Variants of uncertain significance in the era of high-throughput genome sequencing: a lesson from breast and ovary cancers

NGS is a massive parallel sequencing technology that permits simultaneous sequencing of hundreds of DNA/RNA fragments [7]. This technology allows the analysis of single gene- or multi-gene panels enriched for the disease-related genes of interest. Compared to classical standard sequencing methods, NGS simplifies, accelerates, and enlarges the number and variety of sequences analyzed. Indeed, it is possible to sequence different regions (exons and introns), whole exomes (protein-coding regions), or whole genomes, increasing the retrieved information but keeping the costs contained [7]. The choice of the right strategy for genetic testing (e.g., single- or multi-gene panels) and methodology depends on several factors, which include medical purpose (e.g., kind of disorder, etiology, family history), technical considerations and costs [3]. All these issues have been extensively and exhaustively described in many articles that report technical specifications and advances, sequencing purpose, methodological and clinical advantages and relative pitfalls, and we refer to these publications for more detailed information [3, 7‐10].

In oncology, for which research, diagnosis, and therapy demands are more heterogeneous than for monogenic diseases, two additional distinctions are taken into consideration: the genetic testing for germline mutations in cancer predisposition genes, in the diagnosis of hereditary cancer forms, and the genetic testing for somatic mutations in “actionable” genes for therapy evaluation, in sporadic tumors.

In addition to the remarkable benefits brought by the advent of NGS in cancer investigation, this technology has created a paradoxical relative shortage of answers in the face of the massive quantity of information generated by high-throughput technologies [11, 12]. The identification of known somatic or germline mutations accelerates diagnosis, development, and use of drugs that specifically target tumor-driving mutations. Indeed, the use of sequencing data ameliorates patient outcomes by improving the identification of targetable mutations and enlarging the number of patients eligible for specific therapies [11]. However, there are no available drugs for many known mutations or newly discovered variants, yet. While NGS is revealing deep tumor molecular characterization, it may not be sufficient to improve patient outcome or treatment, e.g., in the absence of actionable mutations [12]. Furthermore, NGS has accelerated the identification of many diverse diagnostic sub-groups of patients, but studying populations sufficiently large to be statistically significant is still challenging. This makes rare variants difficult to use for clinical purposes, such as the performance of controlled clinical trials [11]. Another possible challenge in NGS use is the time it may take to receive sequencing results, which might be too long to match clinical needs [11]. Although NGS costs have dramatically decreased in absolute term, this technology is not always financially sustainable in all research and clinical environments [12]. Due to the great number of variants obtained, NGS data are complex and uncertainty may arise in physicians interpreting the results [11, 13]. Moreover, in the last decade, the scientific community has faced the definition of an increasing number of variants that cannot be categorized into strict and clear classifications [6]. The ever-growing accumulation of genetic data generates larger and larger percentages of VUS, and this is especially true of oncological diseases, for which large gene-panel sequencing is often required [6, 14]. Indeed, pathogenic variants often represent only a small percentage (< 20%) of all the variants established in high risk cancer genes, such as BRCA1/2 [15]. A summary of benefits and possible pitfalls related to the use of NGS is reported in Table 1.

Historically, inherited disorders were diagnosed by direct evidence, such as the phenotype, the segregation of the variant with the disease, or personal and family history. With the advent of gene sequencing and the exponential expansion of variants, the requirement for variant classification became essential [16]. Guidelines to define variant pathogenicity in inherited disorders and, later, for somatic and VUS classification were established. Among the first recommendations, there are those by the American College of Medical Genetics (ACMG), that provide five categories of interpretation based on variant reporting and disease association for Mendelian diseases [17‐20]. In addition, they refer to the Association for Molecular Pathology (AMP) workgroup for the interpretation guidelines of somatic variants [16, 21, 22]. Also the International Agency for Research on Cancer (IARC) system subdivides gene variants into 5 classes: classes 1 and 2 include benign and likely benign variants, respectively. Classes 4 and 5 represent pathogenic and likely pathogenic variants, respectively. Class 3 includes VUS (Table 2) [5, 15]. The IARC states that it is essential to discriminate between variants with scarce information (class 3) and variants with strong but not undeniable evidence of disease association (classes 2 and 4) [5, 15]. Class 3 is the most numerous one, comprising about 40% of all variants discovered thus far [23].

