Overview of specifications to the ACMG/AMP variant interpretation guidelines

Abstract

The 2015 ACMG/AMP guidelines established a classification system for sequence variants; however, the broad scope of these guidelines necessitates specification of evidence types for specific genes or diseases of interest. Since publication of the guidelines, both general use and disease-focused specifications have emerged to aid in accurate application of ACMG/AMP evidence types. This protocol summarizes the approaches to, and rationale for, specifying three evidence categories (population frequency data, variant type and location, and case-level data), including available resources and a quantitative framework that can inform the specification process.

INTRODUCTION

In 2015, the American College of Medical Genetics and Genomics (ACMG) and the Association for Molecular Pathology (AMP) published a guideline that provides a framework for sequence variant interpretation (Richards et al., 2015). The guideline defined 28 criteria, with codes that addressed types of variant evidence. Each evidence type or criterion code was assigned a direction, benign (B) or pathogenic (P), and a level of strength: stand‐alone (A), very strong (VS), strong (S), moderate (M), or supporting (P). Combining rules for these criteria were also provided to assign a pathogenicity assertion for each sequence variant based on applied criteria. More than 95% (62/65) of surveyed laboratories reported using the ACMG/AMP five tiers for classifying variants in genes associated with traits with Mendelian patterns of inheritance and 97% (62/64) of laboratories were using approaches for variant interpretation consistent with the guidelines (Niehaus et al., 2019).

The ACMG/AMP guidelines were developed to be broadly applicable across many genes, inheritance patterns and diseases (and thus necessarily generic). They anticipated “that those working in specific disease groups should continue to develop more focused guidance regarding the classification of variants in specific genes given that the applicability and weight assigned to certain criteria may vary by gene and disease” (Richards et al., 2015). The NIH-funded Clinical Genome Resource (ClinGen) consortium was formed in 2013 to develop standards and processes for evaluating genes and genomic variation to enhance clinical validity and utility (Rehm et al., 2015). As a core goal of ClinGen is expert interpretation of variants, ClinGen has convened Variant Curation Expert Panels (VCEPs) that focus on a gene or group of genes. The VCEPS are tasked with providing specifications to the ACMG/AMP guidelines for gene-disease dyads, interpreting variants according to these rules, and publishing the interpretations through NCBI’s publicly available ClinVar resource (Rivera-Muñoz et al., 2018). In addition, the gene–disease dyad criteria specifications can then also be used by the community to enhance consistency in variant interpretation. As of this writing, ClinGen VCEPs have published disease-specific guidance for MYH7-related inherited cardiomyopathy (Kelly et al., 2018), RASopathies (Gelb et al., 2018), hearing loss (Oza et al., 2018), PAH (Zastrow et al., 2018), PTEN (Mester et al., 2018), and CDH1 (Lee et al., 2018), with multiple VCEPs in other disease areas in the process of specifying the guidelines (https://www.clinicalgenome.org/working-groups/clinical-domain/).

In additional to the gene-disease dyad VCEPS, ClinGen also established the Sequence Variant Interpretation (SVI) working group to refine and evolve the ACMG/AMP guidelines for accurate and consistent clinical application, and to harmonize disease‐focused specification of the guidelines by VCEPs (https://www.clinicalgenome.org/svi/). As seen in Table 1, SVI has published recommendations for multiple evidence types, which will be summarized in more detail in the relevant sections below, including suggested removal of the reputable source criteria (PP5, BP6) from the ACMG/AMP framework (Biesecker & Harrison, 2018). Additionally, SVI has published a Bayesian framework which provides a quantitative approach to specifying the ACMG/AMP guidelines (Tavtigian et al., 2018).

Table 1.

ACMG/AMP evidence codes for classifying sequence variants organized by data type and strength

	Benign Criteria		Pathogenic Criteria
	Strong	Supporting	Supporting	Moderate	Strong	VeryStrong
	−18.7	−2.08	2.08	4.33	18.7	350.0
Population Frequency Data	*BA1 G,S BS1 G			PM2G
Variant Type and Location		BP1 BP3 BP7	PP2G	PM1G PM4 PM5	PS1	PVS1G,S
Case-level data	BS2G BS4	BP2 BP5	PP1P PP4G	PM3S PM6S	PS2S PS4G
Functional and Computational data	BS3G	BP4P	PP3P		PS3G,P
Reputable source		BP6R	PP5R

^*BA1 is considered a “stand-alone” criterion and was not included in the Bayesian model. G, criteria requiring gene-specific information; S, criteria with SVI guidance in place; P, criteria with SVI guidance in progress; R, criteria suggested for removal by SVI.

This unit is oriented toward individuals who perform variant interpretation, set policy for diagnostic laboratories regarding implementation of the Richards et al criteria, and VCEPs who are adapting Richards et al for specific gene-disease dyads. We will focus on approaches undertaken by existing VCEPs, the SVI, and other groups to specify the ACMG/AMP guidelines for a disease of interest.

