A combination of different types of evidence incorporating population data, functional studies, clinical data, and predictive tools is necessary for thorough, thoughtful variant classification.
Variant classification criteria may be optimized in a quantitative, gene-specific manner using validated predictors of pathogenicity for genes or conditions with sufficient information.
Large-scale data (genome sequencing of healthy and affected cohorts, high-throughput functional studies, and in silico metapredictors) increase the robustness of evidence used for variant classification and lend themselves to incorporation in quantitative frameworks.
Collaborative efforts by laboratories and disease-specific expert groups reduce variant classification discrepancies and improve the quality of variant interpretation information available to patients and researchers.
Sequence variants may be detected in an individual’s germline DNA by molecular genetic testing performed in diagnostic, predictive, or preventive clinical settings. Interpretation of variants detected via this testing is complex and requires gathering multiple pieces of information to be analyzed in the context of the clinical presentation and the gene or disorder under consideration. This article reviews the history of germline sequence variant interpretation, highlighting the lines of evidence used in current variant interpretation practices. It also considers how developments in sequencing and other technologies have provided data that have enhanced the ability to interpret genetic variants.
Molecular germline genetic testing has evolved from a niche service consisting of single-gene testing for affected individuals exclusively by genetics providers, to multigene panels or exome testing for either affected or unaffected individuals offered by any clinician, and will one day involve routine genome sequencing of healthy individuals as a part of primary care preventive medical practice. Expanding the scope of genetic testing as well as the population undergoing analysis brings new challenges to analyzing results. Inaccurate variant interpretation can have dire consequences for patients. For genes associated with high cancer risks, detection of a variant deemed pathogenic can lead patients to make surgical decisions that are irreversible should the variant classification subsequently change based on new information. Specific medications or therapies to treat seizure disorders or metabolic conditions may be offered to or contraindicated for patients with pathogenic variants in specific genes. These factors drive the need for a careful, evidence-based approach to variant interpretation for individuals pursuing molecular genetic testing regardless of the clinical indication.
A brief history of sequence variant interpretation
Clinical testing for sequence variants and copy number changes may be performed on an individual’s germline DNA to diagnose a hereditary disorder based on clinical phenotype or family history, or may also be undertaken to screen healthy individuals for disease risk. This testing seeks to identify changes, or variants, in the patient’s DNA that differ from a “healthy” reference sequence. Inevitably, the more genes that are analyzed, the more variants are bound to be detected. The challenge then becomes interpretation of these variants: gathering evidence to determine whether they should be considered pathogenic (disease causing) or benign (not associated with disease risk).
Although it may seem logical to assume that a truncating variant (one that disrupts the protein’s reading frame, leading to its decay or removal of an essential functional domain) in a gene for which loss of function is the mechanism of disease would be pathogenic, this does not always hold true. For example, exons 26 and 32 of TSC2 (GenBank NM000548.3), also described as exons 25 and 31 using alternate exon numbering, are subject to alternative splicing, and truncating or frameshift variants in these regions have been identified in adults lacking features consistent with tuberous sclerosis complex . Likewise, it can be tempting to consider a missense variant as pathogenic if it occurs in a functionally important protein domain, but subsequent functional studies may reveal no impact, or the variant may be present at high population frequency. It is therefore necessary to have gene-specific rules to guide the process of variant classification that take into account multiple lines of evidence and combine them to determine variant pathogenicity.
More than 20 years ago, Cotton and Scriver suggested criteria for determining whether a DNA variant was phenotype-modifying or neutral, specifying type of variant, segregation studies, population frequency, and functional studies, along with the extent of DNA analysis performed, as helpful pieces of information for variant classification. These lines of evidence continued to serve as core criteria for other qualitative and quantitative gene-specific variant classification models . However, one laboratory or expert group might place a higher or lower level of importance on the same piece of evidence, leading to discrepant classifications for the same variant, and to confusion for patients and providers .
