Critical Analysis of Genome-Wide Association Studies: Triple Negative Breast Cancer Quae Exempli Causa

1. Introduction

One of the most important discoveries of modern genetics following the release of the complete sequence of the human genome and mapping of all genes is the role of genetic variation in disease incidence. To date, there are 88 million genetic variants that have been characterized. A genetic variant is an umbrella term referring to different types of alterations in a sequence within a certain region of the genome. Although the majority of these variants are unlikely to impact human health, there are some that have been associated with disease. These associations have provided valuable information about human diseases. Such genetic variants include indels (“insertions” or “deletions”), copy number variations (CNVs), translocations, transversions, and single nucleotide polymorphisms (SNPs). For the purpose of simplicity and coherence, we will solely focus on SNPs in the present review.

SNPs are nucleotide variations occurring at a specific location within a genome, and with a frequency greater than 1% in a population. However, it should be noted that these variations do not necessarily cause disease. SNPs at low frequencies are usually referred to as mutations and, in combination with other genetic or epigenetic factors, may be responsible for causing disease. Of the aforementioned 88 million variants characterized in the human genome, 84.7 million are SNPs [1]. As SNPs have two alleles, with some exceptions wherein three alleles have been identified, there are typically two base-pair probabilities (or substitutions) for a given allele within a population [2]. SNPs are frequently used as markers for a genomic region for pursuing genetic studies, with the majority of these SNPs having minimal impacts on biological systems. For example, it has been demonstrated that, while an individual SNP or two SNPs do not appear to correlate with specific cancers, a group of 77 SNPs is found to be highly associated with breast cancer development [3]. In addition to genome-wide association studies (GWAS), the use of next-generation sequencing technologies (NGST) has also afforded the use of SNPs for both genetic mapping and evolutionary studies, as they offer unique advantages over other forms of genetic markers. Firstly, when compared with microsatellites and mitochondrial DNAs, SNPs are less affected by homoplasy, as they correspond to single base nucleotide substitutions, and thus their origins can be explained by mutation models. Secondly, in order to determine parent relatedness and population structure, SNPs have been utilized to quantify genetic variations within an individual genome. Thirdly, SNPs are highly desirable as they are present in large numbers. However, it should be noted that, as with all marker systems, there is a certain level of bias within a panel of SNPs under consideration. In order to overcome some disadvantages of the use of SNP markers in genetic studies, the use of randomly distributed SNPs throughout a genome is highly recommended for any population under investigation [4]. At a minimum, haplotype blocks within a genome should be randomly chosen [4]. Moreover, it is important to point out that SNPs and genetic mutations are neither the same, nor are they synonymous [2].

It is known that GWAS assess and analyze differences or variations in DNA sequences across the human genome to detect genetic risk factors of diseases prevalent in a population under investigation. The ultimate goal of GWAS is to predict either disease risk or disease progression by utilizing genetic risk factors to define the biological basis of disease susceptibility. This also enables the development of innovative and preventative therapeutic strategies [2]. There are two fundamental concepts underlying GWAS, including linkage disequilibrium (LD) and a common disease–common variant (CD–CV) hypothesis. These concepts will be discussed in more detail later on in this review. A basic technical workflow for a typical GWAS is presented in Figure 1.

Therefore, GWAS serve as a tool to identify associations between genetic regions and specific traits of interest. A basic GWAS will evaluate the genetic profiles of hundreds of patients of a well-defined phenotype to those of hundreds of control subjects. In this review, we will present and discuss the most significant findings obtained thus far in GWAS for the triple negative breast cancer (TNBC) disease. We have elected to focus on the TNBC subtype mainly owing to its aggressiveness, young age of diagnosis, and limited treatment options.

2. Fundamental Principles of Genome-Wide Association Studies

As mentioned above, there are two fundamental concepts underlying GWAS. These are linkage disequilibrium (LD) and the common disease–common variant (CD–CV) hypothesis.