Quantitative Approach

Adaptation of a quantitative framework to the ACMG/AMP guidelines will allow for refinement to the relevant strength of each piece of evidence, rather than the current “Met”/”Not Met” approach to each evidence type. Towards this goal, SVI evaluated the ACMG/AMP framework for compatibility with Bayesian statistical reasoning (Tavtigian et al., 2018). A high level of compatibility with the ACMG/AMP combing criteria was observed when scaling the relative strength of the ordered evidence categories to the power of 2.0, as multiple pairs of the ACMG/AMP pathogenic combining criteria were related to each other by this factor, as, for example, two strong criteria were equivalent to one very strong criterion. The resulting relative odds of pathogenicity for Supporting, Moderate, Strong, and Very Strong pathogenic evidence were estimated to be 2.08:1, 4.33:1, 18.7:1, and 350:1, respectively (Table 1). This quantitative framework provides opportunities to further refine evidence categories and combining rules, and can support evaluation of appropriate strength-level criteria modifications to the guidelines. For example, if assessment of a functional assay for PS3 application indicates that ~90% of variants with damaging calls from the assay are truly pathogenic, then a moderate strength level is appropriate (4:33:1 odds or ~81% accuracy) as this assay does not meet the strong strength level threshold (18.7:1 odds or ~95% accuracy). We consistently advocate for conservative interpretations of all forms of evidence to avoid overestimating the pathogenicity of a variant.

POPULATION FREQUENCY CRITERIA

Dataset Ascertainment

Information regarding the population frequency of a variant can be used as evidence both for and against pathogenicity. Currently, the Genome Aggregation Database (gnomAD) is the largest and most comprehensive publicly available datasets of allele frequencies (Lek et al., 2016). However, regardless of the dataset used, it is necessary to understand the ascertainment approach for each dataset to determine whether individuals suffering from a specific disease of interest are expected to be present. For instance, the majority of individuals in gnomAD are older (mean age 54 years old) and efforts have been made to exclude individuals with severe pediatric diseases and their first-degree relatives (Lek et al., 2016). Thus, it is unlikely that such individuals are in the dataset, and most variants in the dataset should not be pathogenic for such disorders inherited in an autosomal dominant pattern. For adult-onset or reduced-penetrance conditions, individuals in gnomAD should be conservatively viewed as general population and not healthy controls. For a disorder like dilated cardiomyopathy, one must assume that there are pathogenic variants in associated genes at least at the level of the population frequency of that disorder.

Allele Frequency Considerations

A variant occurring at a much higher population frequency than expected for the disease it is associated with is such a significant indicator of a benign impact that the ACMG/AMP guidelines awarded this piece of evidence as “Benign stand-Alone” (BA1). This means that if a variant were to meet the BA1 criterion, it could be considered benign without the need for assessing other evidence. The BA1 criterion in the ACMG/AMP guidelines was defined as “Allele frequency is >5% in Exome Sequencing Project, 1000 Genomes Project, or Exome Aggregation Consortium.” In 2018, ClinGen SVI proposed an updated definition of BA1 to “Allele frequency is >0.05 in any general continental population dataset of at least 2,000 observed alleles and found in a gene without a gene- or variant-specific BA1 modification” (Ghosh et al., 2018). This proposed updated definition allowed use of subpopulation groups while avoiding errors that may arise in allele frequency estimation due to population diversity and size (i.e., incorrectly concluding a variant is benign due to a high allele frequency in a small or bottlenecked population). To aid in proper usage of allele frequency data, ExAC and gnomAD have included a “filtering allele frequency” (FAF) annotation, which is the highest true population allele frequency for which the upper bound of the 95% confidence interval of allele count under a Poisson distribution is still less than the variant’s observed allele count (Whiffin et al., 2017). In practice, this annotation functions as equivalent to a lower bound estimate for the true allele frequency of an observed variant given a population size (i.e., the FAF for a given variant will always be lower than the calculated allele frequency). In both ExAC and gnomAD, the highest FAF from any continental population is displayed. To control against incorrect usage due to population diversity or size, gnomAD does not calculate FAF for non-continental populations (Finnish and Ashkenazi Jewish subpopulations) nor is FAF calculated for continental populations with a singleton observation.

Calculating and Applying BA1/BS1 Thresholds

A second important caveat added to SVI’s updated definition of BA1 is the recognition that the proposed >0.05 threshold for calling a variant benign is an order of magnitude higher than necessary for many genes and disorders. To calculate a gene- or disease-specific threshold for BA1 or BS1 (“Allele frequency is greater than expected for disorder”) it is important to consider disease prevalence, genetic (locus and allelic) heterogeneity, and penetrance. With these parameters, thresholds can be determined by calculating the maximum credible population allele frequency for a disease, defined as: (prevalence x heterogeneity) / penetrance (Whiffin et al., 2017). Table 2 displays the varying prevalence, penetrance, and heterogeneity input values used by ClinGen VCEPs to calculate both BA1 and BS1 thresholds. For the majority of VCEPs (Cardiomyopathy, RASopathy, Hearing Loss - Autosomal Recessive, and PAH), BA1 and BS1 calculations were differentiated by varying heterogeneity values (locus vs. allelic), while prevalence and penetrance values were kept constant. Alternatively, two VCEPs (CDH1 and Hearing Loss - Autosomal Dominant) kept heterogeneity and penetrance values constant while varying prevalence inputs between BA1 and BS1. These differences likely reflect different degrees of confidence with the reported values of each variable. While approaches to prevalence, penetrance, and heterogeneity values may have differed among VCEPs, all VCEPs followed our preference of using the most conservative estimates for BA1 calculation and less conservative estimates for BS1 calculation.

Table 2.

Comparison of population frequency thresholds from ClinGen Variant Curation Expert Panels.