Concerning models for VUS classification, great efforts have been made by the IARC Unclassified Genetic Variants Working Group in concert with other groups, such as the Breast cancer Information Core (BIC) for BC predisposition genes [24]. VUS classification is largely independent of the occurrence of the variants in the germline or the soma, and irrespective of the disease being inherited or sporadic. BRCA1 and BRCA2 VUS classifications have been used as models for other kinds of VUS. In 2004, Goldgar and collaborators brought together different sources of evidence, such as frequency in case versus controls, co-occurrence with a known deleterious mutation, co-segregation with the disease in families, occurrence of disease in relatives, and biochemical evidence, such as residue position, conservation and functional assays [24]. The combination of these data can determine the odds of causality by calculating the posterior probability that each variant is pathogenic [24]. In 2008, the Human Mutation journal published a special issue with the title “Assessing mutation pathogenicity in cancer susceptibility genes”. This issue collects many articles curated by the IARC that deeply present and explain all the controversies in and solutions to VUS classification [5, 25‐30]. Briefly, the IARC provides standards for the classification of VUS in high-risk cancer susceptibility genes [5, 25]. This system is based on both direct and indirect evidence, as stated in [24] and, in addition to the elements listed above, takes advantage of the likelihood ratio model calculated for BRCA1/2 in order to use systematically all the available information in a quantitative way [26, 31]. Such information includes tumor pathological characteristics [28], variant functional effects [27], in silico analysis based on sequence-alignment methods, as in missense variant investigation [30].

It is evident that a correct genetic variant classification is essential for managing the genetic information obtained. In some cases, for example, it is mandatory to associate a variant to the correct pathological definition, since it correlates with a specific therapeutic or preventive treatment.

The development of these models and guidelines required the formation of curated databases that integrate as much information as possible. It is in this context that in 2013 the Clinical Genome Resource (ClinGen) project has been launched to create a central resource that defines the clinical validity, the pathogenicity and the clinical usefulness of the genomic information [32]. A clear example of the utility of this resource is represented by the work of Lee and collaborators [33]. Analyzing the HBOC-related genes, the authors defined BRCA1/2 as the only genes with a “definitive” assertion for predisposition to both BC and OC. Instead, ATM, BARD1, CDH1, CHEK2 and PALB2 have a “definitive” association only to BC, while BRIP1, RAD51C and RAD51D only with OC [33]. A key source derived from the ClinGen project is the popular database ClinVar, which archives information on variants with clinical interest [29, 32, 34]. We refer to the next sections for further description.

Two different aspects have to be taken into consideration when dealing with the clinical use of variants and VUS in particular: variant investigation and clinical management.

One of the main databases currently used by the scientific and clinical community worldwide is ClinVar (https://​www.​ncbi.​nlm.​nih.​gov/​clinvar) [34]. It is an international archive of variant-condition interpretations hosted by the NCBI, based on a query-search engine technology and it represents a wide genetic data collection, allowing single variant/gene exploration, interpretation, and sharing [34]. Furthermore, to better understand the impact of genome sequencing on variant definition and achieve a wider perspective, a working group from Utah University developed the ClinVar Miner website (https://​clinvarminer.​genetics.​utah.​edu) [45]. It provides a thorough overview of all genetic variants and it is currently updated to December 2019 with about 1000,000 submissions from ClinVar on more than 670,000 variants. ClinVar Miner recapitulates the great diversity of variants in quantity and quality, for more than 30,000 genes, highlighting three important characteristics: i) the genes with the highest number of submitted variants include the most relevant tumor risk genes, such as BRCA1 and BRCA2, APC, mismatch repair genes, and ATM (Table 5), ii) many submitted variants have conflicting interpretation and/or redundant significance, and iii) a large amount of total variants is represented by VUS (Table 6). In addition, the top 10 genes with the highest number of VUS are enriched in tumor-related genes, whose submitted variants are VUS in up to 40–50% of cases (Table 5) [23, 45]. Concerning HBOC-related genes, up to 20% of BRCA1/2 variants are classified as VUS [46‐48].

It is clear that ascertained tumor-related genes are the object of massive interest and sequencing evaluation, and this is why they accrue a larger number of data and, consequently, of undefined variants [46‐48].

