Creating an atlas of variant effects to resolve variants of uncertain significance and guide cardiovascular medicine

Abstract

Cardiovascular diseases are leading global causes of death and disability, often presenting as interrelated phenotypes of atherosclerotic vascular disease, heart failure, and arrhythmias. They arise from interactions between environmental factors and predisposing genotypes and include relatively common Mendelian lipid disorders, cardiomyopathies, and arrhythmia syndromes. The identification of a pathogenic variant through genetic testing can inform disease diagnosis, risk prediction, treatment, and family screening. However, a major roadblock in genomic medicine is that many variants—especially missense variants—lack sufficient evidence to enable a definitive classification and are therefore deemed “variants of uncertain significance” (VUS). In this Review, we describe how multiplexed assays of variant effect (MAVEs) can enable functional assessment of nearly all coding variants in a target sequence, potentially offering evidence proactively for variants observed later in a patient. We discuss validation, including the role of in silico variant effect predictors (VEPs), and how multiplexed experimental methods are illuminating cardiovascular disease biology and ultimately resolving the VUS problem at scale.

Introduction

Major cardiovascular disease phenotypes – ischemia, heart failure, and arrhythmias – are driven by a complex mixture of common polygenic variation, rare large-effect variants, and environmental factors^1–6. Familial hypercholesterolemia (FH) affects nearly 1% of Americans⁷. Channelopathies such as long QT syndrome (LQTS) and Brugada Syndrome (BrS) can cause sudden cardiac death (SCD) and predispose individuals to drug-induced arrhythmias^8,9–12. Hypertrophic cardiomyopathy (HCM), has been revealed by sensitive imaging to affect 1 in 200 people¹³. Inherited dilated cardiomyopathy (DCM) is now recognized as the predominant cause of “nonischemic cardiomyopathy” (~1/250)^14,15, and many of the underlying genes also contribute to atrial fibrillation susceptibility^16,17. In a recent survey, 32% of 709 patients in a general cardiology clinic had actionable genetic testing results¹⁸. Accurately diagnosing cardiovascular genetic diseases has immediate implications for prevention, treatment, and cost of care. For example, FH is often underdiagnosed⁷, but early genetic diagnosis enables therapy that prevents stroke and early myocardial infarction¹⁹. Early diagnosis of heritable arrhythmias and cardiomyopathies can reduce catastrophic outcomes like SCD^20–22.

Accessible sequencing and the power of genetic diagnosis are propelling a new era of genomic medicine^23,24. However, one factor constraining implementation is a limited ability to determine the pathogenicity of most newly discovered variants, which are therefore frequently classified as variants of uncertain significance (VUS). In patients with possible Mendelian cardiovascular conditions, each VUS annotation represents a failure to illuminate the genetic underpinning of disease. VUS classifications can prevent the identification of at-risk individuals, hinder family screening, or lead to unnecessary clinical surveillance in low-risk patients. As more gene-specific risks and therapies are identified for cardiovascular diseases, VUS also impede the application of potentially life-saving, guideline-driven precision care, such as early SCD prevention in LMNA cardiomyopathy or early treatment of familial hypercholesterolemia with cholesterol-lowering medication^22,25,26.

The scope of the variant interpretation problem is vast. One study found that only 2% of the over 4 million missense variants in the Genome Aggregation Database (gnomAD)²⁷ had a clinical interpretation in the NIH database ClinVar. Furthermore, the most common classification for variants with an interpretation was VUS^28,29. In the eMERGE-Seq study³⁰, sequencing of 27 cardiovascular disease genes in nearly 22,000 participants identified 6,840 unique coding variants, of which 6,095 (89%) either lacked a definitive classification in ClinVar, or were absent from the database entirely. The number of variants in ClinVar has increased significantly due to the growing use and reporting of genetic tests in clinical and population screening settings. As of January 2025, there were 35,395 missense variants in ClinVar across 50 ClinGen-validated Mendelian cardiac disease genes (Figure 1). Of these variants, 79% are VUS, 12% have conflicting interpretations, and only 5% and 3% have P/LP or B/LB classifications, respectively. Figure 1C shows missense variants in three key cardiovascular disease genes, KCNQ1, LDLR, and MYH7, in the UK Biobank and All of Us cohorts. For all three genes, over 75% of all detected variants are VUS or have conflicting annotations in ClinVar, or are absent from ClinVar. Furthermore, given the de novo mutation rate, the size of the human genome, and the global population size, nearly every missense variant compatible with life is already present in dozens of individuals in the living human population^31,32. Currently, individuals with non-European ancestry have a higher rate of VUS in cardiac genes than individuals with European-ancestry³³, emphasizing the need for larger non-European datasets which will enable more accurate and equitable variant classification. Thus, there is a critical need for systematic evaluation of the pathogenicity of even the rarest variants. This need was highlighted by the 2020 NIH National Human Genome Research Institute Strategic Plan, which set a goal of making the term “VUS” obsolete by 2030²⁴.

