Breast cancer whole exome-seq datasets

We downloaded the whole exome sequencing datasets for 103 breast cancer samples (54 samples from Mexican patients and 49 samples from Vietnamese patients) from dbGap website http://www.ncbi.nlm.nih.gov/gap (accession number: phs000369.v1.p1) [12]. In this study, we analyzed only 98 samples because 5 Mexican samples have very low sequencing quality. All the 98 breast cancer samples contain tumor/normal pairs. We assume that germline and acquired somatic mutations (till the diagnosis of cancer) could significantly contribute to the differential phenotype of breast cancers [13]; hence, we did not filter out the mutations that are present in the normal sample. Based on the clinical information provided, we sorted the 98 samples (patients) into five comparison groups that include three clinical subtype groups (ER+ vs. ER-, PR+ vs. PR-, HER2+ vs. HER2-), a grade-based (grade II vs. grade III), and a stage-based (stage II vs. stage III) group. A summary of the classification information for the 98 samples is shown in Table 1 (Stage I and IV, grade I were excluded in the comparisons due to lack of sufficient data). The mutation profile for each patient is often sparse. When comparing one class with smaller sample size (<10) against another class with a larger sample size (i.e. >20) in another, one mutation in the former class will be considered to occur at a rate of more than 10%. Therefore, we set a cut off value of 10 to define the descent sample size, in order to minimize the impact of rare mutations in classes with smaller sample size in the statistical tests. In this case, any class that has less than 10 samples will not be compared separately, if applicable (Table 1). We also performed Fisher’s exact test to check the race compositional differences between each comparison group. Notably, the race composition in grade II and grade III is unbalanced (Table 1 and S1 Table); therefore, we only performed the class comparison for Mexican patients, in order to eliminate the effect of any race-specific genetic variations. (A detailed description of clinical information for all samples is shown in S1 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. A summary of the five comparison groups of breast cancers used in this study.

https://doi.org/10.1371/journal.pone.0119383.t001

Sample quality control, alignment, SNV calling and annotation

We used FastQC [14] and FastX toolkit [15] for quality control of the 98 tumor whole exome sequencing datasets. Short reads with low sequencing quality (Phred score < 20) were removed or trimmed, accordingly. Processed reads were then aligned with Borrows-Wheeler Aligner [16] to the human reference genome hg19. We then applied the Genome Analysis Toolkit [17] (GATK) best practices pipeline [18,19] from Broad Institute for SNV (Single Nucleotide Variant) calling from alignment files, and the pipeline includes multiple steps such as Mark Duplicates, Local Realignment, Quality Score Recalibration and variant calling. After 98 SNV profiles were generated, we used ANNOVAR [20] for functional annotation of all the SNVs. The SIFT [21] score reported from ANNOVAR was used to evaluate the degree of deleteriousness of SNVs.

Scoring the deleteriousness of mutated genes

The SIFT score ranges from 0 to 1. An SNV is predicted to be deleterious when its SIFT score is less than or equal to 0.05. Therefore, we filtered out all the SNVs that have the SIFT score more than 0.05. We calculated the deleterious score, D, using the following function, Where

D_ij: the deleterious mutation score for the i^th gene in sample j;

S_ijkx: the SIFT score for the k^th mutation in isoform x of the i^th gene in sample j;

N_k: the number of isoforms that are affected by the mutation k for that specific gene in that sample.

This scoring function combines the SIFT scores for all deleterious mutations in a gene (including the isoforms, if any) and generates a combined deleterious score for each mutated gene. Therefore, by applying this scoring function, we obtained a matrix with 98 columns (98 patients) and about 17000 rows (~17000 RefSeq genes). Each cell represents how deleterious one gene is mutated for the specific patient. Obviously, the higher the score is, the more deleterious way the gene mutations affect the gene function.

