A systematic review of datasets that can help elucidate relationships among gene expression, race, and immunohistochemistry-defined subtypes in breast cancer

ABSTRACT

Scholarly requirements have led to a massive increase of transcriptomic data in the public domain, with millions of samples available for secondary research. We identified gene-expression datasets representing 10,214 breast-cancer patients in public databases. We focused on datasets that included patient metadata on race and/or immunohistochemistry (IHC) profiling of the ER, PR, and HER-2 proteins. This review provides a summary of these datasets and describes findings from 32 research articles associated with the datasets. These studies have helped to elucidate relationships between IHC, race, and/or treatment options, as well as relationships between IHC status and the breast-cancer intrinsic subtypes. We have also identified broad themes across the analysis methodologies used in these studies, including breast cancer subtyping, deriving predictive biomarkers, identifying differentially expressed genes, and optimizing data processing. Finally, we discuss limitations of prior work and recommend future directions for reusing these datasets in secondary analyses.

KEYWORDS:

• Breast cancer
• gene-expression profiling
• race
• immunohistochemistry status
• disease subtypes
• triple negative breast cancer
• health disparities

Introduction

In clinical practice, breast tumors are classified into histopathological subtypes based on immunohistochemistry (IHC) profiles. These classifications are useful for determining a patient’s diagnosis, estimating a patient’s prognosis,¹ as well as determining appropriate patient-management strategies and therapies.^2,3 The IHC subtypes are defined by the combined expression of three cell-surface proteins: estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER-2).^4,5 These cell surface receptors, which are also referred to as IHC markers, are defined as positive or negative, depending on the level of expression of the receptors. While ER and PR are bound by estrogen and progesterone, respectively, there are no known ligands for HER-2, although it can be transactivated by epidermal growth factor (EGF)-like ligands.⁶ The activation of these cell surface receptors by the binding of ligands, results in a cascade of reactions that culminates in transcriptional regulation of genes involved in modulating biological activities such as cell proliferation and metastasis.^7–9

An understanding of IHC status in relation to gene expression is an important step toward understanding mechanisms behind how these markers drive breast-tumor biology and toward providing better treatment options for breast-cancer patients. Relatively recent research has uncovered “intrinsic” molecular subtypes of breast tumors, identified via gene-expression profiling.¹⁰ These gene-expression patterns reflect differences in transcriptional activity in the tumors across approximately 50 genes; these subtypes have also been correlated to clinical outcomes.^11,12 The intrinsic subtypes, to a large extent, overlap with the IHC markers (see Table 1). Although the intrinsic subtypes and IHC markers are both useful and provide insights about breast-tumor biology, this review focuses on the IHC-based markers because they are used widely in clinical practice and because data for these markers are widely available in the public domain. However, later in the review, we briefly describe research on relationships between IHC markers and the intrinsic subtypes, thus helping to connect these two informative types of molecular data.

Table 1. Immunohistochemistry classifications and related intrinsic subtypes. Here we provide a mapping between the immunohistochemistry classifications used in breast cancer diagnosis/treatment and commonly associated intrinsic subtypes. While these may routinely overlap, there may be instances in which they do not.

Immunohistochemistry Classifications	Intrinsic Subtype
ER+/PR+, HER2−; ER−/PR+, HER2−; ER+/PR−, HER2−	Luminal A
ER+/PR+, HER2+; ER−/PR+, HER2+; ER+/PR−, HER2+	Luminal B
ER−/PR−, HER2+	HER-2 Overexpression
ER−/PR−, HER2−	Basal Type

Within the overall population of breast-cancer patients, a subgroup of tumors referred to as triple-negative breast cancers (TNBC) have been associated with more aggressive clinical behavior and worse outcomes than other breast tumors.¹³ TNBCs lack ER and PR expression, as well as HER-2 amplification; they constitute 10–20% of diagnosed breast cancers.¹⁴ Some of the characteristic features of TNBCs include earlier age at onset, more advanced stage at diagnosis, as well as aggressive tumor phenotypes.¹⁵ There are currently no targeted drug therapies for managing TNBCs; thus they are currently managed using chemotherapies like anthracycline, taxanes, antimetabolites, platinum agents, and novel microtubule stabilizing agents.¹⁶ However, only 31% of TNBC patients experience pathological complete responses after chemotherapy,¹⁷ which demonstrates the need for targeted therapies. TNBCs also have a higher tendency to metastasize^18,19 and are associated with a higher rate of relapse and poorer survival than non-TNBC^20,21 tumors. When compared with TNBCs, tumors that are ER+/PR+/HER2− are more responsive to hormonal treatments,²² while tumors that are ER−/PR−/HER2+ are better treated with anti-HER2 therapies.²³ The incidence of TNBCs is disproportionately higher in African American (AA) women than European American (EA) women¹⁵ with 20.8% of AA breast cancers being diagnosed as TNBC compared to 10.4% in EA women.²⁴ Also, the mortality rate of AA women diagnosed with TNBC is 42% higher than that of EA women.²⁵ On the other hand, Asian women present with less-advanced stages and have better survival (90% after 5 years) when compared against non-Hispanic whites.^26,27 Such disparities have led researchers to study in more detail what factors may be associated with race in breast cancer. One way they have approached this task is via gene-expression analysis.^28,29 Evaluating gene-expression differences between individuals of different ancestry may help to elucidate biological mechanisms behind those differences and thus help to explain why clinical disparities occur.