Figure 1: The Variant of Uncertain Significance problem in cardiovascular genes.

A) Missense variants in ClinVar over time in 50 genes associated with Mendelian cardiovascular disorders. B) Barplot by disease class for the same set of genes shown in panel A. Truncating variants in titin (TTN) are associated with DCM and early-onset atrial fibrillation, but due to TTN’s large size and the unclear impact of TTN missense variants, TTN is listed as a separate category and not included in panel A. ClinVar variants in panel B were categorized into disease classes based on ClinGen-validated gene-disease associations, rather than based on phenotype assertions for individual variants provided in ClinVar. Additionally, in panel B, some genes are present in multiple disease classes. C) Number of missense variants in 3 key cardiovascular genes in the UK Biobank and All of Us. Variants are colored by their Clinvar classification. VUS = variant of uncertain significance, P/LP = pathogenic/likely pathogenic, B/LB = benign/likely benign, Arrhyth. = Arrhythmia, AC = arrhythmogenic cardiomyopathy, DCM = dilated cardiomyopathy, FH = familial hypertension, HCM = hypertrophic cardiomyopathy.

Multiplexed Assays of Variant Effect (MAVEs) are in vitro functional assays that can simultaneously assess the function of thousands of variants in a gene or genomic region of interest. In this Review, we summarize the growing use of MAVEs to help solve the VUS problem for cardiovascular genes. We also discuss variant effect predictors (VEPs), and how they can be integrated with MAVEs to help inform cardiovascular genomic medicine.

MAVEs: functional genomics on a massive scale

Overview of MAVEs

Functional assessments of genetic variants have traditionally focused on a few variants at a time, or occasionally dozens of variants using non-multiplexed assays like automated patch clamping^34–36. As more variants are discovered, the proportion of observed variants with high-quality functional data is rapidly decreasing²⁹. Multiplexed Assays of Variant Effect (MAVEs), which can investigate the function of all possible single nucleotide variants in a target genomic region³⁷, can address issues of both standardization and scale (Figure 2). The MAVE process (outlined in Figure 3 and discussed below) involves saturation mutagenesis of a target gene or region; expression of pools of mutants in an appropriate cell model; en masse multiplexed selection for cellular phenotypes that reflect variant effects on relevant protein function(s) (Figure 4); next-generation sequencing to compare variant abundance in pre- and post-selection pools; and data analysis and display (Figure 5)^29,32,37,38. In the following paragraphs, we summarize key aspects of MAVEs (e.g., mutagenesis approach, cell or model organism type, library delivery method, and assay type). Examples of published published and potential cardiovascular MAVE experiments are presented in Table 1.

Figure 2. Overview of lead candidate genes (left) and assay methods (right) for cardiovascular Multiplexed Assays of Variant Effects (MAVEs).

As described further in the text and Figures 3–5, some assay methods can be implemented across multiple genes and disease domains, while in other cases, gene-specific (“bespoke”) assays are needed.

Figure 3: Multiplexed Assay of Variant Effect approach.

Pools of cells each expressing a single variant are generated using gene editing (top left) or saturation mutagenesis in plasmids (bottom left) which are then integrated into cells for study. The cell pool is then challenged as described in the text (drug challenge, assay of cell surface expression, etc), variant abundance is assayed pre- and post-challenge (e.g., Figure 4), and the data are displayed as a variant effect map (e.g., Figure 5).

Figure 4: Examples of MAVE assays.

A. Cell surface expression assay. A pool of cells is sorted as a function of cell surface expression of the protein of interest. The abundance of each variant across bins of expression is assayed to establish the cell surface expression phenotype. B. Functional assay. In this example, the pool of cells is grown under conditions favoring survival of loss of function variants, as described further in the text. Comparing survival at baseline and after a period of growth is then used as the functional readout.

Figure 5: Variant effect map for KCNE1, a 129 amino acid potassium channel subunit.

A. ClinVar annotations. The amino acid position is shown on the x-axis and the amino acid substitutions on the y-axis. The dots indicate the wild-type amino acid at each position. Pathogenic or likely pathogenic (P/LP) variants are shown in orange, benign or likely benign (B/LB) variants in green and variants of uncertain significance (VUS) in gray. B. A map of function assayed by survival of cells coexpressing the gain of function KCNQ1 variant S140G (details in the text). Loss of function variants are shown in shades of red, gain of function variants in blue, and no change in white; no data were generated for cells shown in gray. Panel B reproduced from⁵².

