Introduction

Breast cancer (BC) is the most commonly diagnosed cancer among women in the world [1]. Estrogen receptor (ER)-negative BCs are characterized by the lack of estrogen receptor expression. They account for 20–30% of all BC, and are more common in premenopausal women and women of African ancestry [2]. Genetic factors, such as pathogenic variants in BRCA1, have been shown to contribute to ER-negative tumors [3]. Triple negative BCs (TNBC) are a subset of ER-negative tumors and lack expression of ER, the progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2). TNBC is particularly challenging to treat since tumors are not responsive to routine endocrine therapy or HER2-targeted therapies such as Trastuzumab and Lapatinib [4].

Currently, over 700 SNPs (P ≤ 5 × 10−8) defining more than 400 loci associated with risk for BC have been identified through Genome-wide association studies (GWAS) (reviewed in [5]). Over 60 SNPs have been shown to be associated with ER-negative disease [6,7,8,9,10,11,12,13,14]. Although there has been progress in assessing the biological mechanisms at BC risk loci, those conferring risk to ER-negative disease risk have been underexplored [15,16,17].

In 2016, Couch et al. reported four loci associated with susceptibility to ER-negative BC (P < 5 × 0−8) with three SNPs (rs67073037, rs6734079, and rs4577244) at the 2p23.2 locus representing novel associations [18]. Interestingly, no association with ER-positive BC was detected for 2p23.2, suggesting that the association is specific to ER-negative disease [18]. The SNP most significantly associated at the locus was rs67073037 (P = 4.76 × 10−9) [18]. Preliminary analysis, including eQTL, chromatin marks, luciferase-reporter assays, and electrophoretic mobility shift assays suggested WD repeat domain 43 (WDR43) and tRNA methyltransferase 61B (TRMT61B) as possible target genes [18]. In 2020, fine mapping of the region conducted by Fachal et al. [19]. identified 81 SNPs with strong evidence of association with extensive overlap with the previously identified set proposed by Couch et al. [18]. None are located in coding regions, but 14 SNPs overlap with transcriptional-regulatory elements in BC cell lines [18]. Here we assess the functional contributions of the set of 81 risk-associated SNPs to identify SNPs, regulatory elements, and target genes likely to be the underlying risk at the locus.

Materials and methods

Bioinformatics-analysis pipeline and datasets

SNPs significantly associated with ER-negative BC at the locus after fine mapping (conditional P < 10−6) were retrieved from [19]. Then, SNPs were functionally annotated by examination of the overlap with histone-modification data for normal human mammary epithelial cell line (HMEC) obtained by publicly available (ENCODE/Broad Institute) chromatin immunoprecipitation (ChIP)-seq data visualized via the UCSC Genome Browser https://genome.ucsc.edu/. These features are indicative of putative transcriptional regulatory regions [20, 21] and therefore SNPs that are located in these regions were retained for further analysis. SNPs that did not overlap with these features were not analyzed further for transcriptional-regulatory activity. Next, to evaluate the potential of each SNP to contribute to the regulatory activity in the region, we used RegulomeDB ranks retrieved from http://www.regulomedb.org/ [22]. RegulomeDB guides interpretation of regulatory variants integrating high-throughput, experimental data sets from ENCODE and other sources, computational predictions, and manual annotations. Variants with lower ranks (predicted high regulatory activity) were prioritized for analysis. Because genes likely to be regulated by the candidate regions are expected to be contained in the same topologically associating domains (TAD), we used the EpiTAD tool [23] and the Yue Lab 3 C browser (http://promoter.bx.psu.edu/hi-c/view.php) with the Lierberman-raw 5 kb resolution data from HMEC (GRCh37/hg19) to identify putative target genes.

We used sequences retrieved from the human genome browser for the 41 bp surrounding rs4407214 as the input for JASPAR (http://jaspar.genereg.net/) [24] for both alleles to retrieve the predicted human-transcription factors binding to the site with a relative profile-score threshold of 70%. MirSNP (http://cmbi.bjmu.edu.cn/mirsnp) was used to identify whether the SNPs overlapped with miRNA “seed regions.” ER status and mRNA expression data from the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) were analyzed using cBioPortal [25, 26]. Samples were divided into two groups based on mRNA levels: >2 standard deviations (SD) above the mean = ‘high’; <2 SD below the mean = ‘low’. For TNBC and non-TNBC cell lines, we used RNA-seq data from the Cancer Cell Line Encyclopedia (CCLE) [27].