Identification of DMGs between breast cancer classes

We identified DMGs between five pairs of breast cancer class comparisons (ER+ vs. ER-, PR+ vs. PR-, HER2+ vs. HER2-, grade II vs. grade III, and stage II vs. stage III) using the univariate t-test at a two-sided significance level of 0.001. Considering the small sample size of grade I and stage I classes, we only performed grade II vs. grade III, and stage II vs. stage III class comparisons (no patients with stage IV breast cancer were present in our list). To adjust for multiple testing, we also reported the false discovery rate (FDR) for each gene identified. The FDR was estimated using the method of Benjamin and Hochberg [22]. This procedure was implemented with the class comparison tool in BRB-ArrayTools [23].

Functional analysis of DMGs and SNVs

We examined the deleterious SNVs present in the DMGs of different breast cancer classes using an odd ratio, which identifies SNVs that are at least 2-fold more frequent in one class over the other between the populations of a two-class comparison. Fisher’s exact test was used to examine the significance of the differences. We then used Pfam (protein family) database [24] and CONDEL [25] to predict the functional impact of those significant SNVs on proteins. Pfam database contains information on evolutionarily conserved functional domains; hence, if an SNV occurs in the domain region, it is more likely to affect the structure and/or function of the protein. CONDEL is a method to assess the outcome of non-synonymous SNVs with the best sensitivity and specificity [25]. It uses the consensus deleterious scores by combining predictions from five different tools that include SIFT[21], PolyPhen2[26], Logre[27], MutationAssessor[27] and MAPP[28].

Protein stability analysis for point mutations

For the DMGs, we analyzed the overall impact of point mutations on protein stability. For feasibility of analysis, we selected a set of 10 relatively rare non-synonymous SNVs that occur either in a functionally annotated (Pfam-A) or evolutionarily conserved (Pfam-B) domain region. We used iprscan version 4.8 [29] for Pfam and PANTHER motif search. We then used two reliable structure prediction tools, RaptorX [30] online webserver and I-TASSER suite [31] standalone version, for protein structure prediction. We ran I-TASSER in parallel mode with the default parameters.

Further, we used three similar and independent tools, I-Mutant-2.0 [32], PopMusic-2.1 [33] and CUPSAT [34] to analyze the overall impact of a point mutation on protein stability. I-Mutant predicts the stability of a point-mutated protein from its primary sequence, while PopMusic 2.1 and CUPSAT predict the same from its 3D structure. We evaluated the overall impact of a point mutation on protein stability based on the consensus results from these three methods; if at least two tools predict the same mutation effect on the protein structure, i.e., destabilizing or stabilizing, then only we accept that result.

Comparison of DMGs in hormone receptor positive vs. negative breast cancer subtypes

Due to many similarities between the ER+/- and PR+/- class comparisons, we are presenting these two classes together in this section. We identified 18 genes with significantly different mutation scores between ER+ and ER- (Fig. 2, Table 2), and 9 genes between PR+ and PR- breast cancer subtypes (Fig. 2, Table 3). Three genes, OR1J2, SKOR1, and DPP3 are commonly identified in both class comparisons. In Tables 2 and 3, genes are listed based on their biological relevance to breast cancer. Of these, CSN3, ERBB2, PPP2R4, CAPZA2, SKOR1, ARID5A, and CPN1 belong to category 1 that contain literature-based relevance to breast cancer. Below, we describe the functional roles of all DMGs under category 1 in each comparison group, while description of all other DMGs can be found in S1 File.

CSN3 (kappa-casein) is involved in myeloid leukemia factor 1-mediated growth arrest and CSN3 deficiency impairs p53 activation, facilitates cell proliferation and affects COP1-mediated p53 degradation [36]. It indicates that mutationally impaired CSN3 could promote cancer growth and progression by dysregulation of the tumor suppressor gene p53. This is consistent with our results that CSN3 has a higher deleterious mutation score in the more aggressive ER- breast cancers compared to that in less aggressive ER+ breast cancers (p-value = 7.06×10^-5).

On the other hand, ERBB2 (also known as HER2/neu), is a well-characterized oncogene that is responsible for development and progression of certain aggressive types of breast cancer. ERBB2 has been shown to be associated with poor prognosis of breast cancers [37]. Overexpression of this gene has been shown to be very crucial in the development and progression of certain aggressive types of breast cancer [38]. Our results corroborate that ERBB2 shows higher mutational load (p-value = 1.57×10^-4) in ER+ breast cancers compared to ER- breast cancers and consequently dysregulated to negate cancer growth and progression in the former subtype.

