A knowledge-based framework for the discovery of cancer-predisposing variants using large-scale sequencing breast cancer data

We downloaded from the TCGA the original Binary Alignment/Map (BAM) files of the normal sample for all the 695 women and men of Caucasian origin diagnosed with breast cancer. We then used 673 (7 men and 666 women) of the 695 samples, considering only WES data and blood-derived normal samples. We analyzed the BAM files, following the exact same Genome Analysis Toolkit (GATK) pipeline and the same level of sensitivity used for the control dataset (see Additional file 1: Data preprocess). We retrieved from the TCGA open access database the available clinical information for these patients, including age and sex, estrogen receptor (ER) status and molecular subtypes.

We used the aggregated results from the ExAC database as control population. This resource aggregates more than 60,000 samples with germline genotype data, of which 33,370 are classified as of European origin. The original data source is both from population studies (1000 Genome Project, HapMap, Exome Sequencing Consortium) and from disease-related studies (including part of the TCGA). To avoid any overlap with tumor samples, we filtered our data against cancer samples, remaining with a total of 53,105 samples, of which 27,173 were of Caucasian origin.

To perform our analysis we took advantage of several resources and custom-made datasets to annotate and filter variants. In particular we used:

Further details can be found in Additional file 1: Annotation data.

We first asked whether known breast cancer-predisposing variants were present in our dataset. We collected a list of 15 known breast cancer susceptibility genes from the literature (Table 1) [5, 21, 22] and checked for variants in the TCGA dataset, considering both known pathogenic and truncating variants (Table 1). We decided to take into account also rare truncating variants, since they are generally considered de facto pathogenic when the gene exerts its oncogenic function via loss-of-function. This is the case for all the known predisposing genes in breast cancer and, in general, for the large majority of CPGs.

We obtained 16 different mutations that cover 36 of our 673 cases (approximately 5%). The frequency of the identified variants in the breast cancer dataset is compatible with a sample of sporadic cases, especially given the fact that many potential pathogenic variations are still not reported in databases like ClinVar [20]. We found no variation for both PTEN and PRKAR1A (Table 1) but it is rare to find mutations on these genes. The cancer syndromes linked to them (Cowden syndrome and Carney complex) are in fact extremely infrequent in the population: the former has an incidence of one in 200,000 individuals [23], the latter a total prevalence of few hundred reported cases [24].

Many CPGs are associated with complex tumor syndromes or have multiple tumor associations in which more than one tumor type can arise [5]. Known examples are the aforementioned BRCA1 and BRCA2 that are linked to both breast and ovarian cancers [25], or the more recent discovery of PALB2, linked to breast and pancreatic tumors [26, 27]. We therefore looked for any variant connected to additional cancers or cancer syndrome genes and, on 11 genes, we found 11 different rare variants that showed a higher MAF in the cases than in the controls (Table 2). The reference list of cancer-related variants was derived from Humsavar, ClinVar, DoCM and CiViC (see Additional file 1: Annotation data). Some of these variants are extremely rare, found in one patient over 673 and therefore they would fail any statistical test trying to assess their enrichment in cancer. Nevertheless, our hypothesis-driven approach allowed us to identify them as candidates among thousands of rare variations.

Among the genes, we found COL7A1, a collagen gene linked to epidermolysis bullosa, a severe skin syndrome with elevated lifetime risk of melanoma [28]. We also detected a variant on RET, a gene connected to MEN2A syndrome that confers an extremely high penetrant risk of thyroid cancer [29]. To our knowledge, RET has been connected to breast cancer only through deregulation in its expression levels [30]. Evidence of a connection to another thyroid cancer-related syndrome (MEN1) was recently demonstrated in breast cancer [31], but a possible link to MEN2A is novel and, if confirmed, would represent an unusual case of a gain-of-function mutation linked to breast cancer risk. Interestingly, we identified three truncating or frameshift alterations on FANCC, FLCN, and MSH6, three loss-of-function genes associated, respectively, to Fanconi anemia (like PALB2, BRCA1, and RAD51C reported in Table 1) [32], renal cell carcinoma [33], and Lynch syndrome [34], with no previous direct connections to breast cancer. Lastly, we discovered AKT1 E17K, a variant linked to many cancer types, including breast cancer, at the somatic level. It is reported in databases such as ClinVar or OMIM [35] (that are generally focused on hereditary genetic traits) because it is considered a high-frequency somatic driver mutation [36]. This gene was also linked to a minority of Cowden syndrome cases along with PIK3CA since it belongs to the same pathway as PTEN, whose mutations are causative of 85% of the cases [23]. This variant is particularly relevant because it represents both an example of a gain-of-function mutation in a breast cancer frequently somatically mutated oncogene and a risk-associated germline variant in our dataset.