	Criteria	Prevalence	Heterogeneity	Penetrance	Threshold
Cardiomyopathy (AD)	BA1	1:200	10.60% L	30%	≥0.001 (0.1%)
	BS1		2% A		≥0.0002 (0.02%)
	PM2	1:500		50%	<0.00004 (0.004%)
RASopathy (AD)	BA1	1:2500	100%	40%	≥0.0005 (0.05%)
	BS1		50% L		≥0.00025 (0.025%)
	PM2	-	-	-	Absent R
CDH1 (AD)	BA1	1:800	100%	30%	≥0.002 (0.2%)
	BS1	1:1250			≥0.001 (0.1%)
	PM2	-	-	-	<0.00001 (0.001%) R
Hearing Loss (AD)	BA1	1:30	5% L/A	80%	≥0.001 (0.1%)
	BS1	1:150			≥0.0002 (0.02%)
	PM2	-	-	-	<0.00002 (0.002%) M
Hearing Loss (AR)	BA1	1:200	7.2% A	100%	≥0.005 (0.5%)
	BS1		4.4% A		≥0.003 (0.3%)
	PM2	-	-	-	<0.00007 (0.007%) M
PAH (AR)	BA1	1:5000	90% L	80%	≥0.015 (1.5%)
	BS1		2% A		≥0.002 (0.2%)
	PM2	-	-	-	<0.0002 (0.02%) M
PTEN* (AD)	BA1	-	-	-	≥0.01 (1%)
	BS1	-	-	-	≥0.001 (0.1%)
	PM2	-	-	-	<0.00001 (0.001%) R

AD, autosomal dominant; AR, autosomal recessive; L, heterogeneity input is based on disease contribution of the most commonly implicated gene/locus; A, heterogeneity input is based on disease contribution of the most common pathogenic variant; R, VCEP defined PM2 threshold based on overall rarity; M, VCEP defined PM2 threshold as allele frequency an order of magnitude below BS1 [for HL-AR, PM2 is an order of magnitude below BS1_Suppoting (not shown)].

^*PTEN VCEP began specification of BA1/BS1 before release of Whiffin et al calculator and thus rationale does for thresholds does not fit it defined parameters above.

With this approach of defining BA1 and BS1 thresholds based on the maximum credible population allele frequency for a disease, these thresholds can then be compared to the FAF for all variants in the genes of interest. If a variant has a FAF above BA1 or BS1 thresholds for a condition of interest, then BA1 or BS1 can confidently be applied to that variant. For example, using allele frequency data from ExAC, the Cardiomyopathy VCEP applied BA1 criterion (≥0.001) to all 46 variants in MYH7 that had a FAF >0.001 (Kelly et al., 2018).

Calculating and Applying PM2 Thresholds

Determining if modifications to PM2 criterion phrasing (“Absent from controls, or at extremely low frequency if recessive, in Exome Sequencing Project, 1000Genomes Project, or Exome Aggregation Consortium”) are necessary can also vary by gene and disease depending on whether or not affected individuals are expected to be present in large population datasets. For adult-onset conditions, affected individuals may be present in the dataset and for these reasons, approaches to calculating a PM2 threshold are not as consistent as those for BA1 and BS1 (Table 2). Approaches for determining a PM2 threshold have included: calculating the maximum credible population allele frequency using more realistic and less conservative values than BA1/BS1 (Cardiomyopathy VCEP), assumption that 0–2 individuals in gnomAD would have disease (0 – 0.002%) based on prevalence and frequency of known pathogenic variants (RASopathy, CDH1, and PTEN), and selecting a threshold that is numerically an order of magnitude smaller than the BS1 threshold (Hearing Loss and PAH).

Selecting a PM2 threshold that is simply the inverse of BS1 (i.e., <0.0002 for PM2; >0.0002 for BS1) is discouraged because in that scenario the difference of a single allele count in a population dataset would move a variant from moderate strength for pathogenicity to strong benign evidence strength. This notion is supported by the Bayes classification framework in which a strong piece of benign evidence has 1:18.7 odds of pathogenicity while a moderate piece of pathogenic evidence has 4.33:1 odds of pathogenicity (Table 1), suggesting a “grey zone” in allele frequency thresholds between BS1 and PM2 in which no criteria is applied is necessary.

Lastly, regardless of the calculated PM2 threshold, “absence” should only be assumed if the variant type is expected to be detected by large population datasets (as large indels and variants in repetitive or low complexity regions may be missed or incorrectly annotated), if read depth at that position is sufficient for an accurate call, and if the individual harboring the identified variant is of a genetic background represented in the dataset. If any of these considerations are not met, then the significance of a variant being “absent” from the dataset is unknown, and this criterion should not be applied.

Future Directions

Given the range of expected variant allele frequencies, moving towards a spectrum approach in which allele frequency data is not limited to Benign Stand Alone (BA1), Benign Strong (BS1), and Pathogenic Moderate (PM2) evidence types is anticipated. The ClinGen Hearing Loss VCEP accounted for this issue by also creating and defining BS1_Supporting and PM2_Supporting strength levels for allele frequency data, recognizing that while a variant allele frequency may not meet a stringent threshold for BS1, for instance, the frequency may be high enough that some evidence towards benign interpretation is warranted (BS1_Supporting). This notion of additional allele frequency strength levels is also included in the Invitae/Sherloc specifications to the ACMG/AMP guidelines which include five allele frequency categories, “Absent” and “Pathogenic range” on the Pathogenic side and “Very high”, “High”, and “Somewhat high” on the Benign side, supporting this spectrum approach to allele frequency data (Nykamp et al., 2017).