Due to requirements of research journals and funding agencies to make research data public, thousands of datasets are hosted in public databases across the world by various institutions. Though these datasets are publicly available, they are scattered across multiple repositories, have been generated using different technologies, and have inconsistent metadata. Even though such data are in the public domain, and publications that arose from them are mostly available, researchers must undertake time-consuming efforts to find datasets that are relevant to a given topic of interest and summarize the data. This review helps to fill this void by identifying gene-expression datasets related to breast cancer, IHC status, and race. We have systematically identified 20 distinct datasets representing a total of 10,214 patient samples that align with these themes. We have examined patterns that cut across 32 publications associated with these datasets, and we provide detailed summaries about the datasets.

Methods

We searched for gene-expression datasets in the public domain related to breast cancer. More specifically, we focused on datasets that included metadata for at least some patients indicating IHC status (ER, PR, HER-2), race, or both. We focused on data that had been generated using high-throughput methods (e.g., microarrays, RNA-Sequencing) and searched the largest available gene-expression repositories: Gene Expression Omnibus (GEO),³⁰ ArrayExpress,³¹ and Oncomine.³² We identified unique datasets within these repositories that matched specific search criteria. For GEO, our search parameters were ((((“her2”) OR “estrogen receptor”) OR “progesterone receptor”) AND “race”) AND “breast cancer”. In ArrayExpress, our search parameters were “Breast Cancer”, “her2”, “race”, “estrogen receptor”, and “progesterone receptor”. Until recently, ArrayExpress maintained a copy of all GEO datasets, so all the data returned from ArrayExpress overlapped with what we already had identified in GEO. Oncomine is a commercial repository, but researchers can perform dataset searches for free after creating an account. We used this resource as an additional search tool. In Oncomine, under the Primary Filters tab > Cancer Type > Breast Cancer, we selected “Breast Carcinomas.” Next, under Sample Filters > Demographics, we selected “Race/Ethnicity”. Finally, under Dataset Filters > Data Type, we selected “mRNA” as the data type. For each dataset that we found in Oncomine, we searched for a publicly available version of the data either in GEO or ArrayExpress. After aggregating our results across these repositories, we performed additional filtering. We excluded datasets that did not have a corresponding publication as well as datasets that lacked metadata on both race and IHC receptor status. We also excluded one dataset that contained IHC status for cell lines because we wished to focus on primary tumors only. We were only interested in race and IHC receptor status, so we did not consider other metadata as filtering criteria.

Results

Datasets

We collected information about the journal article(s) associated with each dataset; this information includes author name, PubMed ID, data source, number of samples, availability of raw data, gene-expression profiling technology used, and the type(s) of metadata available for each dataset (Table 2). After filtering, we identified a total of 20 datasets comprising 10,214 samples. Fourteen of the datasets had raw data available. Fifteen datasets comprising 2231 samples were from microarrays (10 used Affymetrix technologies; 5 used Agilent technologies), while 5 were based on RNA sequencing (Illumina). Depending on the platform used, the raw data are CEL files (Affymetrix microarrays), XML or .txt files (Agilent microarrays), or FASTQ files with read sequences and quality scores from next-generation sequencing data.

Table 2. Gene-expression data sources. This table summarizes all gene-expression datasets that we found in the public domain related to breast cancer and immunuohistochemistry status and/or race. Most identifiers in the data source column reference data series from Gene Expression Omnibus; exceptions are noted. The raw data column indicates whether raw data files (for example, CEL or FASTQ) were available for each dataset. The race status and IHC Status columns indicate whether race and/or IHC status was available for at least some patients in each dataset. NA = Not Available.

Author Name	Pubmed ID	Number of samples	Data Source	Raw Data	Platform	Populations	IHC Status
Brueffer et al.^Citation33	Not indexed in Pubmed	3678	GSE81540	NA	Illumina HiSeq 2000/Illumina NextSeq 500	White* (Sweden)	ER, PR, HER2
Burstein MD et al.^Citation34,Citation35	25208879, 26921331	198	GSE76275	Yes	Affymetrix Human Genome U133 Plus 2.0 Array	White, Asian, Other (Pacific Islander)	ER, PR, HER2
Chang et al.^Citation36	15718313	24	GSE6434	Yes	Affymetrix Human Genome U95 Version 2 Array	Hispanic, Black, White	ER, PR, HER2
Chin et al.^Citation37	17157792	130	E-TABM-158	Yes	Affymetrix High Throughput Array U133AA	White, Black (African American), Other	ER, PR, HER2
Creighton et al.^Citation38	21750966	103	GSE24185	Yes	Affymetrix Human Genome U133A Array	White, Black, Asian	ER, HER2
Curtis et al.^Citation39	22522925	2506	https://www.cbioportal.org/study/summary?id=brca_metabric	Yes	Illumina Human HT-12 v3 Expression Beadchips	White* (UK, Canada)	ER, PR, HER2
Huang et al.^Citation40	24098497	81	GSE48391	Yes	Affymetrix Human Genome U133 Plus 2.0 Array	Asian	ER, HER2
Julka et al.^Citation41	18382427	46	GSE8465	NA	Agilent Human 1A Oligo V2 microarrays	Asian	ER, PR, HER2
Lin et al.^Citation42	19967557	70	GSE19697	Yes	Affymetrix Human Genome U133 Plus 2.0 Array	White, Black (African American)	ER, PR, HER2
Lindner et al.^Citation43	24260093	90	GSE46581	NA	Illumina HumanRef-8 WG-DASL v3.0	White, Black (African American), Hispanic	ER, PR, HER2
Loi et al.^Citation44–47	17401012, 18498629, 20479250, 19552798	249	GSE6532	Yes	Affymetrix Human Genome U133A/B/U133 Plus 2.0 Array	White* (Uk, Sweden)	ER, PR
Lu et al.^Citation48	18297396	129	GSE5460	Yes	Affymetrix Human Genome U133 Plus 2.0 Array	White* (USA)	ER, HER2
Miyake et al.^Citation49	22320227	123	GSE32646	Yes	Affymetrix Human Genome U133 Plus 2.0 Array	Asian* (Japan)	ER, PR, HER2
Pawitan et al.^Citation50,Citation51	16280042, 16813654	159	GSE1456	Yes	Affymetrix Human Genome U133A/U133B Array	White* (Sweden)	ER, PR
Prat et al.^Citation52	24443618	156	GSE50948	Yes	Affymetrix Human Genome U133 Plus 2.0 Array	White	ER, PR, HER2
Rahman et al.^Citation53,Citation54	26209429, 23000897	1119	GSE62944 (The Cancer Genome Atlas)	NA	Illumina Genome Analyzer	White, Black (African American), Asian, Other (Native American)	ER, PR, HER2
Shi et al.^Citation55,Citation56	20064235, 20676074	230	GSE20194	Yes	Affymetrix Human Genome U133A Array	White, Black (African American), Asian, Hispanic, Mixed	ER, HER2
Tabchy et al.^Citation57	20829329	178	GSE20271	Yes	Affymetrix Human Genome U133A Array	White, Black, Hispanic, Asian	ER, PR, HER2
Ulirsch et al.^Citation58	23239149	83	GSE35629	NA	Agilent-012097 Human 1A Microarray (V2) G4110B	White, Black, Hispanic, Other (Native American)	ER, HER2
Wang et al.^Citation59	23512947	30	GSE41400	NA	Illumina HumanWG-6 v3.0 expression beadchip	White, Black (African American) Other	ER