LD is defined as a non-random association of two alleles at two or more loci [5]. In turn, this provides insights into past genetic conditions and constraints, allowing for the determination of whether selection was either natural; epigenetic [6]; or owing to other mechanisms that do not occur in isolation, such as genetic drift or gene flow. If detected throughout the genome when investigating populations, LD mirrors population history, breeding system, and geographic subdivision patterns. When investigating genomic regions, LD reveals history of natural selection, gene conversion, mutation, and other forces that either contribute to or cause gene-frequency evolution. Therefore, detecting LD does not guarantee either linkage or lack of equilibrium. Ultimately, it is the local recombination rate that determines how the aforementioned factors affect LD within a certain genomic region or between paired loci [5]. As to be expected, this is related to the concept of chromosomal linkage, which infers that two markers found on a chromosome will remain physically linked on that chromosome throughout consecutive familial generations. However, recombination events will separate chromosomal segments within a family from one generation to another, and this effect is continuously amplified through several subsequent generations. Inevitably, recombination events will break apart segments of chromosomes carrying linked alleles until all alleles within a population are in linkage equilibrium. Simply stated, linkage disequilibrium involves coupling markers at the population level [2]. Furthermore, the rate of LD decay is dependent on the following factors: population size, number of founding chromosomes within a population, and number of generations for which the population has existed. Therefore, it comes as no surprise that there are different levels and patterns of LD when comparing different human subpopulations. For example, the most ancestral human population is that of an African-descent population, which, owing to the accumulation of more recombination events, has smaller regions of LD. Meanwhile, on average, European-descent and Asian-descent populations have larger regions of LD than African-descent populations. This is attributed to the fact that European- and Asian-descendant populations have been generated by founding events, whereby they have split from the African population, thus inherently changing the number of founding chromosomes, population size, and generational age of the population [2].

In general, closely linked polymorphic SNPs have strong LD between them. The International HapMap consortium has demonstrated that the human genome contains haplotype blocks, within which either most or all are high LD SNPs. Thus, there is a fine-scale pattern of LD present in human populations. Subsequently, it has been assumed that high LD levels detected among SNPs are for those alleles that exhibit increased risks of complex inherited diseases [5]. Interestingly, this has been in fact observed for those SNPs significantly associated with breast cancer when GWAS have been conducted and large numbers of SNPs have been surveyed [7]. However, it is important to take into consideration that LD in GWAS can be generated by either undetected or unknown population stratification. Moreover, GWAS have been successful at uncovering associated SNPs, despite the low overall breast cancer risk within a population, and even in identifying new causative alleles [5]. In fact, five new variants have been found to be associated with familial breast cancer, but only 3.6% of familial breast cancer can be attributed to these alleles [7]. However, it is important to point out that it is the relatively high frequency of these causative alleles that allows for their detection by GWAS.

The CD–CV hypothesis [8] has been developed based on the following two principles: common diseases differ from rare disorders in terms of their underlying genetic architecture, and the discovery of several susceptibility variants for a common disease is of high minor allele frequency [2]. In other words, this hypothesis proposes that common diseases are influenced by genetic variations common within a population [8]. Firstly, this suggests that there has to be a high correlation between allele frequency and population occurrence. Secondly, if common genetic variants influence disease, then the effect size or penetrance for any one variant must be small relative to those exhibited by rare diseases [2]. This further implies that, if the same SNP causes a small change in gene expression that alters disease risk by a small proportion, this creates a scenario wherein the frequency of disease incidence and the causal allele are only lowly-correlated. Thus, common variants cannot yield high effects. Thirdly, disease susceptibility is influenced by multiple common alleles based on the condition that common alleles have small genetic effects, and that common diseases exhibit heritability. Additionally, if an allele of a single SNP slightly increases disease risk, this implies that such an SNP accounts for a small amount of variance of the total variation caused by genetic factors. Consequently, multiple genetic factors synergistically account for the total genetic risk of a common genetic variation [2]. However, it is important not to jump to the conclusion that the entire genetic component of any disease is attributable only to common alleles.

3. Challenges of Genome-Wide Association Studies

While discussing genetic heterogeneity and the potential role of rare genetic variants in complex human diseases, McClellan and King [9] have pointed out some important and interesting criticisms of GWAS. Despite the fact that some of these criticisms have already been addressed, it is important to go through them to better understand these issues, and to improve the outcomes of GWAS.