Major steps forward in harmonizing the approach to variant classification occurred with publications from the American College of Medical Genetics and Genomics (ACMG) in 2000 and 2008 and coauthored with the Association for Molecular Pathology (AMP) in 2015 . The 2015 ACMG/AMP guidelines outline a qualitative approach to variant classification incorporating data from several lines of evidence: population data, computational and predictive data, functional data, segregation data, de novo data, allelic data, phenotype specificity, co-occurrence data, and expert opinion. These guidelines have also been adopted by the UK-based Association for Clinical Genomic Science. Variant curation expert panels developed under the auspices of ClinGen ( www.clinicalgenome.org ) are using the ACMG/AMP guidelines as a framework, publishing modified versions of the guidelines tailored for gene-specific use. Of note, these guidelines were crafted for use in the classification of mendelian disorders, as opposed to more common and complex polygenic conditions or somatic cancer, and are relevant to sequence variants detected by Sanger, next-generation, or other sequencing methodologies as opposed to copy number variants. In addition, the ACMG/AMP guidelines provide guidance regarding the nomenclature used to define pathogenicity and likelihood for each category ( Fig. 1 ).
This article focuses on each line of evidence specified in the ACMG/AMP guidelines ( Fig. 2 ), providing background regarding the utility of each criterion for sequence variant classification purposes. Readers are referred to the ACMG/AMP guidelines for definitions regarding the proposed evidence codes (eg, PVS1, PM5, PP4) that are referred to in this article.
Lines of evidence used in sequence variant interpretation
Predicted Effect on Protein and Disease Mechanism
The first step in variant interpretation is to understand the nature of the gene-disease mechanism. Does the gene act via a loss-of-function (LoF) mechanism or a gain-of-function/dominant-negative effect? Protein-truncating variants, most of which lead to nonsense-mediated decay and haploinsufficiency, in genes with an LoF mechanism have a high prior probability of being pathogenic. This probability is captured in the ACMG/AMP rule code PVS1, which is considered to be very strong evidence supporting pathogenicity. PVS1 also encompasses other types of alterations that have a high probability of leading to haploinsufficiency, such those that affect the initiation site, canonical plus or minus 1 or 2 splice sites, certain intragenic tandem duplications, and single-exon or multiexon deletions. It may be inappropriate to apply a very strong evidence level toward pathogenicity for LoF variants in the following situations:
The gene’s mechanism of disease is not LoF
The variant causes protein truncation but does not result in NMD
The variant results in an in-frame deletion
The variant occurs in a region that is subject to alternative splicing
However, the ACMG/AMP guidelines do not expand on how these caveats would affect use of the PVS1 code. Abou Tayoun and colleagues provide a decision tree for modifying the strength of PVS1 that takes into account the type of variant, location within the protein, and other evidence.
Other ACMG/AMP criteria based on protein impact and knowledge regarding variant spectrum include:
PP2: used to support a possible role in disease causation for a missense variant in a gene in which (1) missense variants are a common disease mechanism, and (2) benign missense variants are uncommon. BP1: used to support a benign role for a missense variant in a gene in which only truncating variants are known to cause disease.
PM4: moderate pathogenic evidence for variants that change the length of the protein; this may include in-frame deletions or duplications or variants that extend protein length.
BP3: supports a benign role for an in-frame deletion of a repetitive sequence in a region with no known function.
BP7: supports a benign role for a synonymous (silent) variants not proved or predicted to affect splicing.
Historically, a polymorphism was defined as an allele present in at least 1% of the general population . Given their high frequency and the rare nature of most genetic disorders, it follows that variants present at high enough population frequency, to be defined as polymorphisms, would be considered harmless, with several notable exceptions ( Table 1 ).
|Gene and Variant||Disorder||Approximate Population Frequencies|
|CFTR c.1521_1523delCTT (p.Phe508del), commonly referred to as delta F508 (ΔF508) (NM_000492.3)||Cystic fibrosis||1 in 25 white people|
|HBB c.20A>T (p.Glu7Val) (NM_000518.5)||Sickle cell disease||1 in 10 African Americans|
|BRCA1 c.68_69delAG (p.Glu23Valfs) (NM_007294.3) |
BRCA2 c.5946delT (p.Ser1982Argfs) (NM_000059.3)
|Hereditary breast and ovarian cancer, Fanconi anemia||1 in 100 Ashkenazi Jews |
1 in 62 Ashkenazi Jews
|GJB2 c.109G>A (p.Val37Ile) (NM_004004.5)||Deafness||1 in 10 east Asians|
Early work to understand the diversity of alleles in global populations was limited by high sequencing cost, quality control limitations, and lack of sufficiently powered sequencing studies in diverse ancestral groups . As the efficiency of sequencing technology has increased, genome-wide studies of larger, apparently healthy, or general population cohorts have allowed better understanding of the global allele diversity. In particular, gnomAD ( gnomad.broadinstitute.org/ ) includes exome sequencing data from more than 125,000 unrelated individuals and whole-genome sequencing for an additional 15,000 persons, with specific allele frequencies available for several subpopulations. Other databases focusing on specific subpopulations are listed in Table 2 .