* Population was inferred from the location where the samples were collected, as described in the respective manuscripts.

The minimum number of samples per dataset was 24 (Chang et al.³⁶), while the maximum number was 3678 (Brueffer et al.³³). The median number of samples per dataset was 156. All 20 datasets provided IHC receptor status for at least some patients (Table 3). In total, there were 6635 ER+ samples, 4192 PR+ samples, and 1309 HER2+ samples. There were 2938 ER- samples, 3631 PR- samples and 6038 HER2- samples. The process of determining IHC status is subjective; sometimes histopathologists are unable to clearly determine IHC status, or they disagree about the status of a sample. This was apparent for some patients in which receptors had been examined but IHC status was labeled as “unknown”.^{33,34,37–39,53} In other cases, IHC status was provided for some markers but not for others. In total, ER status was missing for 591 samples, PR for 1978 samples, and HER-2 for 1720 samples.

Table 3. Datasets with immunohistochemistry information.

Authors	# of Samples	Immunohistochemistry Status
		HER2 +	HER2 -	HER2 Un- known	ER +	ER -	ER Un- known	PR +	PR -	PR Un- known
Brueffer et al.	3678	455	2395	714	2732	1046	243	1626	1321	935
Burstein MD et al.	265	22	226	17	49	216	0	42	223	0
Chang et al.	24	2	22	0	18	6	0	14	10	0
Chin et al.	130	11	78	41	84	46	0	74	54	2
Creighton et al.	103	12	54	37	58	42	3	48	50	5
Curtis et al.	2506	247	1733	529	1825	644	40	1040	940	529
Huang et al.	81	34	47	0	53	28	0	NA	NA	NA
Julka et al.	46	12	27	6	24	21	0	22	23	0
Lin et al.	70	34	36	0	34	36	0	25	45	0
Lindner et al.	90	0	90	0	0	90	0	0	90	0
Loi et al.	414	0	0	0	200	45	82	110	23	281
Lu et al.	129	31	98	0	76	53	0	NA	NA	NA
Miyake et al.	115	34	81	0	71	44	0	45	70	0
Pawitan et al.	159	NA	NA	NA	130	NA	NA	114	NA	NA
Prat et al.	156	114	42	0	52	104	0	35	121	0
Rahman et al.	1119	162	546	226	785	230	102	680	332	103
Shi et al.	278	59	219	0	164	114	0	121	157	0
Tabchy et al.	178	26	152	0	98	80	0	83	95	0
Ulirsch et al.	83	25	50	8	48	35	0	0	0	0
Wang et al.	30	5	21	4	15	15	0	14	16	0
	10214	1309	6038	1720	6635	2938	591	4192	3631	1978

Fifteen datasets had race information for at least some samples. These datasets represented a total of 3007 samples (Table 4), although race status was unavailable for some samples in these datasets. We classified the datasets based on high-level, racial/ethnic categories: Asian, Black, Hispanic, Mixed, or White (not Hispanic). Patients with a race/ethnicity classification that didn’t fit into any of these categories were placed in an “Other” group, while samples with no information about race/ethnicity are listed as “NA”. In total, data were available for 225 samples from Asians, 383 from Blacks, 145 from Hispanics, 2032 from Whites, and 33 from Other. For 186 samples in these datasets, no race/ethnicity information was available. Race information was used for research purposes differently across the studies. For example, sometimes gene-expression data was used to compare races,⁴³while in most studies, race information was collected as one of multiple patient characteristics.

Table 4. Distribution of samples within datasets with race information.