Additionally, concerns have been raised that PM2 is given too much weight in the framework. This concern is supported by findings from ExAC that 99% of identified high-quality variants have a frequency <1% and that 54% of identified high-quality ExAC variants are only seen once in the entire data set, suggesting that rarity is actually common. This concept of absence or within pathogenic range is an important factor during variant assessment but may be better suited as a consideration or caveat for application of other criteria, such as taking into account allele frequency before applying PS4 for observation of the variant in multiple affected individuals or insuring an LoF variant does not have an allele frequency inconsistent with the expected prevalence. However, the combining rules outlined in the ACMG/AMP guideline requires multiple evidence types be met in order to reach a classification, meaning removal or downgrading the weight of PM2 would have a major impact on variant classification. Due to those constraints, a general recommendation that PM2 weight be decreased for all diseases cannot be endorsed until the combining rules or inclusion of PM2 into other evidence types are considered.

VARIANT TYPE AND LOCATION

General Constraint Considerations

To accurately apply ACMG/AMP criteria during variant assessment, knowledge about the variant spectrum for the gene and disease area is necessary (Table 1 - Annotations with “G”). By comparing expected variation, due to gene size and sequence, to observed variation in unselected or healthy populations, data from large population datasets such as ExAC and gnomAD can give insight into the tolerance of our species for specific variant types or tolerance for variation in specific regions (Lek et al., 2016; Samocha et al., 2014). However, this approach reflects tolerance due to reproductive fitness and viability, which may not accurately reflect molecular mechanisms for all genes and diseases, thus necessitating detailed guidance and expert curation (Strande, Brnich, Roman, & Berg, 2018).

Loss of Function Variants

In the ACMG/AMP guidelines, the only criterion designated with Very Strong strength level for pathogenicity is PVS1, defined as “null variant (nonsense, frameshift, canonical ±1 or 2 splice sites, initiation codon, single or multi-exon deletion) in a gene where loss-of-function (LoF) is a known mechanism of disease” (Richards et al., 2015). Given the weighting of this criterion as Very Strong and the potential impact of inappropriate usage, SVI published guidance on application of PVS1 that included a decision tree format to determine the applicable strength of PVS1 by assessing variant type, predicted consequence, and location within the gene (Abou Tayoun et al., 2018). Additionally, this recommendation reiterated the concept that the PVS1 criterion should only be applied for genes where LoF is an established disease mechanism; however determination of whether LoF is an established disease mechanism can be subjective. Several resources, including the ClinGen haploinsufficiency (HI) score (Riggs et al., 2012) and the ExAC/gnomAD probability of LoF intolerance (pLI) score and observed/expected score (Lek et al., 2016), can help determine if LoF is a potential disease mechanism. The HI scores are divided into six tiers based on manually curated evidence, with a score of “3” indicating “Sufficient evidence suggesting dosage sensitivity is associated with clinical phenotype”; however, not all genes have undergone this manual curation (https://www.ncbi.nlm.nih.gov/projects/dbvar/clingen/). The pLI score in computationally derived and measures the intolerance of a given gene to LoF variants in the general population, with a pLI >0.9 suggesting a significantly lower than expected rate of LoFs in the gene (Lek et al., 2016). However, with release of v2.1, gnomAD now recommends using the observed / expected (oe) score to measure expected LoF intolerance, in place of pLI scores (https://macarthurlab.org/2018/10/17/gnomad-v2–1/). While a higher pLI score (i.e., closer to 1.0) indicated strong intolerance of LoF variants, lower oe values (i.e., closer to 0) are indicative of strong LoF intolerance. The change from pLI to oe with a 90% confidence interval (CI) was intended to facilitate easier interpretation and continuity across the spectrum of selection. Additionally, a threshold of <0.35 for the upper bound of the oe confidence interval was suggested as a measure of significant LoF depletion (similar to the pLI >0.9 threshold).

Table 3 compares variant spectrum curations and annotations for 22 genes from ClinGen VCEPs. Comparing HI and oe annotations to ClinGen VCEP curations for PVS1 applicability identifies key considerations. First, disease context is critical given that LoF may be a mechanism for one disease caused by a given gene but not another disease caused by that same gene. For example, for PTPN11 the RASopathy VCEP indicated that PVS1 is not applicable, although oe score is 0.03 (and upper bound of 90% CI is <0.35 threshold) suggesting LoF intolerance and HI score is 3 (“Sufficient evidence suggesting dosage sensitivity is associated with clinical phenotype”). However, the HI score for PTPN11 is based on metachondromatosis, a phenotype unrelated to RASopathy. The PVS1 assessment of applicability from the RASopathy VCEP is specific to assessing variants in PTPN11 with regards to pathogenicity for RASopathy-related conditions. Second, the oe score is only considered a valuable metric for genes that cause a disease inherited in an autosomal dominant pattern via haploinsufficiency, as all genes associated with diseases inherited in an autosomal recessive and PVS1 marked as applicable in Table 3 have an upper bound of oe 90% CI >0.35, suggesting no significant depletion of LoF variation. This makes sense, as there is typically little or no selective pressure against the heterozygous (carrier) genotype for such disorders. ExAC and gnomAD calculate a probability score of a gene being intolerant to biallelic loss of function (pRec); however, those authors caution against using pRec as a metric for determining if recessive genes are tolerant of LoF (Lek et al., 2016). Lastly, as oe scores reflect gene tolerance relative to reproductive fitness, age of onset and disease severity need to be considered when assessing whether depletion of the disease-causing variant type is expected in population datasets. For example, for PTEN, CDH1, and MYO6, the upper bound of the oe CI exceeds the suggested <0.35 threshold for measuring LoF intolerance; however each VCEP indicates PVS1 as applicable given that the expected phenotype is not pediatric-onset or is unlikely to impact reproductive fitness, meaning affected individuals are not expected to be excluded from gnomAD. These data suggest that while oe CI <0.35 can help identify genes expected to be intolerant to LoF variation, this score will not identify all LoF intolerant genes and that oe should be viewed as a continuous value rather than a strict threshold. It is hoped that future implementations of a continuous, fully quantitative system can convert these frequencies into odds of pathogenicity.