Cell lines

We used immortalized cell lines representative of basal ER-negative breast tissue: normal human mammary epithelial cell (MCF10A) (ATCC; CRL-10317) and the breast carcinoma cell line CAL-51 (Dynamics; CSC-C0382). All experiments were conducted in MCF10A cells and for EMSA and enhancer scanning experiments, CAL-51 was also used. MCF10A cells were grown in DMEM/F12 (Invitrogen) with 5% donor horse serum (Invitrogen), 20 ng/ml EGF (Invitrogen), 10 µg/ml insulin (Sigma), 0.5 µg/ml hydrocortisone, 100 ng/ml cholera toxin (Sigma), and 1X Penicillin/Streptomycin (Invitrogen). CAL-51 cells were grown in DMEM (4.5 g/L glucose) with 10% fetal bovine serum. Cells were grown at 37 °C in a 5% CO2-humidified incubator.

Electrophoretic mobility shift assay (EMSA)

To determine allele-specific transcription-factor binding, log-growing MCF10A or CAL-51 cells were used to prepare nuclear extracts for EMSA as previously described [28]. EMSA probes covered each SNP ± 20 base pairs. Probe sequences are shown in Supplementary Table 1. Experiments were performed in at least two replicates using the same lysate (technical replicates) and two independent experiments (biological replicates).

To determine allele-specific transcription-factor binding to rs4407214, EMSA probes were designed to cover 41 bases around rs4407214 (20 bases on either side covering the region chr2:29,118,239-29,118,279), for both the reference and the alternative alleles. Probe for the candidate transcription factor USF1 was designed as reported previously [29, 30]. The USF probe was synthesized to contain the human Cathepsin D proximal promoter position +124/+104 for EMSA experiments [29] (Supplementary Table 1). The sequence is homologous to the adenovirus major late promoter element, which contains a half-estrogen response element (5′-GTACC-3′) and E box (5′-CACGTG-3′) as USF1-binding site [31] (Supplementary Table 1). Probes were suspended in TE buffer and annealed at a concentration of 10 µM, and heated at 99 °C for 5 min. Probes were allowed to cool until RT and stored in aliquots at −80 °C. DNA probes were labeled with ATP [γ-32P] (Perkin Elmer) using T4 polynucleotide kinase (NEB) and cleaned using the QIAquick Nucleotide Removal Kit (Qiagen). Labeled probes were then incubated with nuclear extracts (10 µg), LightShift Poly (dI-dC) (Thermo), binding buffer (10 mM Tris, 50 mM KCl, and 1 mM DTT, pH 7.4) and unlabeled competing probes. The reactions were separated by electrophoresis on 6% nondenaturing polyacrylamide gel, 84 V overnight. Gels were vacuum-dried at 60 °C for 1 h, and high-performance films were exposed for 4–24 h. EMSA experiments were performed in at least two technical replicates and two independent experiments.

RNA-guided site-specific DNA cleavage (CRISPR)

To cleave the genomic region around the SNP rs4407214 (RR1), five guide sequences were identified (20-nucleotide DNA sequences followed by a 5′-NGG or 5′-NAG PAM) (Supplementary Table 1). The Streptococcus pyogenes Cas9 targets 20 nucleotide DNA sequences followed by a 5′-NGG or 5′-NAG PAM sequence [32]. These sequences facilitate RNA-guided site-specific DNA cleavage by CRISPR (clustered regularly interspaced short palindromic repeats)/Cas system. The five guide sequences were cloned into the gRNA Cloning Vector (Addgene ID # 41824) according to the Prashant Mali protocol option B (http://www.addgene.org/static/cms/files/hCRISPR_gRNA_Synthesis.pdf) and were confirmed by Sanger sequencing using the M13-Rev primer (Supplementary Table 3).