PPP2R4, also known as protein phosphatase 2A (PP2A), regulates estrogen receptor alpha (ER-α) expression through modulation of ER mRNA stability; hence, it has been considered as a potential therapeutic target for breast cancer [39]. It has been shown that PPP2R4 is involved in PI3K/Akt signaling pathway, a pathway that modulates the interaction between BRCA1 and ER-α [40]. Mutations of PPP2R4 have been shown to contribute to many cancer types including breast cancers [41], and it has been suggested that PPP2R4 might be a tumor suppressor gene [42]. Our results show that PPP2R4 has more deleterious mutations in ER- breast cancers than in ER+ breast cancers (p-value = 2.09×10^-4), suggesting that the higher degree of loss of tumor suppression function for PPP2R4 in ER- subtype relative to ER+ contributes to worse prognosis in the former.

CAPZA2, named as F-actin-capping protein subunit alpha-2, is regulated by Erbb2 in mouse model [43]. It may be also involved in human Ras-MAPK/P13K signaling pathways, as it is predicted to interact with a retinoblastoma tumor suppressor (pRB) protein [44]. Consistent with this notion, our results show that this gene has a higher deleterious mutation score in ER- breast cancers than in ER+ breast cancers (p-value = 4.02×10^-4).

SKOR1, also known as Fussel-15, is a SKI family transcriptional co-repressor that is identified as a DMG both in ER+/- and PR+/- comparisons. It is also a potential repressor of the BMP signaling pathway [45]. A previous study shows that repressing BMP signaling pathway can efficiently prevent bone metastasis from breast cancer cells [46]. Our results show that the gene has a higher deleterious mutation score in ER+ breast cancers relative to ER- breast cancers (p-value = 7.56×10^-4). Similarly, this gene has a higher deleterious mutation score in PR+ breast cancers than in PR- breast cancers (p-value = 1.14×10^-4).

ARID5A has been identified as an ER-α interacting co-repressor protein. ARID5A represses transcriptional activity of endogenous ER-α in MCF-7 breast cancer cells [47]. This gene has a higher deleterious mutation score in PR- breast cancers than in PR+ breast cancers (p-value = 1.0×10^-3).

CPN1 gene encodes an enzyme that is responsible for C-terminal cleavage of stromal cell derived factor-1α (SDF-1) [48]. SDF-1 functions as a growth factor for immature B-lymphocytes and controls chemokine expression, thereby regulating the destination of metastasizing breast cancer cells [49]. Besides, studies show that CPN1 is an estrogen target gene in zebrafish model [50]. This gene has a higher deleterious mutation score in PR- breast cancers than in PR+ breast cancers (p-value = 5.47×10^-4).

Comparison of DMGs in HER2+ vs. HER2- breast cancer subtypes

We identified 10 genes that have significantly different deleterious mutation scores between HER2+ and HER2- breast cancer subtypes as listed in Table 4 (Fig. 2). Among them, BCAR1, CENPJ, EPS8, KIAA0922, and SP4 are directly related to breast cancer as described below. Literature information for Category 2–4 genes can be found in the S1 File.

BCAR1 is a breast cancer anti-estrogen resistance kinase. Previous studies showed that BCAR1 is responsible for resistance to the anti-proliferative effects of tamoxifen [51,52] and its expression level often positively correlate with ERBB2 expression [53], thereby leading to aggressive tumor progression. Table 4 shows that more deleterious mutations of BCAR1 were detected in HER2+ than in HER2- breast cancer subtypes (p-value = 1.06×10^-5), suggesting that BCAR1 mutations lead to its hyperactivation that correlates with the overexpression of ERBB2. Interestingly, it has been found that higher BCAR1 levels were significantly associated with ER+/PR+ tumors [54].