Table 3.

Missense and LoF annotations and curations per gene from ClinGen Variant Curation Expert Panels

Gene	Disease Area (MOI)	HI Score	gnomAD LoF oe metric (90% CI)	PVS1?	Missense Z score (ExAC / gnomAD)	PP2?
MYH7	Cardio (AD)	0	0.45 (0.35–0.57)	Yes (Mod)	6.54 / 3.93	No
BRAF	RAS (AD)	1	0.1 (0.05–0.21)	No	3.99 / 3.72	Yes
HRAS		0	0.36 (0.16–0.93)	No	2.69 / 1.51	Yes
KRAS		0	0.63 (0.34–1.24)	No	1.36 / 2.32	Yes
MAP2K1		0	0.15 (0.07–0.38)	No	3.43 / 3.11	Yes
MAP2K2		1	0.1 (0.04–0.33)	No	1.48 / 1.87	Yes
PTPN11		3	0.03 (0.01–0.14)	No	3.43 / 3.13	Yes
RAF1		0	0.19 (0.11–0.35)	No	2.82 / 2.46	Yes
SHOC2		-	0 (0.00–0.14)	No	2.57 / 2.97	Yes
SOS1		0	0.07 (0.03–0.14)	No	2.18 / 3.05	Yes
PTEN	PHTS (AD)	3	0.24 (0.13–0.51)	Yes	3.71 / 3.49	Yes
CDH1	HDGC (AD)	3	0.25 (0.15–0.43)	Yes	0.81 / 0.71	No
PAH	PKU (AR)	30	1.12 (0.84–1.50)	Yes	−1.54 / −0.65	No
CDH23	HL (AR)	30	0.38 (0.26–0.57)	Yes	−0.24 / 0.71	No
GJB2		-	2.62 (1.39–1.98)	Yes	−1.07 / 1.17	No
MYO7A		-	0.7 (0.58–0.85)	Yes	−1.44 / 1.07	No
SLC26A4		-	0.89 (0.68–1.18)	Yes	−3.23 / −2.01	No
TECTA		30	0.45 (0.35–0.58)	Yes	2.3 / 1.61	No
USH2A		30	0.76 (0.67–0.86)	Yes	−5.12 / −2.47	No
COCH	HL (AD)	-	0.59 (0.40–0.91)	No	0.34 / 0.68	No
KCNQ4		-	0.22 (0.12–0.41)	Yes	2.73 / 1.83	No
MYO6		-	0.3 (0.22–0.42)	Yes	1.02 / 1.39	No
TECTA		30	0.45 (0.35–0.58)	No	2.3 / 1.61	No

- in HI column indicates awaiting review. MOI, mode of inheritance; AD, autosomal dominant; AR, autosomal recessive; HI, haploinsufficiency; Cardio, inherited cardiomyopathy; RAS, RASopathy; PHTS, PTEN hamartoma tumor syndrome; HDGC, hereditary diffuse gastric cancer; PKU, phenylketonuria; HL, hearing loss. Data from ExAC v1.0 and gnomAD v2.1.1

Missense Constraint and Hotspot/Functional Domains

For missense variation, differences in the expected and observed number of rare missense variants in population databases can be informative re pathogenicity if a gene demonstrates constraint for missense variation and if that constraint is observed across the entire gene (PP2) or only confined to specific regions of the gene (PM1). Based on data from ExAC, a missense constraint Z score >3.09 is equivalent to a p-value of 10⁻³ and is considered a significant threshold when splitting transcripts into constrained (meaning significantly depleted of missense variation) and unconstrained classes (Lek et al., 2016). ClinGen VCEPs have been encouraged to use this information when determining if ACMG/AMP criterion PP2 is applicable for their gene of interest (Table 3). However, similar to PVS1 and oe or pLI scores, caution should be placed on relying on the missense constraint Z score alone to determine if PP2 is applicable. One potential area of concern is applying PP2 for all missense variants in a gene when the constraint observed is isolated to a specific region of the gene. For example, MYH7 has a missense constraint Z score of 6.54 and 3.93 in ExAC and gnomAD, respectively, suggesting significant depletion of missense variation (or missense constraint). However, PP2 was marked not applicable for this gene by the Cardiomyopathy VCEP, as recent studies suggest that this high constraint is actually driven by a statistically significant clustering of pathogenic variants in the head region of the protein (Homburger et al., 2016; Walsh et al., 2017). Based on these data, the Cardiomyopathy VCEP concluded this evidence was most appropriately weighted as moderate (PM1) through application of the critical domain rule (Kelly et al., 2018). Specification of PM1 has required gene-by-gene curation by disease experts, combined with laboratory data to determine regions intolerant to variation. These findings suggest that while missense constraint Z score can help inform application of PP2, additional curation is necessary, specifically with regards to PM1.

