was neuroactive ligand-receptor interaction (adjusted p-value = 6.99e-3) with 15 genes in the list. This pathway does not appear to be closely related to breast cancer. It may be a pathway that is differentially regulated between AS and CA population in general (category two factors).

GXNA was then used to identify differentially expressed subnetworks between AS and CA tumors (13,22). GXNA takes patient expression data as input and then outputs a ranked list of subnetworks that contain interactions involving the differentially expressed genes. GXNA identified a large connected subnetwork with 40 genes. Most of the genes in this subnetwork are down-regulated in AS BRCA patients. It contains some well-known cancer-related genes such as TP53 and ERBB2 (HER2). The network drawing tool at STRING database (23) was used to visualize the GXNA network and to show potential gene product interactions beyond those included in the GXNA interaction file (Figure 2).

Figure 2.

The differentially expressed subnetwork obtained by GXNA.

The 40 genes in the GXNA network were then used as input for humanmine.org server to identify enriched pathways (Table 3). Four pathways were significant (using 0.01 as cutoff for adjusted p-values): Arachidonic acid metabolism pathway (p-value = 1.074-4, 6 genes), long-term potentiation (0.0062, 5 genes), pathways in cancer (p-value = 0.0066, 9 genes), and VEGF signaling pathway (p-value = 0.0092, 5 genes). Among them, long-term potentiation was not found to be related to breast cancer and may be a category two pathway.

Arachidonic acid metabolism pathway

The arachidonic acid pathway constitutes one of the main mechanisms for the production of pain and inflammation, as well as controlling homeostatic function. Pathview diagram with differentially expressed genes colored for arachidonic acid metabolism pathway is shown in Figure 3.

Figure 3.

Differential gene expression in Arachidonic acid metabolism pathway. Green color indicates down-regulated genes in AS BRCA patients and red color indicates up-regulated genes in AS BRCA patients.

Pathways in cancer

Several cancer related pathways are differentially regulated between AS and CA BRCA patients (adjusted p-value = 6.59e-03), which can be seen from the Pathview diagram for pathways in cancer (Supplementary materials, Figure S2). Overall, through visual inspection, we can see that substantially more genes are down-regulated in AS patients than up-regulated, indicating overall the cancers in AS patients are probably not as severe as those in CA patients. This is consistent with the fact that the death rate for breast cancer is much lower for AS than CA.

Vascular endothelial growth factor (VEGF) signaling pathway

The VEGF signaling pathway is also down-regulated significantly in AS BRCA patients (adjusted p-value = 9.17e-03). The VEGF signaling pathway regulates vascular development in the embryo (vasculogenesis) and new blood vessel formation (angiogenesis). Upregulation of angiogenesis is a key step in sustained tumor growth and may also be critical for tumor metastasis (27,28).

Stage-matched and then subtype-matched AS and CA tumors were next compared (Supplementary materials, Table S3-S8). Due to sample size limit, we pooled stage I and II patients as early stage, and stage III and IV as late stage. Selected results from these comparisons are summarized in Table S2 in Supplementary materials. Interestingly, a small number of genes and other transcripts (totally 23) overlapped between early and late stage comparisons (Supplementary materials, Figure S3). Overlapping among overall comparison and subtype-specific comparison is also poor, similar to what was found in an earlier study comparing Caucasian American with African American breast cancer patients (13), indicating the differences within subtypes are likely specific to each subtype and in depth study may reveal more interesting biology of cancer health disparity between the two racial groups. The only gene that is differentially expressed in all the comparisons (overall, early stage, late stage, and subtype-specific comparisons) is DDX11L2, whose biological significance is unclear.

In summary, 267 genes and other transcripts were differentially expressed in the early stage comparison (228 decreased and 39 increased in AS BRCA patients), 137 genes and other transcripts were differentially expressed in the late stage comparison (94 decreased and 43 increased), 507 genes and other transcripts were differentially expressed in luminal A comparison (439 decreased and 68 increased), 49 genes and other transcripts were differentially expressed in HER2 comparison (34 decreased and 15 increased), and 104 genes and other transcripts were differentially expressed in triple-negative comparison (10 decreased and 94 increased). An interesting observation is that for all the comparisons the numbers of down-regulated genes (decreased) are significantly higher than those up-regulated in AS BRCA patients, except for triple-negative subtype, for which the number of up-regulated genes is significantly higher than those down-regulated.