Table 1:

Example published and possible MAVE studies of cardiovascular genes

Gene	Mutagenesis approach	Cell type	Library delivery	Assay type	Reference
Published cardiovascular MAVE studies
CALM1/2/3	POPCode	Yeast	Complementation	Cell survival	38
SCN5A (12 aa)	Inverse PCR	HEK293	Landing pad	Cell survival	71
MTHFR	POPCode	Yeast	Complementation	Cell survival	111
CYP2C9	Inverse PCR	Yeast, HEK293	Transformation, landing pad	Click-seq activity probe, protein abundance	119
KCNJ2 (mouse)	SPINE mutagenesis	HEK293	Landing pad	Surface abundance, membrane potential reporter	62
HNF1A	Commercial synthesis	HUH7	Lentivirus	Downstream marker expression	118
KCNH2	Inverse PCR	HEK293	Landing pad	Surface abundance	65
KCNE1	Inverse PCR	HEK293	Landing pad	Surface abundance, cell survival	52
MYH7 (5 aa)	CRaTER	iPSC-CM	Integration at endogenous locus	Protein abundance, cell survival	58
Potential cardiovascular MAVE studies
Lipoprotein metabolism (e.g. LDLR, PCSK9, LPL, APOB)	Exogenous library synthesis	HEK293 or Hela	Landing pad	LDL metabolism, triglyceride hydrolysis, surface abundance	—
Ion channel (e.g. KCNQ1, RYR2, CACNA1C)	Exogenous library synthesis	HEK293 or Hela	Landing pad	Surface abundance, cell survival, intracellular calcium reporter	—
Cardiomyopathy (e.g. MYBPC3, PKP2, RBM20)	Endogenous locus editing or integration	iPSC-CM	Endogenous locus	BNP reporter, cell survival, cell size, cell morphology	—

Saturation mutagenesis and generation of cell pools

A key step in a MAVE experiment is to generate a large pool of cells, each expressing a single variant in the target genomic region, on which the selection assay will be deployed. Experimental methods for library generation and integration into a model system have been discussed extensively in previous reviews^37,39 and are described in brief below.

The generation of mutant libraries can be accomplished using genome editing at the endogenous locus, or by exogenous delivery and expression of a variant library (Figure 3)^40,41. Libraries can be synthesized with a variety of methods including error-prone PCR, oligonucleotide-targeted editing, or commercial synthesis³⁷. For exogenous delivery, the mutagenized library can be cloned en masse into an appropriate expression vector and integrated at an engineered ‘landing pad’ locus to ensure that a single variant is expressed per cell^42,43,44. Alternatively, CRISPR-based gene editing technologies have introduced a range of options for mutagenizing the genome at the endogenous locus^40,45,46,47. These include saturation genome editing based on replacement of an endogenous segment with a mutagenized donor template; direct base editing (e.g., adenine and cytosine base editors directed using catalytically inactive Cas9 or “dCas9”); or more recently using prime editing, using CRISPR-directed introduction of a mutagenized oligonucleotide at the endogenous locus^45,48. One method, CRISPR-X, achieves relatively wide base-editing windows by combining dCas9 and sgRNA with activation-induced cytidine deaminase-mediated somatic hypermutation⁴⁰. These technologies are especially useful in cell types that are not tractable for genome engineering with large constructs, or for assessing variant function in cellular and disease contexts such as the effect of HCM variants on indices of contractility in cardiomyocytes. One advantage of endogenous DNA-editing methods over cDNA delivery approaches is the ability to measure the functional effects of non-coding variants, which are not present in cDNA constructs.

Functional assays for cardiovascular disease–associated genes

Assays have been developed to investigate disease-relevant phenotypes such as cell surface trafficking, response to a selection intervention (such as a drug challenge), or cellular phenotypes such as contractility or cell size. Some assay methods are broadly applicable to multiple genes, while others are gene-specific “bespoke” assays, optimized to measure pathophysiologic effects unique to variants in a particular gene. Typically, assays are developed based on known pathophysiology and are validated using conventional low-throughput measurements of control variants. Central to assay development is the choice of cell system. In some instances, protein function is examined in uniquely disease-relevant cell types, such as cardiomyocytes derived from induced pluripotent stem cells (iPSC-CMs). On the other hand, human protein variant function can often instead be accurately assessed in tractable heterologous cell models, such as human embryonic kidney (HEK) cells^51,52 or even yeast^38,53.

Multiple MAVEs may be required to provide a comprehensive functional assessment, using assays that either measure overall protein function or specific subfunctions. Here, we present examples with a focus on Mendelian cardiovascular disease genes.