Future Directions

Building on the constraint data from population databases alone, Walsh et al proposed a quantitative approach to PP2/PM1 (concepts merged into PM1 criterion with variable strength) that compared rare variants from hypertrophic cardiomyopathy cohorts to reference population databases to identify variant classes and regions significantly enriched in cases (Walsh et al., 2019). As this approach provides a quantitative estimate of the probability that a rare variant is causative (dependent on the gene, variant class and variant location within the gene/protein), thresholds for Supporting, Moderate, and Strong strength levels for PM1 were proposed based on the SVI’s Bayesian adaptation of the ACMG/AMP guidelines (Tavtigian et al., 2018). This proposal is dependent on the presence and availability of large cohorts of affected individuals to compare to population databases, which may not be available for all disorders; however, the proposal establishes a framework that can help move PP2 and PM1 decision making from qualitative to quantitative.

CASE-LEVEL DATA

Case-Level Data Overview

Case-level data include information regarding the individual who harbors the variant, such as phenotype (PP4 and BS2) and other detected variants of interest (PM3, BP2, and BP5), the individual’s family data, such as segregation (PP1 and BS4) and presence in parental samples (PS2 and PM6), and lastly whether the variant has been detected in other affected unrelated individuals (PS4). Applying these criteria based on a single individual or family may be straightforward; however, when applying these criteria based on evidence from multiple individuals and families, uncertainty can arise in combining observations and determining the applicable strength of evidence. To aid in consistency, ClinGen VCEPs and SVI have begun adopting quantitative approaches to these criteria.

Quantitative Approaches to Case-Level Data

As general recommendations for usage, ClinGen SVI has released quantitative-based approaches for both de novo occurrences (PS2/PM6) and occurrences of the variant in trans with a pathogenic variant (PM3). As stated in the ACMG/AMP guidelines, a de novo occurrence can be used as evidence supporting pathogenicity, dependent on the phenotype in the patient matching the gene’s disease association with reasonable specificity, with the strength of a de novo occurrence varying based on whether tested parental samples are confirmed (PS2) versus assumed (PM6) to be from the biological parents of the patient. It is important to note that “confirmed” and “assumed” for de novo occurrences does not refer to parental samples being “confirmed” or “assumed” to be negative for the putative pathogenic variant. In both scenarios the parent samples have been tested and are negative for the variant; the “confirmed” and “assumed” terms pertain to whether or not the tested samples were shown through identify testing to be the biological parents of the patient. This may be done most simply in cases of trio exome or genome sequencing by confirming that there are very few Mendelian inconsistencies in the trio. Otherwise, biologic parentage can be confirmed by STRP analysis.

Given that a variant may have arisen de novo in multiple individuals with varying parental confirmation scenarios and varying phenotypes, ClinGen SVI established a quantitative approach to determine the appropriate strength level to apply for de novo data (https://www.clinicalgenome.org/site/assets/files/3461/recommendation_ps2_and_pm6_acmgamp_critiera_version_1_0.pdf). With this approach, each proband with a de novo variant is awarded a point value based upon phenotypic consistency and confirmed or assumed de novo status (Table 4B). The combined point value of all probands with de novo occurrences is then compared to the PS2/PM6 point thresholds (Table 4A) to determine the applicable evidence strength level.

Table 4.

Overview of case-level data specifications

4A. Point value thresholds per strength level for PS2/PM6 and PM3 and proband count thresholds per Variant Curation Expert Panel for PS4.
		Supporting	Moderate	Strong	VeryStrong
PS2/PM6 point thresholds		0.5	1.0	2.0	4.0
PM3 point thresholds		0.5	1.0	2.0	4.0
PS4	Cardiomyopathy	2 probands	6 probands	15 probands	N/A
	RASopathy	1 proband	3 probands	5 probands	N/A
	PTEN	1 point	2 points	4 points	16 points
	CDH1	1 proband	2 probands	4 probands	16 probands
	Hearing Loss (AD)	2 probands	6 probands	15 probands	N/A
4B. Points awarded per proband with a de novo variant
Phenotypic consistency	Points per Proband
	Confirmed de novo	Assumed de novo
Phenotype highly specific for gene	2	1
Phenotype consistent with gene but not highly specific	1	0.5
Phenotype consistent with gene but not highly specific and high genetic heterogeneity*	0.5	0.25
Phenotype not consistent with gene	0	0
4C. Points awarded per proband with variant in trans
Classification/Zygosity of other variant	Points per Proband
	Confirmed in trans	Phase unknown
Pathogenic or Likely pathogenic variant	1.0	0.5 (P) 0.25 (LP)
Homozygous occurrence (max point 1.0)	0.5	N/A
Uncertain significance variant (max point 0.5)	0.25	0.0

For disorders inherited in an autosomal recessive pattern, criterion PM3 can be applied if the variant in question has been detected in trans with a pathogenic variant in an individual affected with such a disorder. However, no directions were provided on how to increase PM3 weight based on multiple occurrences or to decrease PM3 weight if phasing is unknown (Amendola et al., 2016). To aid in consistency in application of PM3, ClinGen SVI proposed a quantitative approach to determine the appropriate PM3 strength level to apply for in trans occurrences (https://clinicalgenome.org/site/assets/files/3717/svi_proposal_for_pm3_criterion_-_version_1.pdf). Similar in scope to the de novo point system described above, with this proposal each proband is awarded a point value based upon phasing of the two variants in question (confirmed in trans versus unknown) and classification of the variant on the other allele (Table 4C). Point values from all probands are then summed and compared to PM3 point thresholds (Table 4A) to determine the applicable strength level.