CENPJ encodes centromere protein J that is a co-activator for STAT5 signaling pathway [55] and NF-kappa-B-mediated transcription [56]. Nuclear localization of STAT5 marks a good prognosis of ER+/PR+ breast cancers [57] and could be used as an indicator of anti-estrogen therapy [58]. NF-kappa-B pathway may be involved in the gain of resistance to HER2- targeting agents therapy [59]. Our results suggest that mutations in CENPJ could potentially be the driver events as the deleterious mutation score for CENPJ in HER2+ breast cancers is much higher than that in HER2- breast cancers (p-value = 2.59×10^-4).

EPS8, an epidermal growth factor receptor pathway substrate 8, has been identified as a novel candidate oncogene for breast cancer [60]. EPS8 also decreases chemosensitivity and affects survival of cervical cancer patients [61]. It has been found that small interfering RNA of Eps8, could reduce proliferation and tumorigenesis in Eps8-attenuated HeLa and SiHa cells cultured in dishes or inoculated in mice [61]. Table 4 shows that EPS8 has higher deleterious mutation score in HER2+ breast cancers than in HER2- breast cancers (p-value = 4.61×10^-4), suggesting that its mutations might result in poor prognosis of breast cancers.

KIAA0922 is a novel gene detected in Kazusa cDNA sequencing project [62]. Recent studies on KIAA0922 show that it is a transmembrane 131-like (TMEM131L) protein and it functions as a novel regulator of thymocyte proliferation [63]. KIAA0922 also functions as a novel inhibitor of Wnt signaling pathway [63]. Abnormality of Wnt signaling pathway has been associated with breast cancer [64].

Lastly, SP4 is a transcription factor and down-regulation of this gene is associated with inhibited growth of cancer cells in pancreatic [65], colon [66] and breast cancers [67,68].

Comparison of DMGs in Grade II vs. Grade III breast cancer classes

We identified 7 DMGs between Grade II and Grade III breast cancer subtypes as are listed in Table 5 (Fig. 2). SELP is directly associated with breast cancer [69,70,71]. ANO7 belongs to category 2, and ANKRD18B, ANKRD32 and THAP8 belong to category 3. ADD1 is related to hypertension and SNVs in ADD1 is strongly linked with cancer, but there is no literature evidence showing the involvement of tumorigenesis for this gene. Literature information for Category 2–4 genes can be found in S1 File.

SELP has been a part of an invasive ductal carcinoma gene signature [69]. SELP mediates adhesions for various cells including cancer cells in inflammation, thrombosis, cancer growth and metastasis [70]. High expression of SELP correlates with worse prognosis of human cancer by promoting metastasis of the cancer cells [71].

Although there is no direct evidence for the role of ADD1 in breast cancer progression and tumorigenesis, ADD1 has a significantly higher deleterious mutation score in Grade III breast cancers than in Grade II type (p-value = 3.63×10^-4). Among all patients with grade II and grade III breast cancer, 14 patients have deleterious mutations (rs4961 and/or rs4963) in ADD1, 12 of those have both rs4961 and rs4963 (Fig. 3). A previous study has shown that the carriers of rs4961 were at 1.8 times increased risk for hypertension (CI: 1.32–2.43) [72]. Also, it has been confirmed that rs4963 is tightly linked with rs4961, and thus could also be linked to hypertension [73]. Hypertension has been shown to be one of the common comorbidities in breast cancer patients, and be associated with worse prognosis of breast cancers [74]. Our data shows that 76.9% (10/13) of the grade III breast cancer patients have either rs4961 and rs4963, indicating an increased risk of having hypertension, compared to 16% (4/25) of that for grade II breast cancer patients (Odd ratio is 17.5). Thus, the correlation between hypertension and breast cancer is worth investigating.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The distribution of deleterious SNVs across the compared patient samples.

Five charts illustrate the deleterious mutation distribution in different breast cancer class. Red dot indicates the presence of SNV for the corresponding gene in each sample.

https://doi.org/10.1371/journal.pone.0119383.g003

It should also be noted that almost all the DMGs between grade II and grade III classes have higher deleterious mutation scores in grade III except one gene (THAP8). This suggests that deleterious gene mutations evolve with the progression of cancer.