For RNA-guided site-specific DNA cleavage, 70–80% confluent MCF10A cells, which are diploid, were used for transfection. The five gRNA-vectors were cotransfected with hCas9 (Addgene ID# 41815), and pBABE-puro (Addgene ID# 1764) using Lipofectamine 2000 (Invitrogen) at a 3:1 ratio of Lipofectamine 2000 volume (µL) to DNA (µg). After 24 h, cell cultures were trypsinized, plated at different densities (1:1, 1:9, and 1:20), and selected with 1 µg/mL puromycin (Invitrogen). Clones were isolated with plastic cloning rings and expanded. Excision of the rs4407214 region was confirmed by sequencing (Supplementary Table 3). Eight clones were isolated and characterized. Clones CCΔ1, CCΔ2, CCΔ3, and CCΔ4 displayed a deletion at the targeted locus. Clones CC5, CC6, and CC7 had no deletion at the targeted locus detected in the sequenced amplicon. It is possible that they may harbor genomic changes outside of the examined amplicon. Clone CC8 had a deletion on the 5′ region of the amplicon that did not remove the targeted locus. Clones CC5–8 were chosen to serve as negative controls because they represent cells that were transduced, selected, cloned, and expanded (in parallel with the clones with deletion).

Quantitative RT-PCR

For MCF10A CRISPR clones CCΔ1–4 and CC5–8, RNA was extracted using a Qiagen RNeasy Mini-Prep Kit. Reverse transcriptase reaction was performed with 1 µg of total RNA using the Qiagen QuantiTect Reverse Transcription Kit with genomic DNA removal. Quantitative PCR reaction was performed using TaqMan gene expression assays for RBKS (Hs00223231_m1), BRE (Hs01046283_m1), PLB1 (Hs00290809_m1), FOSL2 (Hs01050117_m1), PPP1CB (Hs01027793_m1), SPDYA (Hs00736925_m1), TRMT61B (Hs00372418_m1), WDR43 (Hs01064086_m1), FAM179A (Hs00416668_m1), C2orf71 (Hs01079277_s1), CLIP4 (Hs00372786_m1), and ALK (Hs01058318_m1). Eukaryotic 18 S assay was used as an internal control (ABI catalog # 4319413E), with TaqMan Universal PCR Master Mix (Life Tech), on 7900HT Fast Real-Time PCR System (Applied Biosystems). RT-PCR was performed with three technical replicates and two biological replicates. The results were analyzed by comparing ∆CT values using Student’s t-test. No clone displayed detectable expression of C2orf71.

Luciferase reporter assays (enhancer scanning)

Genomic tiles (~2 Mb) A–E spanning regions RR2–RR5 containing risk-associated SNPs were generated as previously described [33]. Briefly, forward and reverse primers contained attB1 and attB2 sequences, respectively (Supplementary Table 1). Tiles were cloned in forward and reverse orientations upstream of the SV40 promoter in pGL3-Pro-attb vector to test for enhancer regions. Each clone containing a tile was cotransfected in eight replicates using LipoFectamine 2000 (Life Technologies) into MCF10A or CAL51 cells with pRL-CMV (Promega), an internal control expressing Renilla luciferase, per well of 96-well plates. Luciferase was measured 24 h after transfection by Dual Glo Luciferase Assay (Promega) and two independent experiments were conducted.

To generate the luciferase reporter for the WDR43 3′UTR, the region was synthesized with restriction sites for SpeI, and HindIII (NEB) added to the end of the fragment (Supplementary Table 1). The fragments were synthesized with the reference [GG] or alternative/effect [TC] haplotypes for the rs11680458 and rs1131880, respectively. The fragment was cloned in using SpeI and HindIII (NEB) into the pMIR REPORT miRNA expression reporter vector (Ambion) according to the manufacturer’s protocol. Positive clones were selected by restriction digestion using BamHI (NEB) and confirmed by Sanger sequencing. For the luciferase reporter assay, MCF10A was plated at 10,000 cells per well in 96–well plates 24 h prior to transfection. Allele-specific WDR43 3′UTR pMIR REPORT construct and pMIR-REPORT empty vector (EV) were cotransfected with pRL-TK (Renilla luciferase driven by TK promoter) as an internal control, mimic-hsa-miR-141-3p (Sigma), mimic-hsa-miR-376c (Sigma), mimic-hsa-miR-548ar-3p (Sigma), mimic-hsa-miR-200a-3p (Sigma), and mimic-has-miR-155-5p as a negative control that does not target the 3′ UTR WDR43 region. Transfections were performed using lipofectamine 2000 (Invitrogen) at a ratio of 3:1 lipofectamine 2000 volume (µL) to DNA (µg). After 24 h, luciferase activity was measured using the Dual Glo Luciferase Assay Kit from Promega. Luciferase values were normalized to the internal control from Renilla luciferase values and compared with the mimic miR-155-5p-negative nontargeting control. Luciferase experiments were performed with eight technical replicates and two independent experiments. The results were analyzed using Student’s t-test.