Disease Specification and Alternative Usage of PS4 Criterion

If the prevalence of a variant in affected individuals is significantly increased compared to controls, based on suggested odds ratio or relative risk >5.0, criteria PS4 can be applied. However, for many diseases or individual variants, sufficient data is unavailable to perform significant case-control studies. The ACMG/AMP guidelines provides a method for this data type to still be utilized as PS4 evidence by noting “in instances of very rare variants where case–control studies may not reach statistical significance, the prior observation of the variant in multiple unrelated patients with the same phenotype, and its absence in controls, may be used as moderate level of evidence” (Richards et al., 2015). ClinGen SVI endorsed this notion and recommended VCEPs specifying the ACMG/AMP guidelines for autosomal dominant disorders establish proband counting thresholds for varying PS4 strength levels (VCEPs specifying autosomal recessive disorders were encouraged to use PM3 instead of PS4 for case counts - see above). With this suggestion, VCEPs have specified PS4 usage to include counts of multiple unrelated probands with consistent phenotypes. Differences in the number of probands needed for each strength level has varied by VCEP according the specificity of the phenotype/disease of interest (Table 4A). As PTEN and CDH1 specifications were developed after the Bayes adaptation of the ACMG/AMP guidelines, both VCEPs scaled the relative strength of PS4 to the power of 2.0. Additionally, while the majority of VCEPs defined strength levels based on the number of probands with consistent phenotypes, the PTEN VCEP instead used a point system in which the weight of each proband varied based on the specificity of the phenotypic features (Mester et al., 2018).

Application of Case-Level Criteria

As noted in the relevant sections above, recommendations for PS4, PM3, and PS2/PM6 allow multiple proband occurrences to be combined using a quantitative framework to determine the applicable criterion strength level. Application of these criteria, or any of the case-level data criteria (Table 1), is not limited to evidence from the case currently being assessed. For example, the de novo occurrence of a PTPN11 variant in the literature could meet PS2/PM6 criterion even if the PTPN11 variant did not occur de novo in the individual currently being assessed. Or, identification of an MYH7 variant in an individual with hypertrophic cardiomyopathy could be combined with independent observations of this variant in individuals with hypertrophic cardiomyopathy from another laboratory to determine the applicable PS4 strength level. Additionally, a single proband could be used as evidence for multiple criteria, if those criteria are not overlapping in scope. For example, evidence of an affected proband with a de novo PTEN variant could be assessed with regards to de novo criteria (PS2/PM6) and prevalence in affected individuals (PS4). If there are multiple observations of the variant occuring in trans with a pathogenic variant in affected probands, only PM3 and not PS4 should be applied to those observations; however, if there was an independent statistically significant case-control study PS4 may be applied in addition to the PM3 observational cases.

COMBINING RULES SPECIFICATIONS

Bayesian Approach to Combining Rules

In addition to refining the relative strength of evidence types, another advantage of a quantitative approach to the ACMG/AMP guidelines is refinements to the combining rules. Following recommendations from the ACMG/AMP guidelines that “likely pathogenic” and “likely benign” be used to mean greater than 90% certainty of a variant either being disease-causing or benign, respectively, the Bayesian framework used the following posterior probability (Post_P) ranges for classification (Table 5): Pathogenic Post_Ps >0.99; Likely pathogenic Post_P between 0.90 and 0.99; Likely benign Post_P between 0.001 and 0.10; Benign Post_P <0.001. In modeling the ACMG/AMP guidelines into a Bayesian framework, each of the 18 combining rules outlined in the ACMG/AMP guidelines were evaluated for consistency by determining if the Post_P for each scenario is within the expected Post_P range for that classification. For example, from Richards et al (2015) Table 5, combing criteria for Likely pathogenic (iv) - ≥3 Moderate should result in a Post_P in the Likely pathogenic range (between 0.90 and 0.99). Of the 18 combining rules, only two combinations were mathematically inconsistent with the overall framework. The first inconsistent combing rule, Pathogenic (ii) - ≥2 Strong, resulted in a Post_P of 0.975, which was in the Likely pathogenic range (0.90 – 0.99), not the Pathogenic range (>0.99). The second inconsistent combining rule, Likely pathogenic (i) - 1 Very Strong AND 1 Moderate, resulted in a Post_P of 0.994, which was in the Pathogenic range (>0.99). The Likely pathogenic (i) - 1 Very Strong AND 1 Moderate combining rule is most commonly applied for novel LoF variants (PVS1 and PM1). While this combination may warrant a Pathogenic classification for certain variants, considerations such as confidence in LoF impact based on variant type and location need to be addressed and included in criteria evaluations before this modification can be adopted.

Table 5.