Single nucleotide polymorphisms (SNPs) associated with gene expression in breast cancer

A fundamental question in disease studies is to understand how genotypes affect phenotypes of the disease of interest. We combined gene expression data with eQTL data from GTEx project and SNP allele frequency information from 1000 Genomes project (see Methods for details). We use gene CYP1A1 to illustrate our approach. Since the expression of CYP1A1 is linked to ER (estrogen receptor) (29) and AA (arachidonic acid) metabolism (30), the variants of CYP1A1 identified in this study may be associated with tumorigenesis of breast cancer through metabolism of ER and AA. This analysis not only will help us to identify disease associated genomic variants, but also point to plausible mechanisms for the associations.

We found one variant (ID: rs55646017) that is associated with expression of CYP1A1 with an effect size of 0.36 and a p-value of 1.5e-5. The reference allele is T and alternative allele is C. The frequencies of alternative allele in South Asian population, East Asian population, and European population are 0.29, 0.23 and 0.49, respectively. This agrees qualitatively with the observed expression difference between AS and CA BRCA patients (down-regulation in AS with a log2 fold change of −0.92). This variant may be a promising candidate in follow-up association studies for breast cancer. This approach can be applied to other differentially expressed genes to provide both additional evidences and mechanism explanations for potential associations.

Survival analysis for the differentially expressed genes

Survival analysis results for the 284 differentially expressed genes from the overall comparison were extracted from Broad GDAC of Broad Institute (http://gdac.broadinstitute.org/). We also performed survival analysis using the same approach (logrank test in univariate Cox regression analysis with proportional hazards model) and obtained very similar results (Supplementary materials). Totally, 28 genes have p-values smaller than 0.05 (Supplementary materials). Among them, PTGS2, has a logrank test p-value of 0.026. PTGS2 encodes the COX-2 enzyme, which catalyzes the conversion of arachidonic acid to prostaglandins (PGs) and other eicosanoids. Expression of PTGS2 is also significantly correlated with pathological stage (p-value = 4.40e-05) and pathological T stage (p-value = 0.00011). PTGS2 expression, which is undetectable in most normal tissues, is induced in response to hypoxia, inflammatory cytokines, tumor promoters, growth factors and other stressors (31). PTGS2 is involved in carcinogenesis, immune response suppression, inhibition of apoptosis, angiogenesis and tumor cell invasion and metastasis. Recent studies have indicated that PTGS2 genetic variation is associated with breast cancer susceptibility (32,33). Furthermore, overexpression of PTGS2 in patients with breast cancer is associated with a worse prognosis (34).

Correlation analysis of differentially expressed genes with clinical variables

The correlations between clinical variables and the differentially expressed genes were also extracted from Broad GDAC of Broad Institute (http://gdac.broadinstitute.org/). We examined the following clinical variables: pathological stage, pathological T stage, pathological N stage, pathological M stage, and the number of lymph nodes. Overall, 129 genes are significantly associated with at least one clinical variable (including survival time). Several genes are highly correlated with the stage of breast cancer: MMRN1 (p-value = 1.83e-07), PLIN1 (p-value = 2.60e-05), PTGS2 (p-value = 4.40e-05), and TMEM132C (p-value = 1.67e-05). Two genes are correlated with metastasis (M-stage): KLK5 (p-value = 0.0074) and PTGS2 (p-value = 0.00082). The detailed result is included in Supplementary materials.

Discussion

The heterogeneity of breast cancer (BRCA) is manifested in its racial disparities, as in many other aspects of the disease. Racial disparities may provide a unique angle to investigate the heterogeneity of BRCA. In this study, we compared mRNA expression between Caucasian American (CA) and Asian American (AS) breast cancer (BRCA) primary tumor samples using RNAseq data obtained from TCGA. We matched the patients on age and stage and performed overall, early stage, late stage, and subtype specific comparisons. We have identified a large number of differentially expressed genes and other transcripts for each comparison. We focused on the analysis of overall comparison in this study. The results from other comparisons also deserve more detailed analysis and interpretation, especially the comparison of triple negative patients, which showed unique expression patterns compared to other comparisons. In the comparison of triple negative subtype, AS triple-negative BRCA patients have significantly larger number of up-regulated genes than the CA counterpart while in other comparisons AS patients have significantly lower number of up-regulated genes than down-regulated genes. In the discussion below, down-regulation or up-regulation refers to the regulation changes in AS breast cancer. Pathway and network analysis using significantly differentially expressed genes revealed several differentially regulated pathways between the two racial groups, namely, PPAR signaling pathway, arachidonic acid (AA) metabolism pathway, pathways in cancer, and VEGF pathway.