To summarize our findings, we drew a co-occurrence heatmap of all the aforementioned variants in our dataset (Fig. 2). The sum of all the cases with at least one of these mutations is 110 and approximately covered 16% of our dataset and included 11 non-breast-related pathogenic variants, 12 pathogenic breast-related variants, and four truncating variants on breast CPGs. However, co-occurrent mutations were quite rare: only seven of the 110 samples had more than one variant. Furthermore, variant frequency in the dataset was extremely unbalanced: the top eight variants in Fig. 2 account for 13% of patients, while 19 variants cover the remaining 3%.

We identified approximately 70,000 non-synonymous variants; of these we kept the rare variants which showed a higher prevalence in the cases respect to the controls (approximately 50,000). We then computed their deleteriousness score (DS), by adding up the individual scores obtained through the nine different methods evaluating the functional impact of the mutations, and retained the variants with a DS >0.5 (i.e., those defined as damaging in four or more methods, see Additional file 1: Annotation data) [13]. DS filtering yielded approximately 16,000 variants (Fig. 3). Unbiased inference of causality for such rare variants would require exceedingly large datasets. However, the results described above suggest that a simple comparison of variant allele frequencies in the TCGA cases using sufficiently large control datasets allows re-discovering known risk associations. Thus, it should be possible to transfer this approach to exploratory analyses aimed at identifying candidate loci to be validated externally. Accordingly, we further selected our variations, keeping only perfect matches with the somatic mutations in the COSMIC and cBioPortal databases and 440 variants were finally retained; of these, 183 belong to our list of manually curated target genes reported in Additional file 3: Table S6 (see Additional file 1: Annotation data). Among the 183 variants, we found 37 monomorphic alterations in the ExAC database that represented our control. The most relevant result of this analysis branch is probably the already mentioned variant E17K on AKT1 (rs121434592). The AKT1 gene is a known somatic driver kinase and this mutation was found in the cBioPortal database in 46 different samples from many different tumor types, including breast. E17K is also in the CIViC and DoCM database lists of curated somatic driver mutations [36]. Along with ATM R337C (rs138398778), this variant is in the list of cancer hotspots curated by Chang et al. [37], both variants representing a case of known somatic driver mutation that can be considered a cancer-predisposing variant. In addition, we found other germline variants present in more than two samples in COSMIC or cBioPortal on the following genes: HNF1A, FGFR3, and ASXL1. Interestingly, these genes are included in our list of CPGs or somatic driver genes, and none of them has been connected to breast cancer predisposition before.

We also developed a way to annotate if a variant falls close to at least one of 130 manually selected breast cancer-associated SNPs from the NHGRI-GWAS catalog (see Additional file 1: Annotation data), which are reported as GWAS blocks in Additional file 2: Table S3–4 [14]. A variant in proximity of one these SNPs can be considered in linkage disequilibrium with the SNP itself. We identified 436 variants within the GWAS blocks associated with breast cancer (Additional file 3: Table S7). Using STRING [38], we found that the genes containing these variants had more interactions among themselves than what would be expected from a random set of genes of the same size (p value: 0.000279). Interestingly, three genes in our list (ZNF365, SGSM3, and LSP1) were significantly annotated (using Webgestalt [39] and Disgenet [40]) with the “category mammographic density”, that is a strong risk factor for breast cancer [41]. Among the list of SMGVs, only 73 out of 436 fell in such GWAS regions (Fig. 3 and Additional file 3: Table S7). The overlap of these two groups is apparently random as it is not significantly different from a bootstrap of random overlaps (p value of permutation Z test = 0.19). This result highlights two important aspects. First, the lack of enrichment in somatic mutations in GWAS-associated regions confirms the results of Machiela et al. [42], which show no correlation between regions around cancer-associated SNPs and enrichment in somatic mutations. Second, while GWAS are designed to work on common variants, somatic mutations are usually rare. Thus, these two types of analysis represent two different layers of hereditability. After subsetting for our list of target genes, only six variants in four genes ended up being SMGVs and fell in the GWAS regions. These variants form a list of highly valuable candidates (reported at the bottom of Fig. 3), as theorized by one of the COGS flagship papers [7]. In particular, RAD51B is a known breast cancer-associated gene [43]. TET2, another variant discovered in our dataset, is only approximately 80 kb away from the COGS SNP rs9790517. Notably, TET2 has been already associated with breast cancer at the RNA level [44] and it is considered a known somatic driver in leukemia and melanoma [45]. Another COGS variant (rs132390 on EMID1) is in a low recombination region together with the NF2 R335C variation. NF2 has been associated with hereditary neurofibromatosis syndrome 2 and it is mutated at both germinal and somatic levels in breast cancer [46]. The same COGS SNP was found in LD with CHEK2, a known breast cancer-associated gene [47]. Although our HapMap data do not support this linkage disequilibrium, we found a variant on CHEK2 (rs201206424) at approximately the same distance as the NF2 variant described above (approximately 400 kb) [7]. This CHEK2 variant was also found as somatically mutated in breast cancer in our database. Finally, we found two different alterations on EP300 in LD with the COGS SNP rs6001930. EP300 has a well-established role as a tumor suppressor gene but has been poorly investigated as a breast cancer-predisposing gene [48].