Posterior probability ranges per ACMG/AMP classification term

Post_P Range	Classification
Post_P > 0.99	Pathogenic
0.90 < Post_P ≤ 0.99	Likely pathogenic
0.10 ≤ Post_P ≤ 0.90	Uncertain significance
0.001 ≤ Post_P < 0.10	Likely benign
Post_P < 0.001	Benign

Novel Combining Rules

The Bayesian framework approach to the ACMG/AMP guidelines has also facilitated the development of combining rules not explicitly stated in the guidelines, as Tavitigian et al provided a simple spreadsheet calculator that allows users to calculate a Post_P for any set of criteria (Tavtigian et al., 2018, supplemental table). For example, the ACMG/AMP guidelines included two combining rules for a Likely benign classification, (i) 1 Strong and 1 Supporting and (ii) ≥2 Supporting, which result in Post_P of 0.003 and 0.025, respectively, both of which are within the Likely benign Post_P range (0.001 – 0.10). However, while not stated in the guidelines, one Benign Strong piece of evidence on its own results in Post_P of 0.006, which is within the Likely benign Post_P range. This scenario is most commonly applicable for variants that meet BS1 allele frequency threshold with no conflicting pathogenic data. Currently, three of the six approved VCEPs (Hearing Loss, RASopathy, Cardiomyopathy) have included in their specifications to the ACMG/AMP guidelines a note that in addition to specifying criteria usage, these VCEPs have also specified the combining rules to allow BS1 to reach a likely benign classification, provided there is no conflicting evidence. Additionally, the Hearing Loss VCEP proposed that a combination of one Very Strong criterion and one Supporting criterion reach a variant classification of Likely pathogenic. This combination, which was not included in the ACMG/AMP guidelines combining rules, was considered given that the Hearing Loss VCEP adopted the stringent criteria for evaluating predicted LoF variants proposed by SVI, such that rare LoF variants meeting at least PM2_Supporting have at least a 90% chance of being pathogenic. This additional combining rule was approved by SVI as one Very Strong and one Supporting results in Post_P of 0.988, which falls within the Likely pathogenic range (0.90–0.99).

Classifying Variants with Conflicting Data

With the ACMG/AMP guidelines (Richards et al., 2015), a VUS classification can result from either insufficient evidence or due to contradictory benign and pathogenic criteria. The guidelines did not provide guidance about classifying variants when the strength of conflicting evidence is not balanced. For example, if a variant occurred de novo, with paternity and maternity confirmed (PS2), a well-established functional assay suggests a deleterious impact (PS3), and the variant was absent from population databases (PM2), the variant would reach a Pathogenic classification. If for this same variant, multiple lines of computational evidence suggested this variant had no impact and criterion BP4 was also applied, then pathogenic and benign criteria are contradictory and the classification would now be uncertain, according to strict adherence to the ACMG/AMP Table 5 combining rules. However, a quantitative approach to the guidelines would help identify instances in which both pathogenic and benign criteria are applied but the evidence is predominantly on one side versus the other. For example, for the variant outlined above, if only the pathogenic criteria (PS2, PS3, PM2) were applied, the Post_P is 0.994, which is within the Pathogenic range (>0.99). If benign criterion BP4 were also applied (PS2, PS3, PM2 plus BP4), then Post_P lowers to 0.988, which is now in the Likely pathogenic range (0.90–0.99).

One limitation to this approach is the assumption that all criteria at a given strength level are in fact mathematically equivalent evidence. Richards et al was constrained by the need to have categories or tiers of evidence and precision may have been sacrificed. Determining which evidence types at a given strength level constitutes contradictory evidence may be subjective. For example, the ACMG/AMP guidelines state that when supporting by splicing predictions (BP7) and computation evidence (BP4), “one can classify novel synonymous variants as likely benign” (Richards et al., 2015). This scenario is supported by the Bayesian approach as two benign supporting criteria results in Post_P of 0.025, which is in the likely benign range (0.001–0.10). However, if this synonymous variant is novel, then PM2 is also applicable, which would result in contradictory benign and pathogenic criteria and an uncertain classification (Post_P of 0.10; BP4, BP7, plus PM2). This specific issue could be mitigated by removing or decreasing the weight of PM2, as previously discussed.

FUTURE DIRECTIONS

The Richards et al (2015) ACMG/AMP guidelines established a classification system for sequence variants by specifying evidence types, each with a suggested measure of strength, and providing rules for combining those evidence types to arrive at a classification. This framework was intentionally broad to be applicable for all Mendelian disorders; however, this broad scope created a degree of ambiguity when applying the guideline for variants within a specific gene or disorder. Since publication of the 2015 guidelines, both general use and disease-focused specifications have emerged to aid in accurate application of ACMG/AMP evidence types. The next iteration of variant interpretation guidelines, a joint recommendation from ACMG, AMP and ClinGen, has been tasked with revising the 2015 guidelines to address ambiguities in criteria, re-evaluate the appropriateness and strength of criteria, consider inclusion of gene-specific modifications from ClinGen VCEPs. To accomplish these charges, formal adoption of a quantitative, point-based framework will likely be required. Interpretation of the Richards et al (2015) ACMG/AMP guidelines in a quantitative Bayesian framework has already shown that the existing classification system is fundamentally sound (Tavtigian et al., 2018), indicating adoption of a quantitative approach will not require major disruption of the basic framework already established. Additionally, the pending ACMG/ClinGen recommendation for interpretation of constitutional copy number variants has followed a quantitative approach by specifying numerical values for each evidence type and for each classification term. Given the number of clinical laboratories that perform both sequence and copy number variant testing and classification, harmonization of frameworks would benefit the genetics community. In summary, refinement and specification of the 2015 ACMG/AMP guidelines will help the community move toward more consistent variant classifications, which will improve the care of patients with, or at risk for, genetic disorders.