Authors	# of Samples	Race
		Asian	Black	Hispanic	Mixed	White	Other	NA
Bittner	354	3	20	4	0	316	9	2
Burstein MD	265	3	0	0	0	254	4	4
Chang et al.	24	0	13	6	0	5	0	0
Chin et al.	130	0	21	0	0	94	11	4
Creighton et al.	103	10	16	0	0	77	0	0
Huang et al.	81	81	0	0	0	0	0	0
Julka et al.	46	46	0	0	0	0	0	0
Lin et al.	70	0	20	0	0	48	2	0
Lindner et al.	90	0	43	8	0	25	0	14
Prat et al.	156	1	1	0	0	154	0	0
Rahman et al.	1119	61	176	0	0	730	1	151
Shi et al.	278	18	29	42	3	176	0	10
Tabchy et al.	178	1	13	83	0	0	0	0
Ulirsch et al.	83	1	24	2	0	52	3	1
Wang et al.	30	0	7	0	0	20	3	0
	3007	225	383	145	3	2032	33	186

Thirteen datasets provided both IHC receptor status and race information (Table 2). By examining datasets that have both types of metadata, researchers may be able to shed light on associations among gene-expression levels, IHC receptor status, and race, including relationships between TNBC and race. Hippen 113

Each of the datasets we identified has been analyzed in at least one primary research article. We examined these articles to identify research questions that have been addressed using these datasets and summarized the articles based on common methodologies used to analyze the data (Table 5). In some cases, we also cite articles not associated with these datasets to provide additional context. Our findings are described in the following sections.

Table 5. Themes across journal articles.

Author	Study Title	Theme
Brueffer et al.	Clinical Value of RNA Sequencing–Based Classifiers for Prediction of the Five Conventional Breast Cancer Biomarkers: A Report From the Population-Based Multicenter Sweden Cancerome Analysis Network – Breast Initiative	Predictive biomarkers (breast cancer biomarkers ER, PR, HER2, Ki67, NHG)
Burstein et al.	Comprehensive Genomic Analysis Identifies Novel Subtypes and Targets of Triple-Negative Breast Cancer	Breast Cancer Subtypes (genomic and transcriptomic), TNBC
Hollander et al.	Phosphatase PTP4A3 Promotes Triple-Negative Breast Cancer Growth and Predicts Poor Patient Survival (Data from Burstein)	Differential Gene Expression Analysis (Comparing TNBC and ER + BC)
Chang et al.	Patterns of resistance and incomplete response to docetaxel by gene-expression profiling in breast cancer patients	Differential Gene Expression Analysis (Comparing pre and post docetaxel administration)
Chin et al.	Genomic and transcriptional aberrations linked to breast cancer pathophysiologies	Predictive biomarkers (Combining DNA data and RNA data to predict survival outcome.
Creighton et al.	A Gene Transcription Signature of Obesity in Breast Cancer	Differential Gene Expression Analysis (Comparing obese and non-obese patients)
Curtis et al.	The genomic and transcriptomic architecture of 2,000 breast tumors reveals novel subgroups	Breast Cancer Subtypes (genomic and transcriptomic)
Huang et al.	Concurrent Gene Signatures for Han Chinese Breast Cancers	Predictive biomarkers (Combining DNA data and RNA data to predict breast cancer survival risk.
Julka et al.	A phase II study of sequential neoadjuvant gemcitabine plus doxorubicin followed by gemcitabine plus cisplatin in patients with operable breast cancer: prediction of response using molecular profiling	Predictive biomarkers (drug response for neoadjuvant therapy)
Lin et al.	A gene-expression signature that predicts the therapeutic response of the basal-like breast cancer to neoadjuvant chemotherapy	Predictive biomarkers (drug response for neoadjuvant therapy)
Lindner et al.	Molecular phenotypes in triple negative breast cancer from African American patients suggest targets for therapy	Differential Gene Expression Analysis (Comparing TNBC between EA and AA populations)
Loi et al.	Definition of clinically distinct molecular subtypes in estrogen receptor-positive breast carcinomas through genomic grade	Breast Cancer Subtypes (Transcriptomic)
Loi et al.	Gene-expression profiling identifies activated growth factor signaling in poor prognosis (Luminal-B) estrogen receptor positive breast cancer	Differential Gene Expression Analysis (Different subtypes of ER+ breast cancer)
Loi et al.	Predicting prognosis using molecular profiling in estrogen receptor-positive breast cancer treated with tamoxifen	Predictive biomarkers (prognosis)
Loi et al.	PIK3CA mutations associated with gene signature of low mTORC1 signaling and better outcomes in estrogen receptor–positive breast cancer	Differential Gene Expression Analysis (Identify genes that are differentially expressed between tumors with PIK3CA mutations and tumors without)
Lu et al.	Predicting features of breast cancer with gene-expression patterns	Predictive biomarkers (breast cancer features such as IHC status – ER & HER2, histological features – grade & lymphatic vascular invasion, stage parameters – tumor size & lymph node metastasis)
Miyake et al.	GSTP1 Expression Predicts Poor Pathological Complete Response to Neoadjuvant Chemotherapy in ER-Negative Breast Cancer.	Predictive biomarkers (drug response for neoadjuvant therapy) Single gene analysis using whole genome data
Koboldt et al.	Comprehensive molecular portraits of human breast tumors	Breast Cancer Subtypes (multi omic data); Integrative Analysis (Combining DNA data – copy number arrays, DNA Methylation, exome sequencing, RNA data – mRNA arrays, microRNA sequencing & Protein data – reverse phase protein arrays
Rahman et al.	Alternative preprocessing of RNA-Sequencing data in The Cancer Genome Atlas leads to improved analysis results (Data from Koboldt)	Bioinformatics (Refining data processing methods)
Pawitan et al.	Gene-expression profiling spares early breast cancer patients from adjuvant therapy: derived and validated in two population-based cohorts	Differential Gene Expression Analysis (Identify differentially expressed genes in treated vs untreated breast cancer patients)
Hall et al.	Hormone-Replacement Therapy Influences Gene-expression Profiles and Is Associated with Breast-Cancer Prognosis: A Cohort Study	Differential Gene Expression Analysis (Identify differentially expressed genes in breast cancer patients that have had HRT vs those that have not); Predictive biomarkers (Prognosis)
Prat et al.	Research-Based PAM50 Subtype Predictor Identifies Higher Responses and Improved Survival Outcomes in HER2-Positive Breast Cancer in the NOAH Study	Predictive biomarkers (Survival outcomes)
Shi et al.	The MicroArray Quality Control (MAQC)-II Study of Common Practices for the Development and Validation of Microarray-Based Predictive Models	Predictive biomarkers (ER status and chemotherapy response)
Popovici et al.	Effect of Training-Sample Size and Classification Difficulty on the Accuracy of Genomic Predictors (Data from Shi)	Predictive biomarkers (ER status and chemotherapy response)
Tabchy et al.	Evaluation of a 30-Gene Paclitaxel, Fluorouracil, Doxorubicin, and Cyclophosphamide Chemotherapy Response Predictor in a Multicenter Randomized Trial in Breast Cancer	Predictive biomarkers (Chemotherapy response)
Ulirsch et al.	Vimentin DNA methylation predicts survival in breast cancer	Predictive biomarkers (Survival outcomes)
Wang et al.	Lipid metabolism genes in contralateral unaffected breast and estrogen receptor status of breast cancer	Differential Gene Expression Analysis (Compare ER+ vs ER- breast cancer)