Results

Characterizing the chromatin landscape at the 2p23.2 risk locus

We mapped the 81 SNPs that considered the credible set of SNPs (those most likely to contribute to the phenotype, conditional P < 10−6) at the 2p23.2 locus by Fachal et al. [19] to the human genome and determined their overlap with genes obtained from ENCODE RefSeq to assess their functional relevance. None of the 81 risk-associated SNPs (Supplementary Table 2) were located in coding regions, consistent with the hypothesis that they might impact risk through allelic-specific differences in gene regulation rather than altering protein structure or function [34]. These risk-associated SNPs, spanning a region of approximately 72 kb, either overlapped or were proximal to WDR43, a protein-coding gene related to rRNA processing and ribosomal biogenesis (Fig. 1) [18].

Fig. 1: Chromatin landscape at the 2p23.2 locus.
figure 1

Human Genome browser (hg19). SNPs significantly associated with ER-negative breast cancer at the 2p23.2 locus after fine-mapping (conditional P < 10−6) previously reported by Fachal et al. [19]. The 2p23.2 locus region contains risk-associated SNPs spanning the WDR43 gene at the 2p23.2 locus. Several risk-associated SNPs mapped to ENCODE features in human mammary epithelial cell line (HMEC). The index (most significantly associated in the meta-analysis) SNP, rs67073037, is shown in red. Tiles for luciferase-reporter assays to identify enhancers are shown as orange boxes A–E.

Full size image

We identified five noncoding regions, containing 14 risk-associated SNPs, hypothesized to constitute transcriptional regulatory regions (RR) [18] (Fig. 1). Candidate transcriptional regulatory regions were defined based on tissue-specific features and chromatin-state segmentation from ENCODE ChIP-seq data for H3K4m1, H3K4m2, H3K4m3, H3K9ac, H3K27ac, H3K36m3, H3K79m2, H4K20m1, EZH2, and H2AZ in normal human mammary epithelial cell line (HMEC). Previously, using two breast cancer datasets, we only found significant eQTL associations to TRMT61B association [18]. We searched for eQTL and sQTL associations in GTEx (V8) using the breast (mammary tissue, n = 469) and confirmed that only eQTL associations to TRMT61B are found at the locus (data not shown). Risk-associated SNPs in RR1 are located within the first intron of WDR43, and SNPs in RR2 overlap with introns 8–10 of WDR43. SNPs in RR3 map to a miRNA “seed region” at the 3′ untranslated region (UTR) of WDR43 (Fig. 1). SNPs in RR4 and RR5 are in the intergenic region between WDR43 and FAM179A (Fig. 1).

Regulatory Region 1: transcription factors interacting with rs4407214

We previously reported rs4407214 allele-specific binding of nuclear proteins from MCF10A and CAL-51 cells using electrophoretic mobility shift assay (EMSA) [18]. Evidence from luciferase reporter assays, which showed allele-specific activity, and ENCODE features, is consistent with this region harboring a regulatory region that could act as an enhancer element [18]. To identify transcription factors binding to rs4407214 in an allele-specific manner, we combined bioinformatics prediction of transcription-factor binding with ChIP-Seq data available in the ENCODE project obtained from the University of California, Santa Cruz (UCSC) Human Genome browser repository site.

We conducted a transcription-factor search analysis using the 41-base sequence surrounding SNP rs4407214 using the JASPAR web tool [24]. JASPAR identified 1852 predicted binding sites for 266 unique human transcription factors binding to the region containing rs4407214 [G or T alleles] (Fig. 2A). The rs4407214 SNP was part of the binding site for 188 transcription factors (Fig. 2A) (Supplementary Table 3). Next, we used ChIP-Seq data from the ENCODE Factorbook [35] repository to identify 57 unique transcription factors that have been experimentally shown to bind to the region containing rs4407214 (Fig. 2A) (Supplementary Table 4). Of those, 13 had a ration of supporting/total experiments > 0.8 and only five had more than one experiment (Fig. 2A). Using these stringent cutoffs, we identified USF1 as present in both sets (Fig. 2A).