Comparison of DMGs in Stage II vs. Stage III breast cancer classes

We identified 10 DMGs between Stage II and Stage III breast cancer classes as are listed in Table 6. Similar to the grade class, all the genes in Table 6 display higher deleterious mutations in the worse prognosis class (stage III) supporting the general notion that higher mutational load leads to worse prognosis. Half of these genes are directly related to breast cancer (CPZ, LPPR2, PRCP, UNC45A, and PLEKHG6), MMP20 is related to other types of cancer, while CDH26 and GSTO1 belong to category 3. Literature information for category 2–4 genes can be found in S1 File.

CPZ encodes a member of the metallocarboxypeptidase family. This gene is involved in Wnt signaling pathway [75], and therefore potentially plays a role in prognosis of breast cancer. LPPR2 encodes a lipid phosphate phosphatase-related protein that regulates lysophosphatidic acid (LPA) production and signaling [76], and could promote breast cancer initiation, progression and metastasis [77,78,79]. PRCP encodes a protein that acts as a regulator of cell proliferation and autophagy [80], and is also an anti-estrogen resistant protein in ER-positive breast cancer patients [80]. Autophagy functions as a tumor suppressor mechanism, thereby preventing tumor progression [81]. UNC45A encodes a protein that plays a role in cell proliferation and myoblast fusion, and could increase human breast cancer metastasis [82]. Knockdown of UNC45A mRNA slows down human breast carcinoma cell proliferation and invasion [82]. PLEKHG6 regulates the invasion activity of breast cancer cells [83,84].

Based on the five class comparisons we made in this study, it should be noted that there are several potential limitations in this study. First, results from class comparisons with small sample size were more likely to be affected by rare mutations. Secondly, tumor heterogeneity remains a big challenge for SNV analysis, although tumor heterogeneity did not introduce many false positives in this study. Here, we only reported the most likely genotypes using the GATK tool UnifiedGenotyper. Therefore, any reported deleterious mutations should have decent allele frequency in our samples. Heterogeneity of cancer cells would only neutralize the ability to identify those mutations with lower allele frequency. On the other side, a reported deleterious mutation should be either presented in all subclones, or in one or more subclones that are the dominant population in the sample. Thus, our statistical tests only identified the dominant mutations that are more deleteriously mutated in one group compared to another one. To resolve the tumor heterogeneity issue, the single-cell sequencing technology is a good choice.

Functional analysis of deleterious SNVs

We identified 24 deleterious SNVs that have more than 2-fold difference in the odd ratio while also located inside the functional or conserved domain regions of proteins, from the 117 DMG-associated SNVs (S3 Table). These SNVs are presented in Table 7 (rs12421620 from DPP3 is present in both ER and PR class comparisons). Fisher’s exact tests show that all the odd ratio differences are significant (p≤0.05). For each SNV, we also determined a score that suggests the degree of mutation deleteriousness using the CONDEL software (S4 Table). Fig. 3 shows the presence or absence of SNVs in patients from the five comparison groups of breast cancer. For each class comparison, the frequencies of mutations are highly correlated with the prognostic features. Also, except for the Stage class, all the other classes show contrasting patterns of SNVs (between better and worse prognoses), within each class suggesting their enhancing or suppressing roles in cancer progression. Nine SNVs from ER- vs. ER+ class are predominately present (60.0%) in ER- (poorer prognosis class) while a different set of 6 SNVs (40.0%) is identified in ER+ (better prognosis class). In PR- vs. PR+ comparison, 6 SNVs are significantly present (60.0%) in PR- (poorer prognosis class), compared to 4 (40.0%) in PR+ (better prognosis class). Other classes with poorer prognosis, all have higher number of deleterious mutations, with 4 SNVs (66.7%) in HER2+, 7 (87.5%) in Grade III, and 10 (100.0%) in Stage III. Besides, in Grade II vs. Grade III, Fig. 3 also shows an increased risk of having hypertension comorbidity in Grade III patients because of the higher mutation rate for ADD1 gene in this class (16% in Grade II vs. 76.9% in Grade III).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Differentially occurring SNVs with deleterious mutations in domain regions.