IHC receptor status and breast-cancer patient outcomes

It is widely accepted that the IHC status of breast tumors has prognostic significance;⁶⁰ these markers are commonly used in determining the choice of therapy, including endocrine therapies, chemotherapies, or combinations of drugs.^61–63 Generally, breast cancer patients with ER+/PR+ tumors have a lower risk of mortality when compared to women diagnosed with ER+/PR-, ER-/PR+ or ER-/PR- tumors (p < .05, hazard ratio 95% CI [2.2–2.4]).⁶⁴ Women with ER+ and PR+ tumors also have a greater chance of survival when treated with chemotherapy regimens such as letrozole or tamoxifen.⁶⁵ More specifically, Prat et al.⁶⁶ showed that HER2+ tumors benefit from being treated with trastuzumab-based chemotherapy, with 44% showing a complete pathologic response (pCR), which is defined as the complete and total absence of residual invasive tumor in the breast and axillary lymph nodes.⁶⁶ Altogether, these findings illustrate the importance of IHC status in the treatment, management, and outcomes of breast cancer.

Race and breast-cancer diagnoses and outcomes

The influence of racial differences on the incidence and mortality rates of breast cancer are not well understood. Some studies^67,68 have highlighted potential multifactorial causes of these disparities, such as access to healthcare, socioeconomic status, treatment delays and tobacco use; but adjusting for these factors does not fully explain these disparities.⁶⁹ Even in high-income countries, such as the US and the UK, disparities exist. In the US, AA women have a 41% higher breast cancer-related death rate than do EA women; the 5-year relative survival is 74% in AA women but 86% in EA women.⁷⁰ In the UK, 5-year, breast cancer relapse-free survival is 62.8% for black women, compared to 77% for white women, despite equal access to healthcare.^71,72 The Surveillance, Epidemiology and End Results Program Cancer Statistics Review (1975–2014) reports that up to the age of 45 years, population-based incidence rates of breast cancer are higher for AA compared to EA women; furthermore, the median age at breast cancer diagnosis is 63 years for EA patients compared to 59 for AA patients.⁷³ The median age of breast cancer patients in Africa tends to be younger (60 years) than those in Europe and North America (63 years), although this younger age distribution may reflect the shorter life expectancy for populations in low and middle income countries.⁷⁴ Smith-Bindman et al.⁷⁵ argued that a possible explanation for these differences is that AA women are less likely to receive routine mammography screening and therefore tend to have larger tumors at a more advanced stage, potentially allowing the accumulation of deleterious mutations in the tumors. Among Asian populations, patients from Japan present with less advanced stages of breast cancer and have better survival rates than non-Hispanic whites.⁷⁶ Hispanic women also have a lower incidence of breast cancer than non-Hispanic white women.⁷¹ Finally, populations of African descent are diagnosed disproportionately with TNBCs when compared to populations of European descent,⁷⁷ whose breast tumors are more frequently ER+ and/or PR+.

With these statistics in mind, it is imperative that scientists and researchers develop new strategies and approaches to improve outcomes in these populations that are more adversely affected. Survival rates may be increased and mortality may be reduced through epidemiological efforts to encourage habits that will help in breast cancer prevention, such as early detection. In addition, a deeper understanding of the molecular mechanisms that drive these differences can advance efforts to aid these populations.

IHC receptor status as a variable in gene-expression studies of breast cancer

Medical oncologists often determine treatment and management strategies for breast-cancer patients based on the expression of IHC markers. However, the methodologies used to identify these markers often vary between laboratories, are subjective, and are not always reproducible.^78,79 Among the studies we examined, four evaluated the use of gene-expression data to guide IHC status assessment,^33,48,55,56 Brueffer et al.³³ tried to predict IHC status using machine-learning classifiers and achieved an accuracy of 95.3% for ER, 90.4% for PR and 88.5% for HER2. Lu et al.⁴⁸ were able to predict ER+ and HER2+ tumors with 98% and 93% accuracy using the recursive support vector machine algorithm. Popovici et al.⁵⁵ and Shi et al.⁵⁶ also developed multi-gene predictors for ER+ status, and their accuracy ranged from 87% – 93%. In most other studies, IHC status was collected in addition to other clinical factors but was not directly correlated to gene-expression levels. Some studies used IHC status to assess similarity in the relative risk of breast cancer mortality among different patient subsets. For example, Dunnwald et al.⁶⁴ found that women who had ER+/PR+ tumors had lower risk of mortality compared to women who had ER+/PR-, ER-/PR+, or ER-/PR- tumors; this risk was independent of demographic characteristics. On the other hand, Creighton et al.³⁸ used reported IHC status as a way to estimate the most appropriate therapy to administer, with patients receiving endocrine therapy if they were ER+/PR+. The use of more objective and consistent methods for determining IHC status, such as via gene-expression profiling, could increase confidence in the determination of treatment and management strategies.