One of these noted issues pertains to the fact that some of the genetic variants lack biological functions, and thereby their relative importance is highly diminished. In fact, it has been observed that GWAS are populated by risk variants of no known functions. Thus, the utility and reliability of GWAS have been questioned as most detected SNPs in GWAS are from intergenic regions [9]. Furthermore, GWAS identify approximate locations of loci associated with disease variants rather than attempt to specifically identify functional SNPs. This is attributed to widespread LD between segregating sites within a given human population. In addition, most SNPs in SNP arrays have unknown biological functions, as most SNPs in HapMaps are located in noncoding regions, and SNP arrays usually do not select for SNPs of known functions. Moreover, it is also important to emphasize that GWAS variants are not functional variants that confer risk, also referred to as “risk variants” in the published literature. Thus, 100% of a subpopulation carrying a risk allele does not truly suggest that all subjects of such a population are predisposed to risk. This simply indicates that LD patterns at a target locus are different than those of another subpopulation. Although the majority of thousands of “risk variants” that have been identified from GWAS have no apparent known biological functions, these are explained by using deduction and rationale, as outlined by the CD–CV hypothesis. This suggests that most genotyping platforms select for common variants. Moreover, as evolution has ensured that the most common variants are neutral, it should come as no surprise that most GWAS findings are neutral, originating from factors other than associations with disease risk. On the basis of evolutionary genetics, most alleles are in fact recent, and they are rare [10,11]. It is unclear what is exactly required for a common allele to remain in a population, as mechanisms of evolution can both facilitate and hinder heritability, particularly as they do not occur in isolation. For example, an allele can significantly increase in frequency without any need for selection when either a population bottleneck occurs (genetic drift) or when a subpopulation migrates and integrates with another (gene flow).

In another claim, it has been reported that it is population stratification that results in GWAS hits [9]. Although population stratification or substructure inflates test statistics, this can be readily identified, and adjusted for accordingly. In general, populations differ among each other over many loci and not only for one or two SNPs, which is precisely how whole-genome data are used to identify stratification. This is exemplified by the particularly fine-scale sub-populations in Europe that can be readily separated utilizing whole-genome data. Most importantly, as population stratification is one of the fundamental assumptions taken into consideration by the CD–CV hypothesis, the GWAS community has established methods to deal with population stratification that are fairly effective for common variants. For example, EigenStrat is a multi-dimensional scaling approach for addressing stratification, and is commonly used as a standard practice in the case-control GWAS dataset. Additionally, family-based study designs in GWAS have an advantage in protecting against stratification. Lastly, frequency estimates are dependent on sample size, thus conferring additional variations to such results [10,11].

As with all studies, sample size significantly impacts interpretations of data. Single GWAS analyses are relatively underpowered owing to the fact that they have a limited number of samples, which drastically increases the probability of false-positive findings. Given this, implementing meta-analysis of several GWAS can overcome these small-sample numbers and study-specific limitations, thus providing a more robust statistical analysis and reduced false-positive results. To date, there are many published articles describing the meta-analysis of GWAS [12,13,14]. However, each meta-analysis consists of several stages comprised of analysis set-up, investigating heterogeneity, data storage, and variant selection for any subsequent analysis. There are several parameters and methods employed for meta-analyses, such as p-values, fixed effects, random effects, Bayesian statistics, and multivariate analysis [15]. Using such meta-analysis methods, a new collaboration, iGOGS, has discovered 74 new susceptibly loci for hormone-dependent cancers [16,17,18,19]. However, there are other consortia that have used this method for identifying other SNPs relevant for each type of disease, such as BCAC [20], ISC [21], and MAGIC [22].

The use of GWAS for cancer research studies has encountered several challenges, including the following: sample size; high numbers (430) of significant SNPs for cancer; association of several SNPs with multiple cancer localizations; implications of identified genes in several key signaling pathways involved in cancer; modulation of some pathways by lifestyle and environment; and, lastly, the fact that most studies are conducted using European populations, thus limiting extrapolation of these findings to other populations [23].

4. Genome-Wide Association Studies on Triple Negative Breast Cancer

5. Risk Factor Scores for Prediction of Breast Cancer

The association of disease with particular alleles remains unclear, as it should. With the availability of large numbers of published GWAS identifying disease susceptibility loci, the following studies are specifically related to breast cancer [7,98,99,100,101,102,103,104,105]. Such susceptibility loci infer the relative increased risk of a disease at a given locus, but albeit not necessarily the particular locus involved in the expression of the disease itself. This hypothesis has contributed to an interesting research paradox. If we are to assume that the disease locus is necessary for disease expression, there is little to no evidence to support such a finding. Conversely, if we are to detect an association between a disease and an allele, there is little supportive evidence beyond such an association. Unfortunately, the association itself is not sufficient for any clinical value. For these associations to be useful to a clinician or a patient, they must be correlated to predictive analytical models. One such predictive model is the predictive risk score. In this model, it is proposed that the more factors, such as environment, that are taken into consideration and accounted for by the predictive risk scores, the better the prediction and the higher the clinical value.