Protein abundance:

A major class of loss-of-function variants are variants that disrupt protein stability or subcellular localization^54–57. One method, Variant Abundance by Massively Parallel Sequencing (VAMP-seq)⁵¹, involves the creation of a library of variants in a protein fused to a fluorescent protein such as GFP. Variants that disrupt protein abundance result in lower fusion protein abundance and lower fluorescence; pools of variant-expressing cells are sorted using Fluorescence Activated Cell Sorting (FACS) and deep-sequenced⁵¹. This method can be applied to a broad set of cardiovascular proteins where such variants are a major disease mechanism. For example, a multiplexed study of 113 variants in the HCM-associated gene MYH7 used VAMP-seq to identify pathogenic variants that disrupted β-myosin heavy chain protein abundance.⁵⁸ This assay accurately classified 3 benign and 3 pathogenic control variants. Another method, the abundance protein fragment complementation assay (aPCA), uses a fragment of an essential enzyme fused to the protein of interest⁵⁹. This technique translates differences in protein stability into changes in growth rate in yeast cells; pools of mutant-expressing cells can thus be assessed by deep sequencing. aPCA has been applied on a massive scale to assess the stability of thousands of variants in 500 protein domains⁶⁰. Although protein abundance assays can be readily applied to many genes to identify function-disrupting variants, a limitation of this assay type is that most genes likely have function-disrupting variants that do not affect protein abundance.

Cell surface abundance:

Many cardiovascular disease genes are cell surface proteins, including ion channels associated with arrhythmias and receptors and other proteins that regulate lipid processing. A decrease in cell surface abundance, typically due to misfolding and/or mistrafficking, is a common loss-of-function mechanism for cell surface proteins⁶¹. Cell surface proteins can be labeled using a fluorescently-labeled antibody against the protein or or an epitope tag (e.g. HA or myc tag; Figure 4A) on live cells. This cell surface abundance assay is applicable to a large set of cardiovascular genes, including key hyperlipidemia and arrhythmia genes such as LDLR, LPL, SCN5A, KCNE1, KCNQ1, KCNH2, and KCNJ2. A comprehensive scan of cell-surface abundance of KCNJ2 missense variants identified regions especially sensitive to mutation, including regions in the transmembrane helices and ER and Golgi export motifs⁶². A cell surface abundance assay of all possible KCNE1 variants identified known loss of trafficking variants associated with LQTS and predicted new loss-of-trafficking variants⁵². This study also identified glycosylation and N-terminal cysteine mutations that affect cell surface trafficking, and surprisingly revealed that large C-terminal truncations do not disrupt trafficking to the cell surface. Nearly 50% of possible variants in a KCNH2 hotspot region disrupted trafficking and MAVE trafficking scores were strongly correlated with automated patch clamp measurements of protein function^63,64. Extension of this trafficking assay to over 18,000 variants across KCNH2 revealed that the MAVE data, in combination with other variant features, predicted severe cardiac events⁶⁵. As with protein abundance assays, a limitation of cell surface trafficking assays is that cell-surface genes likely have function-disrupting variants that do not affect trafficking.

Ion channel function can also be assessed by MAVE-compatible voltage-sensitive fluorescent markers. For instance, overexpression of the cardiac potassium channel gene KCNJ2 in HEK293 cells leads to a decrease in the resting membrane potential. A membrane potential-sensitive fluorescent reporter was used in a MAVE to comprehensively measure the function of KCNJ2 variants⁶⁰. This dataset was compared to the cell-surface trafficking MAVE described above to identify variants that specifically altered channel function without affecting trafficking. For example, the lipid PIP₂ activates KCNJ2, and the “function-only” variants included residues involved in binding to and responding to PIP₂.

Cellular growth or survival:

Functional variant assays based on cellular growth have been the basis of many MAVE studies (Figure 4B). For example, Findlay and colleagues⁴⁷ exploited the observation that functional BRCA1 is required for survival of HAP1 cells, a haploid cell line. After saturation genome editing of BRCA1 exons encoding critical functional domains, high-throughput sequencing of variant pools was used to identify loss of function and normal function variants. The assay nearly perfectly validated 375 known pathogenic or benign variants and discovered ~700 new loss of function variants out of 3,893 missense variants tested. This approach is currently being deployed for additional genes that are essential in the HAP1 cell line^66–68. A map for the three genes CALM1, CALM2, and CALM3, which encode the identical calmodulin protein product, was generated based on cellular growth and survival in a yeast selection assay³⁸. This map has already been used to interpret clinical variants in 3 patients presenting with variants in CALM1/2/3 and arrhythmias (see below for additional details)⁵³. In other work, a multiplexed study of 113 MYH7 variants in iPSC-CMs identified a cell survival phenotype that could accurately classify 8 control benign and pathogenic variants⁵⁸.