VUS are difficult to classify for three main reasons: i) lack of sufficient population-based statistical evidence, ii) scarcity of functional evidence, and iii) different evaluations by clinicians and researchers. In the first case, VUS might be not so rare, but found in many different pathological conditions and population subgroups, impeding appropriate statistical evaluations and classifications. The second reason is mainly due to the nature of the variant itself: VUS are mainly missense or synonymous substitutions, substitutions of biochemically similar residues, or in-frame insertions/deletions. They may be found in non-coding regions, at less conserved residues, at splicing boundaries, or in less functionally relevant domains compared to true pathological variants. Thus, the impact of such VUS on the proteins and their functions are more difficult to uncover, compared to nonsense mutations. This explains the scarcity and complexity of in vitro assays [15, 40], but strengthens the need for experimental solutions when dealing with VUS. The third reason is due to different approaches taken by scientists and clinicians. For scientists, VUS, non-affecting variants such as polymorphisms, and novel/uncharacterized variants are noteworthy because they can unveil peculiar genetic and protein alterations involved in biochemical processes, although they are not always informative for clinical purposes. On the contrary, medical genetic counselors mostly consider affecting and pathogenic variants with documented involvement in the disease. Furthermore, different laboratories do not necessarily adopt the same standardized reporting format [49]. These divergent approaches create a gap in knowledge and make VUS difficult to use, overlooking potentially disease-relevant information.

In order to improve VUS (re)classification, it has been suggested to introduce genetic testing on family members, while implementing in vitro test development. In this scenario, functional assays, supported by clinical databases and predictive biochemical algorithms, are the most reliable approaches to reach an accurate VUS classification.

Through great efforts, several experimental approaches have been developed in the last few years to determine variant functions in pathological processes. In this context, it is important to choose the appropriate model system to recapitulate the biochemical alterations and their biological consequences [40]. Missense substitutions and VUS are more challenging than truncating mutations since, in the former, the effects on protein structure and function can be less evident. Similarly, experimental assessment of enzymatic catalytic domain activity is easier than the evaluation of, e.g., proteins involved in signaling affected by positive and negative feedbacks [40, 46].

The largest part of functional assays developed thus far concerns the study of BRCA1 and BRCA2 [50, 51]. The importance of these experiments has raised with the increase in the number of VUS and the advent of personalized therapies. The assignment of specific treatments depending on the mutational profile of the tumor has great impact on the management and prognosis of patients. Variant experimental analysis has led to in vitro treatments with specific agents that have been progressively introduced in the clinical practice. An example are the poly (ADP-ribose) polymerase (PARP) inhibitors (PARPi), which were demonstrated to be effective in breast and ovarian cancers bearing mutations not only in BRCA1 and BRCA2, but also in genes involved in the DNA damage homologous recombination pathway, such as PALB2, RAD51, ATM, ATR and CHK2 [52, 53]. The observation that also other “BRCAness” genes were involved in PARPi sensitivity opened the possibility of investigating other kinds of malignancies with dysfunction in the DNA damage response, where these genes are high-to-intermediate risk factors [54].

Each of these in vitro assays is different from the others, depending on the experimental model, the type of mutation analyzed, the protein region in which it occurs and the function it alters. They range from homology-directed DNA repair assays [55, 56], to in vitro transactivation of specific domains and measurements of ubiquitin ligase activity [57]. Studies with conditional knock-out cells allow to evaluate the capacities of HBOC-related gene variants to rescue lethality, DNA repair, and resistance to PARPi [48, 53, 58, 59]. Other assays focus on micronucleus formation and centrosome amplification in mutated cell lines [60, 61], on the restoration of resistance to damaging agents by BRCA2 re-introduction [60], or on increased chromosome breakage after γ-irradiation [62]. Variants in non-coding regions, accounting for about 98% of the genome, are acquiring more importance and are mostly classified as VUS. Many studies measure splicing defects by minigene construction or DNA transcript analysis [63, 64], while others are developing assays to investigate interactions among enhancers, promoters or transcription factors [65].

The advent of new genome editing technologies allowed functional testing with amplified potentiality. The direct manipulation of gene sequences allows the analysis of the biological effects of hundreds of single nucleotide variants (SNVs) in specific protein regions by multiplexed experiments. One example of this is saturation genome editing, which takes advantage of the CRISPR/Cas9 technology [66]. Findlay and collaborators improved this system to build a library of homology-directed repair templates with all the possible SNVs in 13 exons of BRCA1 for simultaneous analysis. They were able to evaluate 3893 SNVs in their native genomic context and assigned a functional score, predictive of pathogenicity. Of all the submitted variants, 72.5% were deemed functional, and 21.1% non-functional. Missense variants were mainly functional (70.6%), but represented the mutational class with the largest percentage of intermediate classification (8.1%), supporting the evidence that missense mutations are the most difficult to define. All these data were confirmed by the high grade of overlapping between the functional score and the ClinVar classification (i.e., 162 out of 169 pathogenic variants scored as non-functional and 20 out of 22 benign variants scored as functional) [67].