The following section will discuss predictive models used in some GWAS related to breast cancer, although not specifically related to TNBC, but still correlated to receptors of specific subtypes.

In the general population, it is only a small proportion of breast cancer patients who actually exhibit rare mutations in certain genes, such as BRCA1 and BRCA2 genes, that confer the highest risks of developing breast cancer. GWAS have discovered multiple common breast cancer susceptibility variants, which individually present small risks, but their combined effects can be substantial. One method to quantify the combined effects of common variants is to use polygenic risk scores, as these genomic profiles permit stratification of women based on their risks of developing breast cancer [106]. Currently, GWAS have identified 170 breast cancer susceptibility loci [107,108]. According to genome-wide heritability estimates, these loci account for only 40% of the heritability exhibited by all common variants on a genome-wide SNP array. This indirectly suggests that discrimination by the predictive risk score (PRS) could be improved by including more variants, and thereby widening the significance threshold. Additionally, these variants must differ based on breast cancer subtypes; for example, estrogen receptor (ER)-positive versus ER-negative. Therefore, these subtype-specific PRSs may allow for improved prediction of the disease, as well as allow for the selection of women for preventative treatments, which would be extremely beneficial for more aggressive breast cancer subtypes, such as TNBC.

Mavaddat et al. [3] developed PRSs based on 77 established breast cancer susceptibility SNPs in a study of 33,673 breast cancer cases combined with 33,381 control women of European origin. They found that women in the highest 1% of the PRS had a threefold increased risk of developing breast cancer compared with women in the middle quintile (OR = 3.36, 95% CI = 2.95–3.83). In addition, they reported that ORs for ER-positive and ER-negative disease were 3.73 (95% CI = 3.24–4.30) and 2.80 (95% CI = 2.26–3.46), respectively. Moreover, they found that lifetime risk of breast cancer for women in the lowest and highest quintiles of the PRS was 5.2% and 16.6%, respectively, for women without family history, and 8.6% and 24.4%, respectively, for women with a first-degree family history of breast cancer.

These findings highlight the potential for combining PRS and other known risk factors for risk stratification. Moreover, the risk strata defined by PRS have allowed for the evaluation of risk reduction strategies. There are various other studies that have used similar approaches by combining PRS with environmental, modifiable, or non-modifiable risk factors. For example, Garcia-Closas et al. [16] have used breast cancer as a representative model to demonstrate that genetic information, when combined with other factors, provides layered levels of risk stratification that could facilitate personal decision-making or population-based prevention programs. The breast cancer model is uniquely advantageous because it already has established modifiable risk factors, such as menopausal hormone therapy, options for chemoprevention (e.g., endocrine therapy prevents estrogen receptor positive breast cancer), and screening strategies for early detection. As mentioned above, methods for detecting low-frequency causative alleles are required. Maas et al. [109] have attempted to address this by evaluating combined risk stratification utility of common low-penetrant (small effect size) SNPs and epidemiological risk factors. A total of 17,171 cases and 19,862 controls sampled from the Breast and Prostate Cancer Cohort Consortium (BPC3), along with 5879 women participating in a 2010 National Health Interview Survey, have been taken into consideration and studied. This model is used to map the distribution of an absolute risk for the population of Caucasian women in the United States after adjusting for competing causes of mortality. The degree of stratification of absolute risk can be attributed to the following: non-modifiable factors, including SNPs, family history, and components of menstrual and reproductive history; along with modifiable factors, including body mass index (BMI), menopausal hormone therapy, alcohol consumption, and smoking. This suggests that this model permits the identification of population subsets with increased risks that would benefit from altering modifiable factors using risk-reduction strategies. One of the interesting discoveries is that women in the highest decile of risk are attributed to non-modifiable factors; that is, those women with low BMI, who do not smoke or drink, and who do not use menopausal hormone therapy (MHT) appear to have risks comparable to those of average women in the general population [109].