|The Greater Middle East Variome||Genome sequencing of 2497 Middle Eastern individuals||igm.ucsd.edu/gme/|
|Al Mena||Genomes and exome sequencing from 2115 Arab, Middle Eastern, and North African individuals||clingen.igib.res.in/almena/|
|Institute of Precision Medicine||Genome sequencing from 100 healthy volunteers and exome sequencing from 648 normal individuals; most persons of Singaporean Chinese ancestry||beacon.prism-genomics.org/|
|Korean Variant Archive||Exome sequencing of 1055 healthy Korean individuals||www.kobic.re.kr/kova/|
|Integrative Japanese Genome Variation||Genome sequencing of >3500 Japanese individuals||ijgvd.megabank.tohoku.ac.jp/|
Population data may provide evidence leading to either a benign or pathogenic classification. A variant with population frequency greater than 5% can take a variant directly to benign through application of the ACMG/AMP stand-alone benign criterion BA1. Strong evidence toward a benign classification may also be achieved when a variant is present at a high enough allele frequency to be inconsistent with disease causation. No allele frequency threshold for BS1 (strong benign evidence) application was purposefully set so that cutoffs could be established in a disease-specific or gene-specific manner. Whiffin and colleagues published a strategy to set disease-specific population frequency cutoffs taking into consideration condition prevalence, allelic and genetic heterogeneity, and disease penetrance, and created a freely available online calculator for maximum credible allele frequency calculation ( cardiodb.org/allelefrequencyapp/ ) . Absence of a variant within a large unaffected population cohort may hint at disease association, and is designated as moderate evidence toward pathogenicity by ACMG/AMP (PM2), although caution must be taken regarding this. Although the emergence of population-specific allele frequency data is vastly improved by resources such as gnomAD, not all populations are well represented. Thus, if the particular patient is from a population that may not have been captured in existing population resources, it remains possible that the variant in question is a benign finding endemic to the patient’s ancestral group. In addition, rarity in and of itself may not be a specific predictor of pathogenicity .
For highly penetrant conditions for which onset is expected before adulthood, identifying a variant in multiple unaffected individuals suggests a benign classification (ACMG/AMP BS2). However, the natural history of the disorder as well as the demographic features of the unaffected control population being studied are important considerations.
Studying organisms and genes that have acquired changes in their nucleotide sequences is a fundamental practice in biology. In vivo or in vitro functional studies, assessing germline variant impact, is often an important aspect of making gene-disease associations and, once established, can be used in the classification of that alteration. For rare variants, robust, clinically validated functional studies are often the key to classification, especially when other lines of evidence, such as cosegregation or case-control data, are scarce .
A well-established, statistically validated functional study can be weighted as a strong line of evidence for or against pathogenicity (PS3 and BS3, respectively). The ACMG/AMP guidelines recommend that, for a well-established study, the following be considered:
The biological environment
Whether the assay reflects the full spectrum of functions of the protein
The validation, reproducibility, and robustness of the assay
Assays with high positive and negative predictive values that are reproducible and functionally appropriate can be weighted as strong evidence. However, it can be challenging to find studies that meet these criteria, and the ACMG/AMP guidelines do not provide guidance on what data are needed to deem an assay well established. After evaluating the evidence at hand, expert groups may decide to consider the body of functional evidence as strong if multiple studies replicate the same result. Alternatively, the weight of functional evidence could be decreased if appropriate functional studies are available but have not been statistically validated or replicated .
To keep up with the increasing rates of variants identified by multigene panel testing, functional studies that are both high throughput and clinically validated will be of critical importance. An early example of large-scale functional analysis was published in 2003 by Kato and colleagues , who used comprehensive site-directed mutagenesis in a yeast-based assay to evaluate all possible missense substitutions caused by single nucleotide variants throughout the p53 protein. As new technologies such as saturation mutagenesis are more widely used, multiplexed functional assays for large numbers of variants in single experiments will become more readily available . Depending on the sensitivity and specificity of the assays, these could have a significant impact for variant classification, particularly for rare variants that lack other helpful evidence.