https://doi.org/10.1371/journal.pone.0119383.t007

The deleterious mutation shown in Table 7 for ERBB2 is rs1058808. A previous study has shown that rs1058808 may be associated with higher Body Mass Index (BMI) for high risk of endometrial cancer [85]. Although the association between this SNV and the risk of breast cancer is not identified as statistically significant in these studies [86,87], our results show that this mutation is preferably present in the ER+ compared to the ER- subtype (odd ratio 0.23, p = 0.00624). Another mutation, rs2480452 in PPP2R4 is predominantly present in the ER- subtype (40.7% in ER- vs. 5% in ER+ with odd ratio of 13.06, p = 4.29×10^-4). Our protein stability analysis also suggests that this mutation is destabilizing PPP2R4 protein (Table 8). As mutations of PPP2R4 are significant in the pathogenesis of breast cancer [41], especially in different ER status patients, the downstream effect of this SNV on protein stability is further investigated in the next section.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Pfam and Panther motif analysis for breast cancer related mutated genes and overall impact of mutation in protein stability.

https://doi.org/10.1371/journal.pone.0119383.t008

Table 7 shows SNVs that have significantly different occurrence frequency between different breast classes. For example, rs11225089, rs61751507, and rs79167802 occur more frequently in the PR- than PR+ class; rs1139916 occur more frequently in the HER2- than HER2+ class; rs76504036 occur more frequently in the grade III than in grade II class; rs11540666 and rs2229437 occur more frequently in the stage III than in stage II class. These SNVs might be related to tumor evolution and contribute to different prognosis of breast cancer subtypes.

There are some SNVs that are differentially occurring between the comparison groups but not present in the functional domain regions of proteins (S3 Table). However, it is possible that these SNVs are present in the inter-domain or loop regions, but still have an effect on the structure of protein or otherwise affect a protein’s ability to bind and interact with other proteins.

Protein stability analysis

For feasibility, we selected 9 relatively rare occurring SNVs from Table 7 to analyze the consequences of point mutations on protein stability. We carried out hmmpfam/hmmpanther motif search with iprscan, to assess if the SNVs are part of the functional motif or domain region (with high confidence at an E-value < = 10^-4) (Table 8). Then, we compared two protein structure prediction tools, I-TASSER and RaptorX, using the known PDB structure of DPP3 (PDB ID- 3FVY) to determine which method is more reliable and accurate for protein structure prediction. The Root-Mean-Square Deviations (RMSD) of atomic positions between the known DPP3 PDB structure and the I-TASSER or RaptorX predicted models are 0.45 and 7.1, respectively, indicating that I-TASSER is performing far better than RaptorX. We repeated the structure prediction twice for each protein, in order to check if we can get the same structure for each run or not. I-TASSER always gave the same result while RaptorX often gave slightly different results for several proteins. Thus we used I-TASSER for further analysis of all proteins.

Further, we analyzed the impact of point mutations (SNVs) on protein stability by using I-Mutant 2.0, PopMusic2.1 and CUPSAT tools (S5 Table). Our results suggest that 7 out of 9 SNVs tested have destabilizing effect on proteins. In contrast, the other two SNVs (present in GFM2 and ANKRD32 proteins) have a stabilizing effect (means no significant change to structure or function) after mutation (Table 8) (Fig. 4). In mutated DPP3 protein, negatively charged Glutamate residue (E) got replaced with positively charged Lysine (K) at position 690. Structure analysis of DPP3 suggests that mutant protein has almost similar structure to normal protein, except that the C-terminus has its helix structure changed to a loop structure because of the point mutation (Fig. 4). It has been reported that the C-terminal structure of this protein can play a big role in substrate binding in DPP3 [88]. As the mutation occurs close to the substrate binding residues, K666 and R669 [88], we hypothesize that the altered structure at C-terminus affects substrate binding and consequently alters protein function.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Superimposed structures of normal (green) and mutated (yellow) DPP3 protein chains.

Amino acid change at 690th position for DPP3 leads to the structural changes at the C-terminus (in red square) region of the mutant protein. Normal residue (E) at 690th position is shown in blue and the mutated residue (K) is shown in red.

https://doi.org/10.1371/journal.pone.0119383.g004