Many Mendelian cardiovascular genes are not essential in commonly-used tractable cell lines, limiting applicability of this relatively straightforward cellular survival MAVE approach. However, it is often possible to impose environmental or genetic contexts that sensitize cells to the function of the gene of interest. For example, when the fatty acid synthase gene (FASN) is knocked out in HAP1 cells, the loss of LDLR (which encodes the low-density lipoprotein (LDL) receptor) causes a more pronounced growth impact⁶⁹. Other examples involve SCN5A and KCNE1, ion channel genes linked to arrhythmias. A scan of a 12 residue voltage sensor region in the sodium channel SCN5A was performed using a triple drug selection assay in which veratridine and brevetoxin inhibited normal channel inactivation, and the sodium-potassium ATPase inhibitor ouabain sensitized cells expressing normal or gain of function channels to sodium overload and cell death^70,71. Similarly, a KCNE1 map was generated in HEK293 cells stably expressing the KCNQ1 variant S140G, which exhibits gain of function properties when coexpressed with KCNE1 (Figure 4B)^72,73. Cells expressing normal function KCNE1 variants were depleted from the cell pool, whereas cells expressing KCNE1 loss of function variants remained in the cell pool.52 This assay generated functional data for 2,534 variants, of which only 111 of which are annotated in ClinVar, most of which are VUS (Figure 5).

Cardiomyopathy phenotypes:

Many pathogenic HCM- or DCM-causing variants cause specific contractile phenotypes that are difficult to assess in non-cardiomyocyte cells, and iPSC-CMs are a promising system in these cases. Direct measurements of hypercontractility or hypocontractility — primary cellular correlates of HCM and DCM — are possible in single iPSC-CMs⁷⁴, 2D monolayer, or engineered 3D tissue of iPSC-CMs⁷⁵, but these measurements are difficult to scale to multiplexed assays. However, cardiomyopathy is also associated with other phenotypic changes, including altered cellular size, stiffness, sarcomeric structure, and expression of markers such as BNP. Assays to detect these phenotypes may be useful for genes harboring variants that cause sarcomeric hyperactivation (MYH7, MYBPC3, TNNT2 and TPM1)^76,77 or aberrant sarcomeric and cytoskeletal stretch-sensing (TTN, LMNA, DSP)^78–80. For example, a non-multiplexed study of 51 TNNT2 variants identified changes in contractility in iPSC-CMs expressing pathogenic variants⁷⁴. The same study also distinguished variants using a fluorescent reporter for expression of NPPB (encoding BNP), an approach that is compatible with MAVE experiments. Passive size-based cell sorting could also provide a direct measure of single iPSC-CM hypertrophy⁸¹. High speed microscopy and advanced image analysis using deep learning can define static and dynamic cell morphology characteristics⁸² and may enable active sorting of cells based on multi-feature morphology assessment^49,58. Pathogenic variants in cardiomyopathy genes might also cause changes in protein abundance or cell survival in iPSC-CMs that are compatible with the corresponding scalable assays described above⁵⁸.

Uptake and secretion:

The ability of cultured cells to effectively take up LDL and other lipoprotein particles that have been labeled with a traceable fluorophore facilitates a range of scalable functional assays for dyslipidemia genes⁸³. For instance, cells expressing a disrupted LDLR gene exhibit decreased internalization of fluorescently labeled LDL⁸⁴. This approach could be expanded to other genes, such as PCSK9 and LDLRAP1, which are involved in LDL uptake, as well as other lipoprotein subtypes and receptors (e.g., very-low-density lipoprotein and VLDLR).

The general MAVE approach (Figure 3) relies on culturing a large number of cells that each express a different variant. This is not ideal for secreted proteins like PCSK9 because, in a heterogeneous culture, the impact of a secreted protein variant becomes decoupled from the genotype of its cell of origin. To address this issue, the secreted protein can be tethered to the surface of the secreting cell, for example, using a glycosylphosphatidylinositol anchor⁸⁵. Alternatively, secreted proteins can be bound to their cell of origin by expressing them as a fusion with a selected transmembrane domain. This strategy was applied to coagulation factor IX, a secreted serine protease linked with hemophilia B⁸⁶.

Gene expression and splicing assays:

Most MAVEs and clinical classification efforts to date have focused on protein-altering coding variants. Non-coding regions are often more difficult to interpret, due to their large size and the complexity of their effects on regulation of gene expression or function. However, non-coding regions are increasingly recognized for their contribution to disease, both through common risk loci detected through genome-wide association studies and rare, large-effect variants. Regulation of gene expression and aberrant splicing can also be examined using variant effect mapping.

Enhancers can be studied with Massively Parallel Reporter Assays (MPRAs). MPRAs involve the creation of a library of candidate enhancer or promoter regions, followed by multiplexed expression in cells and high-throughput sequencing. MPRAs can be applied to inform disease processes for a variety of common and rare cardiovascular disorders⁸⁷. MPRAs of candidate cardiovascular enhancers can be performed in cell lines or in vivo, e.g. by transduction of mouse hearts^88,89. The LDLR promoter has been comprehensively scanned, identifying loss of function regulatory variants that may be linked to FH⁹⁰. A recent MPRA study of over 680,000 candidate enhancer sequences included saturation mutagenesis of an enhancer of PKLR, a gene linked to pyruvate kinase deficiency and hemolytic anemia⁹¹.