In Silico Predictors
In the absence of functional studies, computational in silico tools may provide a lower-confidence prediction as a proxy. These tools can be grouped into those that examine potential splicing defects and those that predict whether an amino acid change, mostly missense, is likely to damage the protein. Under the ACMG/AMP framework, in silico predictors are given the lowest level of evidence (supporting) when multiple predictors are in complete concordance. However, these guidelines do not specify how many predictors should be explored or which should be used, which can lead to variant classification discrepancies between clinical laboratories because using different predictors may increase the rates of variants classified as being of uncertain significance if 1 discordant prediction prevents application of the PP3 (pathogenic evidence) or BP4 (benign evidence) criteria . Sun and Yu found that the predictive accuracy increases until an optimal performance is reached using 2 or 3 tools, but that sensitivity began to decrease when additional models were incorporated.
Variants that cause splicing defects can occur in introns or exons, and they may disrupt the consensus splice site, alter splicing regulatory sequences such as exonic and intronic splice enhancers and silencers, activate a cryptic splice site, or create an entirely new alternate splice site. Computational splicing predictors based on a variety of algorithms, including position weighted matrix, neural network analysis, maximum entropy distributions, and other machine learning methods, have been developed to predict potential splice defects . Most of the models are designed to predict changes in the strength of the splice sites for U2-type introns (GT-AG), which account for 99% of introns . Although U12-type introns (AT-AC) are rare, clinicians should be careful not to use models designed for U2-type introns to predict U12-type splice sites.
Predicting the pathogenicity of missense variants relies on a combination of features including evolutionary conservation and structural and/or functional features of the amino acid change . Each computational tool is based on different principles and has its own strengths and weaknesses , but improved performance may be seen when complementary models are combined to create a metapredictor . Given that different metapredictors may include some of the same individual predictors in their algorithms, selecting one that performs the best after testing on a set of validated pathogenic and benign variants is the preferred strategy compared with using results from several different metapredictors to avoid circularity. However, complete avoidance of circularity may be impossible, because predictors might have been trained on the variants or proteins currently under analysis, leading to overfitting .
Phenotypic Specificity and Case-Control Studies
Identifying a rare variant in an individual with a phenotype highly specific to a rare disorder may provide evidence supporting pathogenicity, in line with the ACMG/AMP PP4 criteria. For disorders with genetic heterogeneity, it may be wise to ensure that, to the extent possible, other candidate genes have also been analyzed with no pathogenic variants identified. Targeted exome studies may be particularly helpful toward this goal . The rarity of the variant is also an important consideration. If a variant is present at significant allele frequency in the general population, 1 (or more) affected individuals with the relevant phenotype will be found to harbor the variant based on chance alone.
Case-control evidence may apply when multiple affected individuals share the same variant. In order to warrant evidence toward pathogenicity, particularly for strong evidence as awarded by the ACMG/AMP PS4 criteria, cases and controls must include well-phenotyped populations; be well-matched with respect to age, gender, ancestry, and other potential confounding factors; and should include enough individuals (be sufficiently powered) to detect statistically significant odds ratios . However, given the rare nature and allelic heterogeneity observed in most mendelian diseases, it is uncommon that a case-control study for a rare mendelian disorder is sufficiently powered. For this reason, the ACMG/AMP criteria allow for a relaxed use of the PS4 case-control criteria to count multiple unrelated probands showing phenotypic specificity with the same variant.
Cosegregation with Disease
Cosegregation of a variant with disease is a valuable quantitative data point used in assessing the pathogenicity of germline variants, particularly for rare mendelian disorders. This concept is similar to genetic linkage analysis in that likelihood ratios are derived similarly to logarithm of the odds scores in linkage, but specifically concerning the segregation of the variant and not a linked marker . The seminal linkage models, developed from the 1950s through the 1980s , are based on the degree of relationship of the probands and the genotype and disease status of the tested relatives. The derived likelihood ratio is also a function of an assumed penetrance that includes the risk of disease and can be stratified by age and/or sex as well as the frequency of the disease allele. Jarvik and Browning proposed a simpler approach for use with qualitative frameworks such as the ACMG/AMP guidelines. Given that segregation supports linkage to the allele as opposed to a specific variant, the strength of evidence assigned for segregation data may be adjusted based on the number of genes associated with the disease and degree of certainty that all potential disease-causing variants were identified during the testing process.