These functional assays still share some pitfalls, such as poor feasibility in daily clinical practice. They are strictly dependent on the experimental model, often represented by commercial cell lines that do not recapitulate patient biological processes, and they are time-consuming. Extensive gene editing requires whole research groups and these assays are not applicable to all the different kinds of variant and to every different gene. Nevertheless, they show concrete efficacy in VUS classification and they have directly addressed the need for experimental approaches to the classification of cancer-related gene variants.

One important gene involved in both rare genetic and neoplastic diseases is the ATM gene. The acronym derives from Ataxia-Telangiectasia (A-T), a rare autosomal recessive genetic disease caused by biallelic mutations in the ATM gene that determine an early-onset disorder characterized by cerebellar ataxia, immunodeficiency and predisposition to cancer [68].

The majority of ATM mutations causing A-T are nonsense mutations, leading to truncated proteins and severe-to-complete loss of function. ATM is a very large gene with many variants, about 40% of which (more than 2000) are represented by VUS (Table 7 and Fig. 1) mostly represented by missense, in-frame, or synonymous mutations.

Up to 2% of the whole human population is heterozygous for a pathogenic ATM variant and it has been demonstrated that healthy carriers, such as the parents of A-T patients, are more susceptible to tumor formation, with a 5- to 9-fold increased risk of breast cancer in women [69, 70]. Heterozygotes are also sensitive to damage from radio- and chemo-therapy [69‐73]. Moreover, it has been demonstrated that rare missense variants of moderate-risk genes, such as ATM, confer an increased risk for early-onset breast cancer [74]. Differently from A-T patients, who carry almost exclusively nonsense truncating mutations, cancer patients bear more heterogeneous types of variants. These include out-of-frame mutations with premature stop-codon formation, splicing-affecting variants, synonymous, missense and intronic mutations, with much less evident functional impact, and are enriched in VUS [71, 75]. In the last decade, it has been shown that cell models with ATM mutations that cause low or no protein expression are more sensitive to PARPi, supporting clinical implications similar to those of BRCA1/2 mutant carriers. Such mutations have thus become promising markers for positive PARPi response [76‐79]. As a confirmation, a phase II double-blind trial for the treatment of gastric patients with paclitaxel with or without PARPi olaparib showed that ATM^low patients receiving the combination of the two drugs had an improved overall survival, compared to the whole population and the group receiving paclitaxel only [80]. Given the role of ATM as an intermediate risk factor for breast, hematological, and pancreatic cancers and as marker of PARPi response, the classification of ATM variants is essential for appropriate diagnostic and therapeutic management, and has thus been routinely included in the tumor gene panel for NGS.

To underscore the importance of recognizing clinically-relevant variants, here we highlight how functional analyses may improve ATM VUS classification. Our group recently developed a simple functional test based on the ATM-dependent mitotic centrosomal localization of p53 (p53-MCL). This test is able to detect the presence of pathogenic mutant ATM in peripheral blood mononuclear cells (PBMCs) [81]. The p53-MCL test takes advantage of the activity of ATM on p53 during mitosis, when ATM phosphorylates p53 at Ser15 and allows its translocation to the centrosomes [82]. This mechanism is important for mitotic surveillance [83] and it has been shown that cells mutated in ATM, such as A-T cells, lack this localization [81]. The p53-MCL test is based on an immunofluorescence assay that analyzes the localization of p53 at the centrosomes in PBMCs during mitosis, which is quantitatively impaired in cells carrying pathogenic ATM mutations [81]. This was first demonstrated in a large group of A-T patients together with their parents, as obligate healthy carriers of pathogenic mutations, compared with wild-type healthy donors [81]. A preliminary study performed on BC patients carrying germline ATM missense variants or small deletions in both exonic and intronic regions showed that p53-MCL was impaired, highlighting its potentiality for the detection of ATM pathogenic mutations also in BC [84]. These data suggest that the p53-MCL test could represent a new functional tool to easily help in the classification of ATM VUS in neoplastic diseases.