The large majority of CPGs exerts their function via recessive loss-of-function variants [5]. We selected from our dataset all the truncating variants occurring below 1% in the control population: in total 2372 different truncating events on 1865 different genes. On this reduced dataset, we looked for any imbalance between control and case frequency in any of the truncating spots with a gene-wise testing procedure (see section Additional file 1: Tumor suppressor-like analysis). After testing and correcting for false discovery rate (FDR) [49], we looked for potential tumor suppressor-like genes in our list of 758 target genes. Only 90 genes had at least one truncating variant with a frequency in the control cohort below 0.01; of these, 38 passed the p value threshold (Additional file 3: Table S8). As a proof of concept, known breast cancer-predisposing genes like BRCA1, BRCA2, and CHEK2 were selected by our procedure. Other known breast cancer-predisposing genes, such as TP53 or PALB2, were instead not found truncated in our dataset because they are too rare for our detection power in a non-familiar selected dataset (Table 1) [50]. Nevertheless, TP53 has one missense variant included in the list of the 183 variants overlapping with somatic mutations, and this particular variant was never reported as pathogenic before (rs138729528). Among the 41 significant LOF candidates, FGFR3, PIK3CB, HNF1A, and KMTC2 were also highlighted as somatically mutated by the previous analysis, but, in this case, we were able to add a possible loss-of-function role. Interestingly, the genetic ablation of the protein encoded by PIK3CB was described to increase ductal branching and tumorigenesis and could lead to mammary gland hyperplasia in transgenic models of breast cancer [51]. In addition, the landscape of the somatic mutations of PIK3CB and FGFR3 could be an indication that their inactivation might promote cancer progression. In fact, PIK3CB and FGFR3 had a higher presence of somatic truncating events (26.1% and 15.7% respectively, as reported by cBioPortal), compared with the number of truncations that are present in a typical gain-of-function oncogene like KRAS or PIK3CA (approximately 1%). The majority of the genes in the tumor suppressor-like list harbors one to two different truncation points. CRIPAK, however, appears to be an exception with 27 different truncations in various points of the gene body. The abundance of frameshifts and nonsense alterations at various points of the protein can be partially explained by the fact that CRIPAK is an intronless gene containing multiple repetitions of a 31 bp sequence and, like other genes with this feature (e.g., CDR1 or AD7C-NTP), tends to accumulate these variations for evolutionary reasons [52]. In fact, a recent publication on the impact of loss-of-function mutations on coding genes scored CRIPAK among the most tolerant genes (0.99 on a scale between 0, low tolerance, and 1, high tolerance) [53]. In addition, due to the repetitive nature of its coding sequence, it is easier to make alignment mistakes [54]. We think it is most likely a false positive result.

In the last part of our work, we moved from a pure case-control study to a more association-like study. Using the algorithm described in the Additional file 1: Age-dependent polygenic model, we determined the main characteristics of pathogenic and non-pathogenic variants, as shown in Fig. 4a, and we used these characteristics to classify the test set of unknown variants (all the remaining non-synonymous variants). The overall model on the training set reported a very low error in the classification process (out-of-bag error equal to 3.5%), with an area under the receiver operating characteristic (ROC) curve of 0.84 (see Fig. 4b). For example, as shown in Fig. 4c, the random forest model has the tendency of assigning high RF scores to the most deleterious variants, as a clear linear trend is visible between the DSs and the RF scores (see in Fig. 4c). The majority of the known pathogenic variants (the red dots) fall into the top two DS categories, compared to the non-pathogenic variations (the blue dots), which appear to fall in every category without a specific pattern. However, the DS is not sufficient to classify the training set properly, and only the integration of the other features results in a very low classification error (Fig. 4a, b).

After the first dimensionality reduction step, we retained 4045 variants entering the second step, and therefore reducing the risk of an inflated dimensionality (Additional file 1: Age-dependent polygenic model). We regressed the age at the initial pathological diagnosis to the genotype of our subjects in order to obtain a list of variants negatively associated with age at onset. The controls in this procedure are therefore not included. The final output in Table 3 shows 19 variants, ordered by the number of times a feature is retained with a negative beta in a model in at least 10% of the 100 models; 15 variants had a negative beta in more than 50% of the 100 models (see Additional file 1: Age-dependent polygenic model).