Fig. 2: Allele-specific transcription factor binding.
figure 2

A ‘Subway chart’ showing the in silico analysis pipeline for determining candidate transcription factors binding to rs4407214 using predictions from JASPAR database and ENCODE Factorbook data. B, Electrophoretic mobility shift assays using nuclear extracts from CAL-51 (left panel) and MCF10A cells (right panel), respectively, for competition between unlabeled [T] or [G] alleles with a labeled probe for USF1 and USF2. Red arrows indicate allele specific binding. Control lanes (hot probe only and nuclear extract plus hot probe) are loaded as two technical replicates side-by-side. Cold competitor lanes are loaded as three technical replicates side-by-side.

Full size image

Next, we determined whether USF1 displayed allele-specific binding activity using electrophoretic mobility shift assays (EMSA). Nuclear protein extracts from CAL-51 and MCF10A were incubated with DNA probes containing consensus sequence for USF 1 (Supplementary Table 1). To evaluate the binding-competition capability of the SNP rs4407214 [T] or [G] alleles, probes for each allele were combined with nuclear extracts in the binding reaction. The SNP rs4407214 [T] allele did not compete for the binding with the USF factor (Fig. 2B, lanes 5–7) unlike the [G] allele (Fig. 2B, lanes 8–10) in both MCF10A and CAL-51 cell lines. These results suggest that USF1 modulates the activity of the enhancer/promoter element that includes rs4407214 and support the notion that the SNP rs4407214 contributes mechanistically to ER-negative BC risk.

Deletion of putative regulatory region 1 containing rs4407214 via CRISPR–Cas9

To evaluate the role of RR1 as an enhancer on genes at the 2p23.2 locus, we removed the genomic region around SNP rs4407214 via the CRISPR/Cas-9 system. Our screening revealed four MCF10A clones (CCΔ1–4) in which the region around rs4407214 was deleted, and four clones (CC5–8) that retained the region and served as negative controls (Fig. 3A).

Fig. 3: Deletion of regulatory region 1.
figure 3

A The region around rs4407214 was removed using CRISPR/Cas-9 system with the indicated guide RNAs (gRNAs) in the first intron of the WDR43 gene. Clones were screened using Sanger sequencing and aligned to the region to ensure the removal of rs4407214. B RT-qPCR using TaqMan assays for the expression of 11 genes spanning 2 Mb. Values for CRISPR clones containing the removed region (CC∆ 1-4) were compared with the negative controls, or clones with intact regions (CC 5-8). Significant change (p ≤ 0.05) in gene expression between the CC∆ and CC clones is indicated.

Full size image

Next, we compared expression levels of genes in a 2-Mb region at the locus (Chr2: 28,000,000 to 30,000,000; RBKS, BRE, FOSL2, PLB1, PPP1CB, SPDYA, TRMT61B, WDR43, FAM179A, C2ORF71, CLIP4, and ALK) in clones CCΔ1–4 compared with our CC5–8 with 18 S acting as a control-housekeeping gene. No clone showed detectable expression of C2ORF71. Quantitative TaqMan assays revealed a significant change in expression of phospholipase B1 (PLB1) (P = 6.6 × 10−3), which was consistently downregulated in all CCΔ clones (Fig. 3B). This indicates that RR1 acts as an enhancer to PLB1, which is located 400 kb telomeric to WDR43. We did not detect any significant change in expression of WDR43. We also did not observe any significant changes in TRMT61B, a methyltransferase for N(1)-methyl adenine on tRNAs, expression between CC clones and CCΔ clones as we anticipated based on the previous eQTL analysis [18]. These data indicate that the regulatory region around SNP rs4407214 targets the expression of PLB1 in a mammary gland epithelial cell line.

Contribution of the Regulatory Regions 2–5 to the regulation at the 2p23.2 locus

Next, we examined the functional roles of the SNPs in RR2-RR5. We identified RR2, RR4, and RR5 as regions with evidence of enhancer activity using ChromHMM [36], a multivariate hidden Markov model that models chromatin marks to identify different chromatin states (Fig. 1). We also used RegulomeDB which assigns functions based on a collection of regulatory information on a set of variants, with lower ranks associating with higher regulatory function. Most candidate SNPs in RR2–RR5 had weak evidence for functionality defined by high (≥4) RegulomeDB ranks, except rs11677283 and rs4666144 (Table 1).