To address the cis effects of variants in splicing regions of disease genes, previous studies have identified coding variants that lead to reduced transcript expression by RNA sequencing⁴⁷. In some cases, nonsense-mediated decay-affected variants can be identified using genotypes and expression levels derived by RNAseq⁹². Exploration of the cis splicing effects of intronic or exonic variants in genes can be performed with pooled minigene splice reporter assays⁹³. In addition, single cell RNA sequencing can be used to quantify the splicing of specific transcripts⁹⁴, although existing methods for simultaneous DNA and RNA sequencing in single cells remain too resource-intensive for broad use in variant effect mapping⁹⁵, These latter methods may be more fruitfully deployed for examination of pathogenic missense variation in trans splicing regulators, such as RBM20 and RBFOX2 in cardiomyocytes⁹⁶.

Other methods to probe variant-to-function relationships in regulatory regions are emerging. A single cell imaging and gene expression platform (LipocyteProfiler) was applied to variants in the 2p23.3 locus (previously linked to insulin resistance⁹⁷) and identified mitochondrial function phenotypes during early differentiation which progressed to altered lipid-droplet formation in mature visceral adipocytes⁹⁸. More generally, methods that can evaluate variants within risk loci will continue to be important for predictive cardiovascular medicine.

Data analysis and display

In MAVE experiments, the abundance of each variant is quantified by high-throughput sequencing in the pre-selection pool, either in the delivered library or in the cell pool. After phenotypic selection using one of the assay approaches described above, the abundance of each variant is again quantified by high-throughput sequencing and compared with the pre-selection pools. This quantification can be accomplished either by direct sequencing or by identification of barcodes paired with known variants^38,64,71. Variant effect scores are then calculated using standardized and reproducible data analysis pipelines⁹⁹. These scores are then normalized to internal controls, e.g., rescaling scores so that synonymous variants have a median score of 1 and nonsense variants have a median score of 0. Larger numbers of replicates allow more accurate estimation of random error in scores, although even two replicates allows error estimation via trends that relate pre-selection variant abundance to replicate agreement^38,99,100. Scores (and estimated errors) can then be visualized as a “variant effect map”, as illustrated in Figure 5.

Validation and implementation of MAVEs

Using MAVEs in clinical variant classification

MAVE datasets are already being used in clinical settings to help interpret cardiovascular variants. A MAVE for the three identical calmodulin genes, CALM1, CALM2, and CALM3, was performed using a yeast selection assay³⁸. This map was deployed to interpret clinical variants in three patients, each presenting with arrhythmias or sudden unexplained death and distinct variants in CALM1, CALM2, or CALM3⁵³. For two of the patients, the MAVE data supported the loss of function of the variants, providing rapid evidence for variant reclassification towards pathogenic. For the third patient, the MAVE data, combined with additional clinical and family information indicating a phenotype atypical for calmodulin deficiency, supported the conclusion that the CALM1 variant was not the major cause of the patient’s phenotype.

Catalogs of variant function will be increasingly useful for annotating clinical variants and reducing the burden of VUS. MAVEs generate a large amount of variant data using a single standardized assay, thus minimizing the between-lab variability associated with ‘small-batch’ approaches while also facilitating assessments of assay quality. Notably, some “medium-throughput” assays such as automated patch clamping can also generate large, standardized functional datasets¹⁰¹; correlation between MAVE data and these types of datasets can be an important validation step⁶⁵. In the initial ACMG/AMP guidelines, the PS3 criterion was applied if a well-validated functional assay showed a damaging effect or the BS3 criterion was applied if a well-validated functional assay showed no damaging effect¹⁰², with each of these criteria being applied at the “strong” level. The ClinGen Sequence Variant Interpretation (SVI) working group recommended a more nuanced approach, in which functional assays are validated with a large set of accepted pathogenic and benign controls, and this functional evidence can be applied at variable evidence strengths towards variant classification (Box 1)¹⁰³. One potential caveat of MAVEs is that no single assay interrogates all functions of a protein. Because of this, “edge case” variants may be discovered with borderline assertions or disagreements between MAVEs and other data sources such as patient datasets or computational predictors. These variants can be further assessed in other model systems such as zebrafish or medium-throughput assays such as automated patch clamping.

Box 1: Calibrating MAVE assays using benign and pathogenic controls.

Box 1: Calibrating MAVE Assays Using the ClinGen SVI Approach.