The biological factors associated with the breast cancer health disparity between Asian and Caucasian Americans

The difference in incidence and mortality rate of BRCA between AS and CA populations is likely the aggregated effect of multiple factors, including genetic, lifestyle, and environmental factors, with some favoring one race and others favoring the other. The lower incidence and mortality rate of breast cancer in AS population is likely because the factors favoring AS outweigh those favoring CA. With a set of carefully chosen and optimized analysis methods, our study revealed a complex picture with interconnected genetic and environmental factors leading to a plausible explanation for the health disparity between AS and CA populations.

We first list the observations/findings made in previous studies (1,8,29,35-41), followed by the evidences we have collected in this study before drawing the hypotheses.

Observations/findings made in previous studies:

1. Asian Americans have significantly lower incidence and mortality rate of breast cancer than Caucasian Americans (1).

2. The incidence rate of AS increases with the time they live in US (1,8).

3. Unhealthy diet with more meats leads to more consumption of arachidonic acid (35). US have more meat consumption than Asian populations (36).

4. A variant (SNP) on PPARgamma is associated with severe obesity in Chinese population (37).

5. Inhibitors of the 20-HETE-producing enzymes of the Cytochrome (CYP450)4A and CYP4F families can block the proliferation of breast cancer cell lines (38). Many variants of CYP family of genes are known, which are linked to different activities toward their substrates (39).

6. PLIN1 (perilipin) positivity was found to be an independent poor prognostic factor for metastatic breast cancer (40).

7. Metabolic activation of carcinogenic estrogens plays a critical role in estrogen-induced carcinogenesis (41). Estrogen is metabolized in breast tissue by the product of CYP1A1 gene (29).

8. PTGS2 (COX-2) is involved in carcinogenesis, immune response suppression, inhibition of apoptosis, angiogenesis and tumor cell invasion and metastasis (31). Recent studies have indicated that PTGS2 genetic variation is associated with breast cancer susceptibility (32,33). Furthermore, overexpression of PTGS2 in patients with breast cancer is associated with a worse prognosis (34).

Evidences obtained in this study (down- or up-regulation refers to the expression changes in AS breast cancer):

1. Arachidonic acid metabolism pathway is differentially regulated between AS and CA BRCA with some key genes down-regulated.

2. PPAR signaling pathway is significantly down-regulated. A key down-regulated gene is PPARgamma.

3. Cancer related pathways activated by PPAR signaling pathway are also down-regulated, such as VEGF signaling and WNT signaling pathways (Figure S2).

4. CYP4F, the cytochrome P450 gene that metabolize arachidonic acid into 20-HETE, a carcinogen, is down-regulated. The other two genes in this pathway, CYP4A and CYP2U, are marginally down-regulated.

5. PLIN1, a downstream gene of PPARgamma, is down-regulated. PLIN1 is also strongly correlated with the stage of breast cancer (p-value = 2.60e-05).

6. LEP (leptin) and IL6 (Interleukin-6) are both down-regulated. These adipokines are both secreted by adipocytes and can act in either autocrine or paracrine manners to increase production of aromatase, which is directly related to increased synthesis of estrogen (42).

7. CYP19A1 (aromatase), the enzyme for synthesizing estrogen, is down-regulated.

8. CYP1A1, the enzyme for metabolizing estrogen in breast tissues (29), is down-regulated.

9. PTGS2 is down-regulated in AS BRCA patients. Survival analysis showed that PTGS2 is marginally significant for survival. Correlation analysis showed that PTGS2 is significantly correlated with pathological stage and T-stage of breast cancer.

The combined discoveries in both previous and current studies lead us to the hypothesis that breast cancer health disparity between CA and AS populations is driven by the interaction between environmental (lifestyle) and a few specific genetic factors (Figure 4). The fact that the incidence rate for AS increases with the time lived in the US suggests a role the environmental factor plays in this health disparity. Asians tend to have a healthier diet with less meat consumption compared to Americans. Asians are less likely to suffer from obesity than Caucasian Americans. The average BMI is also lower in the Asian population. The down-regulation of many genes in the AA metabolism pathway in AS BRCA (Figure 3) may be due to the less AA content in their tumor tissues (and possibly normal tissues as well). The genetic variants in CYP family genes metabolizing AA and its metabolites may also play a role in the differential regulation of the AA metabolism pathway (39,43,44). Recent studies have suggested that AA and its metabolite pathway play a critical role in the development and progression of breast cancer (45-47). It was also suggested that AA-activated mTOR signaling also plays critical roles in the angiogenesis and tumorigenesis of breast cancer (48). Taken together, AA metabolism may play a key role in the breast cancer health disparity between AS and CA populations. More experimental studies would be required to elucidate how genetic and environmental factors cooperatively affect AA metabolism and its downstream pathways.