Table 1 Candidate functional SNPs in regulatory regions RR2-RR5.
Full size table

Five genomic tiles (Tiles A–E) covering these regions and containing risk-associated SNPS were assessed for enhancer activity in both forward and reverse orientation. Only tile D (RR4, rs4666144) displayed activity in at least one orientation in both cell lines (Fig. 4A) (Table 1). Tile D displayed allele-specific activity depending on the rs4666144 allele [C] or [T] but only in one cell line (Fig. 4B). We then tested whether any of the five risk-associated SNPs in these regions displayed allelic-specific binding of EMSA. SNPs rs11672283 and rs4666144 showed evidence of allele-specific binding in CAL51 nuclear extracts (Supplementary Fig. 1). Overall, evidence for allele-specific enhancer activity was present for rs4666144, albeit weakly. Next, we investigated other potential mechanisms of regulation within the locus.

Fig. 4: Reporter assays.
figure 4

A Enhancer scanning activity detected by luciferase assays using tiles A, B, and C in RR2, tile D in RR4, and tile E in RR5 in MCF10A and CAL-51 cells. Red boxes indicate significant (p < 0.05) difference in luciferase activity in relation to the no-enhancer control pGL3. B Allele-specific enhancer activity detected by luciferase assays using tile D in RR4 in MCF10A and CAL-51 cells. Red boxes indicate significant (p < 0.05) difference in luciferase activity between the two alleles. C Luciferase reporter assay for allele-specific miRNA regulation. Overview of 3′ UTR region of WDR43 construct containing “seed regions” (underlined) for miRNA binding to SNPs rs11680458 and rs1131880. Luciferase assays transfected MCF10A cells with either an empty vector (EV) or construct containing either the [GG] or [TC] alleles for rs11680458 (D) and rs1131880 (E), respectively (WDR43 3′UTR). MCF10A cells were cotransfected with miR-141 and miR-548 against rs11680458, and miR-376 against rs1131880 along with miRNA control (miR-155) and a Renilla sp. luciferase internal control. Significant differences in a two-tailed test (p < 0.05) are shown (ns, nonsignificant).

Full size image

Regulatory region 3: allele-specific miRNA-mediated regulation

SNPs rs11680458, rs1131880, and rs12465505 mapped to the 3′ untranslated region (UTR) of WDR43 (Fig. 1). Thus, we explored it as a possible mechanism of regulation, which we refer to as Regulatory Region 3 (RR3). SNPs rs11680458 and rs1131880 were predicted by MirSNP database [37] to occur within miRNA “seed regions” (Fig. 4C). The rs11680458 reference allele [G] was predicted to be recognized by the miR-141 and 200, while the effect allele [T] was predicted to be recognized by miR-548. The rs1131880 effect allele [C] was predicted to be recognized by the miR-376 and miR-577 (Supplementary Table 5). Our working hypothesis was that one or both SNPs might alter miRNA-binding sites affecting the expression of WDR43.

To assess the effects of rs11680458 and rs113188 alleles, we compared luciferase expression controlled by allele-specific sequences of the WDR43 3′-UTR. We designed the insert with both alleles to match the most frequent haplotype present in CEU populations, retrieved by the LDHap tool [38]. The reference haplotype for both rs11680458 and rs1131880 [GG] was present in the same construct, while the effect haplotype [TC] was present on a different construct (Supplementary Table 1). We cotransfected MCF10A cells with expression vectors for miRNA mimics miR-141, miR548, miR-376, or nontargeting negative control miR-155, a pMIR-REPORT Luciferase vector (Ambion) containing the WDR43 3′UTR region with either haplotype set or an empty vector, and a Renilla sp. luciferase driven by a TK promoter as an internal control (Fig. 4C). The pMIR-REPORT empty vector did not show a statistically significant change in luciferase expression when miR-141, miR-548, or miR-376 were cotransfected in MCF10A cells (Fig. 4D, first three samples; Fig. 4E, first two samples). We detected statistically significant increases in cells transfected with the pMIR-REPORT vector containing the WDR43 3′UTR with the reference haplotype [GG] and the mimic miR-548 or with the risk haplotype [TC] and the mimic miR-141 or miR-376 (Fig. 4D–E). These results suggest that variation in the risk-associated SNPs rs1131880 and rs11680458 can control the expression of WDR43 by modulating transcript stability.

Role of target gene in ER-negative disease

Our analysis shows that PLB1 and WDR43 are regulated by risk-associated SNPs at the locus. BC data from METABRIC [25] show that ER-negative status is more frequent in patients with high WDR43 expression (79.82%) when compared with low WDR43 expression (3.23%). The same is not true for PLB1 (27.9% versus 19.1%) (Supplementary Fig. 2A). Both PLB1 and WDR43 expressions are higher in TNBC cell lines from the Cancer Cell Line Encyclopedia (Supplementary Fig. 2B) [27]. Further studies are warranted to determine how changes in expression of PLB1 and WDR43 contribute to the cancer phenotype.