Example calibration of two MAVE assays. P=pathogenic, B=benign. The left image shows MAVE assay results for 30 pathogenic and 30 benign controls, with the first assay having better performance than the second assay. The middle panel shows two 2×2 matrices summarizing MAVE score results by variant classification. The right panel shows Log Likelihood Ratios for pathogenicity and benignity (LLR_P and LLR_B), and how those convert to evidence strengths in the ACMG classification scheme.

The ACMG/AMP guidelines originally categorized functional evidence from well-established assays as either PS3 (abnormal assay result) or BS3 (normal assay result), both at a strong evidence level⁹⁷. More recently, the ClinGen Sequence Variant Interpretation (SVI) group recommended calibrating functional assays using a large set of control variants known to be benign or pathogenic⁹⁸. This calibration allows the calculation of a likelihood ratio of pathogenicity (LLR_P, sometimes called “OddsPath”⁹⁸). Depending on LLR_P values, the strength of evidence for pathogenicity or benignity can range from supporting to strong or even very strong (for pathogenicity). MAVEs are well-suited for this approach because they generate data for many control variants, enabling implementation of the ClinGen SVI framework¹⁰⁰. For example, in the figure, two MAVE assays are calibrated: one achieves PS3_strong and BS3_strong, while the other achieves PS3_moderate and BS3_moderate. A limitation of the original “OddsPath” approach is the use of fixed score thresholds, which apply the same strength of evidence to all variants above a threshold. Newer methods allow finer calibration, assigning different levels of evidence strength to variants based on their LLR_P values, even within the same dataset^53,124.

For some genes, large sets of control variants are available in ClinVar. However, for many genes, the number of reference variants is limited—especially for benign variants. This is partly due to reporting bias and the difficulty of proving a variant has no disease impact. Control variants can be supplemented with additional pathogenic variants from disease-focused datasets like SHARE¹²⁵ or with variants designated benign because they are present at high frequency (sometimes in only some ancestries) in resources like gnomAD^27,98. Finally, misannotated variants in reference sets can affect calibration accuracy, often underestimating evidence strength. The calibration approach described here can also be adapted for other types of data, such as computational variant effect predictors¹¹⁷.

The performance of a MAVE dataset is often assessed against existing clinical variant classifications (e.g. by precision-recall analysis). A recent method evaluated the performance of computational variant effect predictors based on their ability to infer human phenotypes in two large population-based cohorts, UK Biobank and All of Us¹⁰⁴. Applying this approach to assess MAVE data could facilitate performance benchmarking across multiple MAVE datasets and could help interpret variants found in ancestrally diverse cohorts¹⁰⁵.

A long-term goal of MAVE science is to improve variant classifications and accelerate the use of genetics in clinical care. A recent analysis of MAVE datasets for germline cancer variants revealed that 15–69% of VUS could be reclassified with the inclusion of MAVE data, depending on the gene/dataset¹⁰⁶. A variant effect map for calmodulin was used to prospectively inform variant interpretation, diagnosis, and patient care⁵³. “Best practice” guidelines have been established for reporting MAVE data¹⁰⁷. These guidelines are designed to enhance the quality of MAVE datasets and broaden their clinical applications across populations of diverse ancestry^105,108,109.

MAVEs also offer significant potential for gene-disease pairs with limited or conflicting evidence by generating robust datasets to address critical gaps in variant interpretation. For genes with few known pathogenic variants, such as KCNE1 or JPH2, MAVEs can identify a broader set of function-disrupting variants to clarify their impact on disease risk (Figure 4)⁵². To maximize their utility, gene prioritization for MAVE studies should consider clinical relevance, evidence gaps, and assay feasibility.

The ACMG/AMP system is primarily designed to classify large-effect variants, but there is growing recognition of low-penetrance or ‘modifier’ variants that do not fit neatly within the 5-tiered framework¹¹⁰. MAVEs offer some promise in modeling these variants, as low-penetrance variants may have moderately dysfunctional molecular phenotypes that can be measured in MAVE experiments. Additionally, low-penetrance common variants can interact synergistically with large-effect variants, as exemplified by MTHFR, which is linked to an inherited disorder of folate metabolism. The common partial loss-of-function MTHFR variant p.Ala222Val interacts with a key environmental factor (dietary folate yielding folinate) to modulate the phenotype. Combinatorial MAVEs have been used to model these interactions, examining the effects of p.Ala222Val, other MTHFR variants, and folinate concentration¹¹¹.