Race as a variable in gene-expression studies of breast cancer

Although many studies provided race information alongside gene-expression data, race was used as the primary variable only in the study by Linder et al.⁴³ Focusing on gene-expression data from TNBC patients, their unsupervised analysis revealed a basal-like gene signature that was differentially expressed between AA and EA patients. AA tumors also had a higher genomic grade when compared to EA tumors.

There are conflicting reports on gene-expression differences between AA and EA women. Some studies have used gene-expression data (although not publicly available) to compare different racial populations. For example, Sturtz et al.⁸⁰ confirmed that the frequency of TNBC was significantly higher (P < .001) in African Americans (28%) when compared to Caucasians (12%), although principal component analysis (PCA) did not detect significant gene-expression differences between these populations.

There have also been documented differences in mortality rates between populations. For instance, Huo et al.⁸¹ analyzed data from 154 AA patients and 776 EA patients in The Cancer Genome Atlas (TCGA). In addition to identifying 142 genes that were differentially expressed between these populations, they observed different survival outcomes between breast-cancer patients of European ancestry and African ancestry, confirming prior evidence that patients of African Ancestry have worse overall outcomes. They also observed other molecular differences between these populations, with AA patients having more TP53 mutations and fewer PIK3CA mutations than EA patients.

Chavez-MacGregor et al, evaluated gene-expression data for 376 samples, but reported no significant differences in RNA expression among tumors from black, white and Hispanic individuals.⁸² They reported that unsupervised clustering of significant probe sets showed differences according to races; yet, their validation set did not confirm this pattern. A possible explanation for these conflicting reports is the small number of samples that were used in their analysis, thus emphasizing the need for large-scale, gene-expression analysis studies that compare different ancestral groups.

Analysis methodologies

Differential gene-expression analysis

Several studies have attempted to identify genes that are expressed differently between breast-cancer subgroups. Chang et al. compared expression profiles of patients who had experienced resistance or an incomplete response to docetaxel against patients who responded well to this drug.³⁶ Their data suggest that the expression of genes in the mTOR pathway may be responsible for a lack of response to docetaxel. Creighton et al.³⁸ examined the relationship between obesity and breast-cancer outcomes. In comparing obese and non-obese patients, they discovered gene-expression patterns associated with obesity that correlated with increased insulin-like growth factor (IGF) signaling. This was associated with worse outcomes in the obese patients.

Lindner et al.⁴³ sought to shed light on gene-expression differences between EA populations and populations of African descent in TNBC patients. They performed an unsupervised analysis and identified a basal-like gene-expression signature that was differentially expressed between these populations. This basal-like signature was strongly associated with increased expression of insulin-like growth receptor 1 (IGF1). Thus, therapeutic regimes that target the IGF1 pathway may be beneficial to AA patients.

Overexpression of HER2+ in breast tumors may lead to tamoxifen resistance,⁸³ though the mechanisms of this resistance are poorly understood. Loi et al.⁴⁴ performed a differential gene-expression analysis and concluded that independently of HER2 overexpression, activation of growth factor (GF) signaling pathways in patients with ER+/HER2+ tumors contributes to poor prognosis for patients with this molecular subtype.

When identifying differentially expressed genes, sample sizes should be considered. Small sample sizes can lead to false negatives and inconsistent findings between studies. Ching et al. used a simulation analysis to estimate sample sizes that are needed for one-factor, two-sample designs such as those described above.⁸⁴ They showed that statistical power is conditional upon the dataset and experimental conditions. For example, when replicates in a single group were homogeneous and differences between the groups were large, a power value of 0.8 was attainable for sample sizes as small as 5 or 10. However, for more complex phenotypes, such as differences between human populations, sample sizes of at least 25 were necessary. With the exception of Chang et al., which used 24 samples, differential-expression analyses exceeded these thresholds. When statistical power is lacking, researchers may be able to combine datasets in gene-expression meta-analyses.⁸⁵

Predictive biomarkers

Multiple studies have attempted to use gene-expression profiles to predict patient characteristics and outcomes. Brueffer et al.³³ predicted IHC receptor status for breast tumors using single- and multi-gene classifiers. They attempted to discriminate between tumors that exhibited either positive or negative expression of these markers. The accuracies of the models were as follows: ER, 95.3% ± 2.4%; PR, 90.4% ± 2.9%; and HER2, 88.5% ± 3.8%. In addition, they made predictions for Ki67 (84.9% ± 3.4%); and NHG, (73.8% ± 3.9%),^86,87 two other proteins that are sometimes used to make clinical decisions for breast-cancer patients. Being able to predict IHC status from gene-expression patterns could help in cases where pathologists are unable to determine IHC status unequivocally, thus informing possible treatment courses. Predicting response to therapeutic regimens can also help inform physicians about which regimens a patient will respond to or be resistant to. Chang et al.⁸⁸ predicted responses to docetaxel using a linear classifier and achieved 88% accuracy (95% C.I. = 68–97%). Julka et al.⁴¹ predicted responses to neoadjuvant drug combinations: gemcitabine + doxorubicin and gemcitabine + cisplatin. Using a k-nearest neighbors classification approach, they achieved a level of accuracy between 73–78%. Ayers et al.⁸⁹ achieved 78% accuracy (95% C.I. = 52–94%) for predicting responses to paclitaxel followed by a combination of fluorouracil + doxorubicin + cyclophosphamide. Although the reported accuracies for these studies are somewhat comparable, it is difficult to make legitimate comparisons from one study to the next because validation strategies (e.g., cross validation or use of an independent test set) are not standardized across studies. Furthermore, different quantitative methods were used in these prediction attempts. These included diverse machine-learning algorithms, but there was no consensus method as each group selected algorithm(s) based on what they deemed to be most suitable to their needs.