In summary, we explored five candidate-regulatory regions, RR1–RR5, at the 2p23.2 defined by the presence of SNPs associated with risk to ER-negative BC (P < 5 × 10−8) at the 2p23.2 locus. Two regions, RR1 and RR3, risk-associated SNPs (part of the credible set of 81 SNPs) displayed allele-specific activity and are therefore likely to contribute to risk. Our data suggest that PLB1 and WDR43 are targets for RR1 and RR3, respectively.

Discussion

Although over two hundred loci have been identified through GWAS for BC susceptibility, few have been thoroughly dissected via post-GWAS functional analysis. Among these loci, ~20 have been associated with risk for ER-negative breast tumors [10, 18]. These loci associated with ER-negative BC remain relatively unexplored and little is known about the molecular mechanism driving risk. Here, our objective was to identify likely causal SNPs, those that contribute to increased risk for the ER-negative BC subtype at the locus via allele-specific effects, and illuminate biological mechanisms driving risk at the locus.

Our strategy (Supplementary Fig. 3) started from the assumption that SNPs mechanistically involved in driving risk at the locus would be represented in a set containing 81 credible SNPs from fine mapping analysis of the locus [19]. Integration with publicly available datasets of genomic features (e.g., coding regions, open-chromatin regions, and histone marks of enhancers or promoters) revealed five candidate-functional regions narrowing the set of credible causal SNPs to 12 SNPs located in regions with evidence for promoter/enhancer activity and 2 additional SNPs in a microRNA-binding sequence (Supplementary Fig. 3). We then rigorously evaluated these regions and their risk-associated SNPs in experiments designed to identify allele-specific effects. Experimental analysis identified three transcriptional-regulatory regions, RR1, RR3, and RR4 mediating risk at the locus, although the latter region only had weak evidence of activity.

RR1 contains one risk-associated SNP (rs4407214, Pmeta = 7.63 × 10−09) with allele-specific effect on binding to the upstream stimulator factor (USF) family of proteins. USF1 and USF2 encode 43–44 kDa proteins that function as both homo- and heterodimers to regulate transcription of target genes [31, 39]. Removal of the sequence surrounding rs4407214 led to a downregulation of PLB1, a phospholipase involved in the metabolism of choline. Surprisingly, we did not observe changes in gene expression from the predicted eQTL target TRMT61 [18]. Interestingly, breast carcinomas have increased concentrations of choline containing compounds and choline kinase inhibitors are selectively cytotoxic to tumor cells [40], which identify PLB1 as a promising candidate biomarker for triple-negative BC. Our working hypothesis is that rs4407214 is critical to the activity of an enhancer region and the effect allele [G] increases recruitment of transcription factor USF1 to the enhancer, leading to changes in expression levels of PLB1 with a predicted protective effect on BC risk.

Risk-associated SNPs in RR3 operate through mechanisms different from RR1. In this region we uncovered allele-specific effects of risk-associated haplotypes on the expression of WDR43 by modulating binding of microRNA (miRNA) to the 3′ UTR of the WDR43 transcript. WDR43 is involved in rRNA processing [18]; however, its role in BC is not defined.

Taken together, our data suggest that risk to ER-negative BC associated with the 2p23.2 locus is driven, at least in part, by two common variants, rs4407214 and rs1131880, operating through distinct mechanisms acting on PLB1 and WDR43. While the risk allele in RR3 is expected to lead to degradation of WDR43 transcript, ER-negative tumors from METABRIC are associated with high WDR43 expression, which may reflect a compensatory mechanism. Further studies on the role of these genes in BC are needed to determine their biological relevance.

The limitations of this work include the possibility that regulatory landscapes may differ significantly during development, the incomplete regulatory profiling information from normal human mammary-gland cells, and the possibility of not accounting for unknown very rare alleles or with low quality of imputation that contributes to the phenotype. Despite these limitations our strategy identified two SNPs operating to modulate ER-negative BC risk. Similar examples of risk loci containing multiple SNPs contributing functionally to risk have been uncovered in analysis of loci conferring risk to inflammatory bowel diseases [41]. Future studies will likely focus on dissecting the individual contributions from the two SNPs identified in this study.