De Novo Occurrence
Identification of de novo variants has been greatly accelerated by the emergence of exome and genome trio sequencing, in which concordant analysis of the affected proband along with both parents can identify genetic variants unique to the proband while concordantly excluding nonpaternity and/or nonmaternity, providing evidence supporting a strong pathogenic criterion (PS2) per ACMG/AMP . Before the advent of trio testing as a means to confirm parentage, separate analyses, such as microsatellite marker studies, could prove parentage but may not have been pursued. A nonpaternity rate of 1.9% has been identified among men with high paternity confidence, and higher rates (some greater than 10%) have been reported in other studies . Thus, the ACMG/AMP guidelines give a lower, moderate evidence strength (PM6) for an assumed de novo variant in the absence of studies confirming parentage.
Although de novo occurrence can be helpful evidence supporting pathogenicity, any newborns may have up to 100 de novo variants within their genomes, with 1 or 2 predicted to alter the coding sequence . For genes with expansive coding sequences, de novo variation has a higher a priori risk of occurring by chance . Thus, the requirement that the patient’s phenotype be consistent with the disease caused by the gene in question is key to the appropriate application of de novo criteria. The ClinGen Sequence Variant Interpretation Working Group has crafted guidelines for tempering the strength of evidence applied for de novo occurrence when phenotype specificity is lacking or the disorder has high heterogeneity, or for increasing strength when multiple probands with de novo occurrence are identified ( clinicalgenome.org/site/assets/files/3461/svi_proposal_for_de_novo_criteria_v1_0.pdf ).
Other types of evidence that may play a role in variant classification include:
Presence in a mutational hotspot functional domain: moderate pathogenic evidence may be considered if a variant occurs in a critical functional domain or region known to harbor multiple pathogenic missense variants without any benign variation, indicating functional importance (ACMG/AMP PM1).
Allelic data: whether a second known pathogenic variant occurs on the same (in cis) or opposite (in trans) allele as the variant of interest. Use of this criterion depends on the inheritance and mechanism of disease.
For autosomal recessive disorders, a variant of interest appearing in trans with a known pathogenic variant supports pathogenicity (ACMG/AMP PM3).
For dominant conditions for which biallelic pathogenic variants are incompatible with life or cause a severe phenotype, finding a variant in trans with a known pathogenic variant supports a benign classification (ACMG/AMP BP2).
Variants in cis or with unknown phase require a thoughtful approach. A variant occurring in cis with a variant predicted to result in nonsense-mediated decay would not be expressed, or 2 missense variants occurring on the same allele may work synergistically. If phase is unknown, other observations of the same 2 rare variants together in multiple individuals suggests they are in cis; likewise, observations of each variant independent of the other, or co-occurring with other variants, suggest they are in trans.
Co-occurrence with other pathogenic variants: observation of a variant in an individual harboring other pathogenic variants explaining the patient’s phenotype is an evidence against pathogenicity (ACMG/AMP BP5). Caution should be exercised for phenotypes with significant heterogeneity (eg, breast cancer) because there are multiple reports of individuals with rare pathogenic variants in more than 1 breast cancer predisposition gene .
Present relevance and future avenues to consider or to investigate
The Shift to Quantitative and Bayesian Approaches
This article outlines a primarily qualitative, categorical approach to variant assessment. This approach is purposefully general so that the ACMG/AMP criteria can be used for classification of variants in any mendelian disorder. For well-studied genes with validated functional studies, in silico models, and clinical correlates, it is possible to make gene-specific variant classification criteria using a quantitative approach. Calls for variant interpretation to be driven by a quantitative, bayesian approach have echoed in the literature for more than a decade, when bayesian approaches began to be used to classify variants in cancer predisposition genes . More recently, Tavtigian and colleagues developed a bayesian framework that takes the current ACMG/AMP categories and shifts them from a qualitative to a quantitative structure ( Table 3 ). Among the benefits of this system are the opportunity to combine pathogenic and benign criteria when determining classification, as well as permitting clinicians to more clearly recognize where on the spectrum of benign to pathogenic a variant of uncertain significance may be.