Figure 4.

The environmental and genetic factors that may explain the health disparity between Caucasian American and Asian American breast cancer and a potential link between obesity and breast cancer risk. (1) Unhealthy diets cause higher AA (arachidonic acid) content in the body, which cause the activation or up-regulation of several pathways including AA metabolism, VEGF signaling, and PPAR signaling, which are related to cancer development and migration. (2) Known genetic variants in CYPs may cause the differential expression of these genes and different metabolic patterns of AA, which, in turn affect AA content and AA metabolism pathway. (3) A PPARγ variant exists in Chinese population that is associated with severe obesity. The variant may also exist in people in western countries. The variant may cause the differential regulation of PPAR signaling pathway in the two race groups. (4) Both estrogen production and metabolism are important for breast cancer risk and genetic differences between AS and CA population may affect the estrogen metabolism, which in turn, may have affected their breast cancer incidence rate.

A PPARgamma variant was found to be associated with severe obesity in the Chinese population (37). The variant may also exist in the CA population and be associated with obesity in that population. PPARgamma and its downstream gene, PLIN1 (perilipin), are all down-regulated in AS (Figure 1). Genetic factors, such as PPARgamma variant, may also play a role in the health disparity.

A well-known link between obesity and breast cancer risk is estrogen (49). However, the molecular level mechanism for the association has been poorly understood. We found that estrogen signaling pathway was not significantly differentially regulated between AS and CA breast tumors, indicating the mechanism for the estrogen linkage may not be a simple one. It has been shown in a previous study that metabolic activation of carcinogenic estrogens plays a critical role in estrogen-induced carcinogenesis (41). Estrogen is metabolized in breast tissue by CYP1A1 (29). We found that CYP1A1 is down-regulated in AS BRCA patients, which may have also contributed to the BRCA health disparity. This observation suggests that among obese women, those with a CYP1A1 variant that metabolizes estrogen more effectively may have higher risk of breast cancer than those with a CYP1A1 variant that metabolizes estrogen poorly. This hypothesis can be tested in future studies. The lack of BMI and obesity information in TCGA makes a more in depth study on the link between obesity and breast cancer risk infeasible. Future research using datasets with such information will likely deepen our understanding in this important public health issue.

The Asian population represents 60% of the human population worldwide. A comparison between Asian and Caucasian Americans could also help researchers to understand the genetic differences between Caucasian and Asian populations and how such differences may affect their breast cancer risk and responses to treatment. Several of our discoveries have implications for drug discovery and development of diagnostic and prognostic tools. We discuss two of them below.

Arachidonic acid metabolism or PPAR signaling pathway as potential drug targets

Both the AA metabolism and PPAR signaling pathways are significantly differentially regulated between AS and CA breast cancer. The most notable changes are some down-regulated genes in both pathways in AS breast cancer. Clinical trials for drugs targeting those genes/proteins that are significantly differentially expressed between breast cancer cells of the two racial groups would likely see different outcomes depending upon the racial makeup of the subjects. These clinical trials should consider race as a confounding factor. It may also be beneficial to have a personalized medicine component which takes the expressions of the drug targets or those associated variants into account.

PLIN1 as a prognostic factor

In a previous study, PLIN1 positivity was found to be an independent poor prognostic factor for metastatic breast cancer (40). To use PLIN1 as a biomarker in clinical practice, it needs to be validated in a real clinical setting. Due to the large expression difference of PLIN1 between the breast cancers of CA and AS populations, the chances for it to pass clinical validation may be quite different in western and Asian countries. Again, clinical trials involving PLIN1 may benefit from having a personalized medicine component which takes the expressions of the biomarker or those associated variants into account.

In summary, our study revealed wide-spread transcriptome differences between Caucasian and Asian American breast cancer and discovered a few promising candidates for future studies aiming at developing more personalized strategies for breast cancer.