Rudolph et al. [107] specifically investigated how combining PRS with environmental risk factors would improve risk prediction; however, integrating PRS into risk prediction models entailed the evaluation of their joint associations with known environmental factors. Specifically, joint associations of those described by Mavaddat et al. [3], consisting of 77 SNP PRS with factors of reproductive history, alcohol consumption, menopausal hormone therapy (MHT), height, and BMI, were evaluated. These analyses were based on datasets from 20 studies consisting of up to 23,104 invasive breast cancer cases along with similar numbers of controls. Both global and tail-based goodness-of-fit tests in logistic regression models were performed with outcomes expressed for overall breast cancer, as well as for ER status. The best non-multiplicative joint associations with 77-SNP PRS were obtained for alcohol consumption (p-interaction = 0.009), adult height (p-interaction = 0.025), and the combined MHT (p-interaction = 0.038) specific to an ER-positive status [107].

More recently, Mavaddat et al. [65] have enhanced their earlier investigations, wherein they have generated PRS from the largest accessible genome-wide association datasets, optimized for predicting the ER-specific disease, to empirically validate PRS in prospective studies. It is reported that PRS is capable of improving stratification, specifically for screening for ER-specific PRS. Therefore, it could inform treating physicians to preventively target the use of endocrine therapies. Clinical translational studies within the framework of present screening protocols are required to assess the risks and benefits of including PRS.

6. Conclusions

Even though GWAS are not disease-specific, they provide important and new knowledge for both research and clinical studies. GWAS catalogues or databases offer valuable SNP libraries that assist in gaining a better understanding of a target human disease in correlation with other diseases, and also facilitate in developing new research strategies that could aid clinicians. Although most GWAS focus on European populations, these methodologies could easily be transferred to studies of other population subtypes.

As for TNBC patients, there are large numbers of identified SNPs that correlate with cancer risk, but very few SNPs that correlate with either survival or prognosis. In addition, developing a better understanding of how these alterations relate to gene and protein expression in TNBC will help in translating these findings in specific tests that could aid clinicians either in screening patients at risk for TNBC or in identifying new drugs that could improve patient outcomes by using personalized medicine.

The use of high throughput technologies would bring significant advantages to GWAS and databases, as these would correlate SNPs with other biological relevant data, and subsequently translate this complex information to patient care. Furthermore, GWAS can be used for TNBC by identifying correlations of different identified SNPs with miRNA sites useful for targeting to determine their influence on either prognosis or progression of TNBC in patients. Additionally, a functional consequence of miRNA-associated polymorphisms is variability in disease susceptibility. Furthermore, PRS can potentially improve stratification for screening, especially when combining family history, other risk factors, and risk prediction models.

In conclusion, interpretations of GWAS findings remain challenging; however, identifying and using SNPs specific for TNBC, we may be able to elucidate and better contextualize these types of studies. However, further studies should be undertaken on various other population subgroups, as well as pursuing analysis of general or healthy populations in order to overcome the limitations of our current knowledge.

Author Contributions

M.-A.J. conceived the content of the review and wrote the article, while L.-A.P. was involved in the focus of this article as well as contributed to section of this review, reviewing and writing the article. M.B. carefully checked the literature review and contributed to writing and formatting of the article, as well as editing, while A.-A.Z. prepared all tables and figures. A.P. and S.S.K. was involved in editing and modifying some of the content. I.B.-N. was responsible for coordination of the topic, analysis of data, and interpretation. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Funding Agency—Ministry of Research and Innovation, through the project PNCDI III 2015–2020 “Increasing the performance of scientific research and technology transfer in translational medicine through the formation of a new generation of young researchers”; ECHITAS, by the Executive Unit for Financing Higher Education, Research, Development, and Innovation (UEFISCDI), Romania, project title: “Addressing the complex exposome profile in hormone-dependent cancers of the breast and prostate and its influence on tumoral genome”; ACHILLE, grant ID: PN-III-P4-ID-PCE-2016-0795; no. 164/2017; POSCCE 709/2010 grant with the title, “Clinical and economical impact of proteome and transcriptome molecular profiling in neoadjuvant therapy of triple negative breast cancer (BREASTIMPACT)”; as well as Maria-Ancuta Jurj won a scholarship from the European Social Found, Human Capital Operational Programme 2014-2020, project no. POCU/380/6/13/125171.

Conflicts of Interest

The authors declare no conflict of interest.