Among the studies we evaluated, two used molecular data integration to predict patient characteristics. Huang et al. and Chin et al. both integrated gene-expression measurements with genome copy-number data. Huang et al.⁴⁰ identified genes with copy number variations that were differentially expressed based on ER and HER2 status. These genes were then used to derive gene-expression signatures associated with disease-free survival. They built a survival prediction model based on 16 genes. Chin et al.³⁷ identified gene markers that predicted reduced survival as well as those that could act as possible therapeutic targets. In addition, their research showed that patterns associated with copy-number abnormalities differed among gene-expression based tumor subtypes. However, no gene markers overlapped between these two studies. Future studies that continue the trend of integrating data from multiple assay types may lead to a better understanding of how different types of molecular aberrations contribute jointly to the development of breast cancer and may improve the concordance of analysis results.

Bioinformatics (data processing methods)

Publicly available data have been used to evaluate processing methods for RNA-Sequencing data. Rahman et al.⁵³ observed in experimental replicates that a commonly used pipeline for quantifying gene-expression levels resulted in high variability among cell lines in which HER-2 was activated. This pipeline had been used to process data from TCGA, which is used widely for research analyses; these findings raised concerns that inferences made from TCGA data would be inaccurate. In addition, the pipeline produces normalized gene-expression values, whereas the DESeq2⁹⁰ algorithm commonly used for differential-expression analyses requires that expression levels be represented as raw read counts. To address these concerns, they used the Rsubread package⁹¹ to reprocess the data for 9,264 tumor samples across 24 cancer types, including 1119 breast tumors.⁵³ In this way, IHC status was used as an external validation measure, which may be a useful approach in other studies. Finally, the authors curated clinical data from TCGA, including IHC status and race information, as a way to make it easier for the scientific community to make inferences regarding these variables.

Breast cancer subtypes

The IHC classification scheme presently in use, while useful, is still too broad.⁹² For example, individuals of the same IHC subtype may not benefit from the same regimens. Several studies,^54,63 have used gene-expression data in an attempt to derive more specific breast cancer subtypes. The use of gene-expression data to define subtypes may allow individualized treatment by characterization of molecular profiles unique to these subtypes. Thus specialized treatments could be tailored to these profiles in a more specific way than using IHC status alone.

Burstein et al.³⁴ focused their analysis on TNBCs. Using the Differential Expression via Distance Summary algorithm,⁹³ they identified a subset of genes associated with TNBC status and then used non-negative matrix factorization⁹⁴ to aggregate TNBC patients into four clusters. These clusters formed the basis for discrete TNBC subtypes, each with a different prognosis. They labeled these subtypes as luminal androgen receptor, mesenchymal, basal-like immunosuppressed (BLIS), and basal-like immune-activated (BLIA). BLIS tumors had the worst prognosis, while BLIA tumors had the best prognosis. This classification differs from that of Curtis et al.³⁹ who used joint clustering of copy-number and gene-expression data to define an ER+ subgroup, a luminal A subgroup, a subgroup that includes both ER+ and ER- cases, and several subgroups that consisted of mostly ER+ cancers as well as HER2-enriched (ER-) cases. In contrast, Loi et al.⁴⁵ focused on ER+ breast cancer cases. Using an algorithm called the “gene expression grade index” (GGI), which summarizes the similarity between gene-expression profiles and tumor grade, they attempted to separate ER+ tumors into two distinct and prognostic molecular subtypes. Even though using the GGI method showed some similarity with previous studies,^95,96 they suggested that it provides specific benefits; first, it is simply derived by averaging the expression levels of 97 genes; secondly, the biological function of the selected genes are well understood due to their correlation with histologic grade.

Although these differences in classification may be due to the methodologies as well as the selected datasets used, there is some overlap between the different classification schemes. Identifying biomarkers in these ways may provide a basis for directed clinical research and make it easier to develop targets for more effective treatment.

Relationship between IHC status and intrinsic subtypes

This review has addressed what can be learned about breast cancer based on IHC subtypes. We also mentioned the “intrinsic” subtypes, which are based on gene-expression levels and have shown promise for use in clinical settings.^95,97 Unfortunately, these two systems for subtyping breast tumors are sometimes inconsistent with each other, reaching agreement levels of 75% or lower for some subtypes.^98,99 By integrating IHC status with gene-expression data, it may be possible to derive aggregate subtypes that combine evidence across these two systems. Indeed, ER status is already used to guide the gene-centering normalization approach used for assigning tumors to the intrinsic subtypes.^10,100 However, this approach relies on a balance between ER+ and ER- tumors. Raj-Kumar et al.¹⁰¹ applied principal component analysis to gene-expression data from Curtis, et al.³⁹ and Koboldt et al.^53,54 to evaluate overlap among the subtypes assigned by these two systems and to derive integrated subtypes that ensure an ER balance. Their approach shows promise as a way to reconcile these methods and derive more clinically relevant subtypes.