We noticed several desirable features of the final set of variants. First, without imposing a filter on the control MAF, we selected for rare variants in the population, so that all our 19 variants have, in the control set, a MAF way below the 1% threshold and, in the case set, a MAF for the corresponding variants higher than in the controls. Furthermore, more than a half of these variants were not present in the ExAC dataset. Second, 13 of the 19 variants are classified as truncation events and all the other six missense events have a deleteriousness score higher than 0.8, thus judged as almost certainly damaging. Third, we noticed a double enrichment in variants also found as somatic variants, confirming the importance of evaluating somatic events overlapping with germline mutations.

None of the genes found using this procedure belongs to the list of target genes, none are found within low recombination regions of breast cancer-associated SNPs, and there are very few literature reports of a known involvement in cancer, thus making their selection a completely novel finding (Table 3). Excluding variants on TMCO3, TRIOBP, PYGL, and CST4, all the remaining 15 variants involved one single sample from our dataset; since they are so rare, a simple statistical approach would probably not detect them. As briefly mentioned before, this set of variants and genes are mostly not known to be in cancer, except for PKHD1, a gene involved in polycystic kidney disease and associated with a higher risk of renal cancer [55]. Another known pathogenic variant in PKHD1 has also been mentioned in the first section of the results. Other genes with some evidence of cancer involvement among those reported in Table 3 include STIM2, which was associated to allelic loss in 4p in several tumor types, including breast [56], and FOXP4, an important member of the forkhead box transcription factors, which are involved in tumorigenesis and cell growth [57]. Although not directly implicated in tumorigenesis, other genes that are part of families involved in cancer appear to be promising candidates. These include SERPINF2, a member of the serpin family with a clear role in cancer cell survival [58], PARD6A, a member of the PAR family involved in cell cycle gatekeeping and interacting with major cancer pathways like MAPK and PI3K [59], and HIST2H2AB, part of the cluster 2 of histones whose parallel family in cluster 1 is highly mutated in many cancer types [60, 61].

We aggregated all the identified variants to check for possible interactions between germline and somatic mutations. We collected 27 variants from known CPGs, 183 variants overlapping with somatic mutations, 41 truncating variants from 38 TSG-like genes and 19 variants from our polygenic model for a total of 254 unique variants in 169 genes. We first checked for a sustainability of the signal of our variants by examining whether a germline variant carrier was more or less prone to have a somatic hit on the same gene (Additional file 1 Gene-wise interactions). Three genes demonstrated a certain propensity to have both somatic and germline hits: HNF1B, MSH6, POLR1A (Additional file 3: Table S9). HNF1B has been shown to harbor biallelic inactivation from germline and somatic hits in renal carcinoma [62]. On the 36 MSH6 carriers of at least one of the three variants we identified, two of them were found carrying also a secondary somatic mutation on the same gene, one of them was a stop-gain truncating mutation. A two-hit hypothesis on MSH6 has been proposed in [63, 64]. Furthermore, although not significant because of the small sample size, the somatic mutational burden of the patients carrying a MSH6 variant was slightly higher compared to non-carriers (median for carriers: 33 mutations, median for non-carriers 30 mutations). This is expected from a gene belonging to the mismatch repair pathway. POLR1A, which is a core subunit of RNA polymerase 1, has never been shown to necessitate a biallelic inactivation.

Considering the clinical data of our cohort and extracting the estrogen receptor status (ER), the human epidermal growth factor receptor 2 status (HER2), and the progesterone receptor status (PR) for each patient, we tested the possible association between variant carriers and a particular molecular subtype. As a proof of concept, we initially checked if BRCA1 carriers were associated with ER-negative tumors as previously described [65]. The frequency of ER-negative tumors in non-altered BRCA1 patients was approximately 20%. Among the BRCA1 carriers of a pathogenic, highly deleterious or truncating variant (five samples), we found three ER-negative, one positive and one unknown status (3/4 = 75%). By running Fisher’s exact test between BRCA1 carriers and BRCA1 non-carriers distributions, we found a significant enrichment of ER-negative tumors (p value = 0.025). We expanded this analysis by including the following subtype categories determined by ER, HER2 and PR status: Luminal tumors (ER+/HER2-), HER2+ tumors and triple-negative tumors (ER-, HER2-, PR-). Among the 169 genes taken into consideration in the section above, only two showed a significant imbalance of the distribution of molecular subtypes between carriers and non-carriers (SMOX, a spermine oxidase, q-value = 0.02 and CNOT3, one of the subunits of CCR4-NOT transcription complex, p value = 0.047 - Table S10). For both genes, carriers had more frequently HER2+ tumors compared to non-carriers, (from 2.5% to 25% for SMOX and from 2.4% to 9% for CNOT3). We did not find any particular connection between these two genes and HER2-positive tumors in the literature, thus we think that our result requires further investigation.