Discussion

We have summarized 32 studies that reported findings from 20 distinct datasets representing a total of 10,214 patient samples. We have illuminated diverse types of research that have been done with these data and emphasized findings about relationships among IHC receptor status, race, and gene-expression levels in breast cancer. To the best of our knowledge, no prior review has emphasized this combination of topics. Some research articles have reported findings from multiple datasets related to breast cancer and IHC receptor status. However, these have been restricted to relatively few datasets or have had a relatively narrow focus. For example, Roelands et al.¹⁰² curated a compendium of 13 publicly available breast cancer datasets covering 2,142 cases, but they did not make inferences regarding race or IHC status. Karn et al.¹⁰³ extracted data from 15 publicly available datasets for 394 TNBC samples and an independent validation set of 261 samples; using this data they derived prognostic gene signatures. But they did not take race into account when developing these signatures. Gong et al. attempted to identify breast cancer targets using 5 datasets¹⁰⁴ covering 257 breast cancer patients and 98 normal controls by identifying differentially expressed genes between these groups and then mapping them to genes in the Thomson Reuters Integrity Database (a database that hosts drugs, their gene targets, and diseases associated with them); again, that study made no reference to race or IHC status.

The studies associated with datasets in this review differed widely in the research questions that they addressed. For the few studies that addressed the same research questions, we observed a mix of concordance and discordance. Brueffer et al.³³ and Lu et al.⁴⁸ achieved similar levels of accuracy when predicting IHC status using machine-learning classifiers, even though these samples were processed on different expression-profiling platforms. Huang et al.⁴⁰ and Chin et al.³⁷ combined gene-expression measurements with copy-number variations. Both identified instances in which copy number was correlated with gene-expression levels, and both studies associated these observations with patient survival; however, the genes that they identified were inconsistent with each other, perhaps in part because their analysis methodologies differed considerably.

Different studies use different gene-expression profiling platforms (e.g., microarray and RNA-Sequencing), data-processing methodologies, prediction algorithms, etc. These differences can present challenges, especially when researchers wish to combine datasets. Perhaps most notably, batch effects and other technical artifacts introduce noise.¹⁰⁵ Software tools are available for modeling and removing batch effects and other unwanted variation;¹⁰⁶ however, adjusting for platform differences remains a challenge. For differential gene-expression analyses, tools exist for combining datasets from multiple platforms; these include rank-based methods and p-value aggregation techniques.^85,107 However, for other types of analyses, such as multi-study machine learning,¹⁰⁸ methods are less well developed.

Publicly available datasets differ widely in the clinical metadata associated with them.¹⁰⁹ Curation efforts such as the MetaGxBreast¹¹⁰ package can help to alleviate this problem, providing harmonized datasets that can be analyzed more readily. However, this resource is limited to a relatively small subset of available breast-cancer datasets and lacks metadata about race. Expansive, curated resources that place more emphasis on underrepresented populations could help to provide new insights into the interplay among race and breast-cancer subtypes. Furthermore, data collection and sharing could be improved – and better secondary research could be facilitated – if researchers provided a standard set of clinical variables alongside gene-expression data. For breast cancer, such variables might include IHC status, the patient’s race and ethnicity status, age at diagnosis, tumor grade, and treatment responses. Better still would be to provide data about patients’ environmental contexts, but few studies collect or share environmental variables alongside gene-expression data. Environmental factors can alter gene-expression levels and thus should not be overlooked in disease studies.¹¹¹ Approaches used by agricultural and evolutionary geneticists can help shed light on how an individual’s environment influences gene-expression levels and how these factors interact with a person’s genetic background. For example, researchers might study individuals who are genetically similar but live in diverse geographical regions and have differing lifestyles.¹¹¹

Most studies in this review focused on gene-expression levels in tumor tissue. When evaluating differences in expression between two races, questions may arise as to whether those differences are specific to breast tumors or whether those differences are due to the patients’ genetic background alone. With the exceptions of two datasets, none included gene-expression measurements for normal tissues. Rahman et al. provided a reprocessed version of data from The Cancer Genome Atlas (TCGA). Curtis et al. described data from the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC). In both cases, normal data were provided alongside tumor data, enabling researchers to perform diverse types of research with the data; however, to our knowledge, neither dataset has been used to evaluate gene-expression differences between races in non-tumor conditions. We suggest that where possible, in future studies, authors include non-cancer gene-expression profiles as a way to control for baseline differences between groups.

We live in an era when gene-expression data are widely available in the public domain. Exciting opportunities exist for reusing data to derive new insights. For example, by examining patterns that span multiple datasets, researchers can increase sample sizes and ask new questions that would otherwise be infeasible to ask. However, even the process of finding available datasets can be a challenge.¹¹² By summarizing publicly available gene-expression datasets that have IHC and/or race metadata and providing an overview of findings that have already been made, this review lays the groundwork for future studies that curate, combine, and analyze these datasets. For example, identifying genes that are differentially expressed across many studies between IHC-based subgroups or different populations could provide more insights about the mechanisms that drive these subtypes and guide efforts to develop targeted treatments. As we have noted, IHC and race/ethnicity status can be ambiguous to determine in some cases. IHC status often cannot be assigned a definitive positive or negative value, and individuals’ ancestries are often mixed or unknown. Using gene-expression data to impute missing values or clarify ambiguous values could make datasets more complete¹¹³ and be useful in clinical settings. In addition, researchers may be able to develop more robust prognostic signatures – via larger sample sizes and taking covariates such as IHC status, race, and age into account – and ensure that the models generalize to independent datasets.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.