Insight into underlying biology

High-throughput in vitro studies of variant function can not only enhance variant classification, but also serve as powerful tools to reveal disease biology. These datasets readily identify residues, motifs or domains that are highly intolerant to variation. Pairing this map information to experimentally-derived and computed protein structures^112–114 can provide insights into underlying mechanisms. Functional scores that incorporate disease-cohort and population variation are being used to identify novel functional domains in genes causing arrhythmogenic cardiomyopathy^80,115, to probe the functional consequences of missense variation in channelopathies^35,64, and to inform quantitative models of pathogenicity and penetrance^71,116,117. MAVEs can provide insight into unexpected mechanisms of protein function that lead to follow-up investigations. For example, the recent KCNE1 map highlighted the high number of loss of function variants at protein-protein (KCNE1-KCNQ1-calmodulin) interaction sites⁵². Variant effect maps can also reveal underlying variant mechanisms, i.e. to determine whether a variant causes loss or gain of function, or affects the stability or function of a protein. MAVEs can also reveal the disease implications of variants in genes not traditionally considered Mendelian cardiovascular disease genes. For example, loss of function variants in HNF1A are linked to monogenic diabetes. An analysis of HNF1A gain of function variants identified by a HNF1A MAVE in biobank cohorts found an unexpected association between these variants and the hepatic secretion of atherogenic lipoproteins¹¹⁸. MAVEs of the cytochrome P450 genes CYP2C9¹¹⁹ and CYP2C19¹²⁰ identified novel function-altering variants predicted to impact drug metabolism. These studies have implications for prescribing the anticoagulant/antiplatelet drugs warfarin and clopidogrel, which are affected by CYP2C9 and CYP2C19 genotypes, respectively.

The Atlas of Variant Effect Alliance has developed a controlled vocabulary for MAVE studies¹⁰⁷, implemented in the MaveDB repository¹⁰⁹. This vocabulary enables a standardized description of key MAVE features, including library generation methods, cellular systems, phenotypic assay types, and sequencing methods. As with all in vitro models, it is essential to critically evaluate whether MAVE assays accurately reflect the biological processes and mechanisms relevant to the disease being studied. MAVEs report the specific functional feature(s) being assayed in vitro across large numbers of variants. The extent to which variants that alter function in a MAVE generate clinical phenotypes, i.e., penetrance, is highly variable and dependent on multiple and incompletely understood factors. These include genetic background (currently often captured by polygenic scores) and environmental influences; in addition, the severity of protein dysfunction captured by a MAVE may influence disease penetrance. Public resources like ClinVar¹⁰⁸ and MaveDB¹⁰⁹ make MAVE data accessible in standardized formats, which will help clinical labs integrate this evidence into variant classification efficiently.

Variant Effect Predictors in clinical variant classification

Overview of Variant Effect Predictors

In the last decade, Variant Effect Predictors (VEPs) have begun to offer an increasingly reliable means of predicting variant effects. We note that our use of the term “VEP” does not refer to the Ensembl Variant Effect Predictor (VEP) annotation tool, but to the general class of variant computational prediction tools. These models make predictions based on patterns derived from input features (e.g., evolutionary, biochemical, or population-level evidence) and can be broadly categorized by whether they were developed using supervised or unsupervised machine learning methods (see Box 2). Because they can be scaled to provide exome-wide predictions, computational predictors can be readily applied to a large set of human genetic variants—comparatively, MAVE functional assessments have to date covered fewer than 1% of human genes³⁷.

Box 2: Variant effect predictors.

The most important input for essentially every well-performing VEP is sequence conservation. This feature has been traditionally derived from multiple sequence alignments, and indeed VEPs have generally performed best for regions that are well aligned across many species¹²⁶. Since the earliest examples^{115,127–129}, most VEPs have used a protein-specific multiple sequence alignment, with the frequency of each residue at each position informing the probability that a given substitution is damaging. Here, substitutions are more likely to disrupt function when they are at conserved positions and also involve amino acid exchanges that are less-commonly seen in aligned homologous sequences (as captured in substitution matrices such as BLOSUM62¹³⁰). Later methods integrated other protein information such as biochemical features, site annotations (e.g., binding sites), and (often-predicted) secondary and tertiary structure¹³¹. More recently, deep learning models have been employed to capture higher-order (between-position) dependencies between amino acid substitutions through the course of evolution¹²⁶. Most VEPs have been “supervised” in the sense that their models have been optimized to perform well for a training set of variants that have been labeled according to pathogenicity. One recent high-performing VEP, AlphaMissense, did not use pathogenicity labels, but instead optimized models using the absence of variants in a population-level cohort (gnomAD) as a proxy for pathogenicity. An important class of “unsupervised” methods (e.g., ESM1b¹¹³) exploits conservation information by ‘learning the language’ of protein sequences drawn from many species, without ever using pathogenicity annotations. Given that these VEPs do not require multiple sequence alignments, they have greater potential to estimate variant effects within intrinsically-disordered or other less well-aligned regions.

Some VEPs can predict the impact of noncoding variants, including both single-nucleotide variants and short insertions and deletions. Some such VEPs predict the effect of variants genome-wide^132,133 while others are optimized for sequence subtypes such as 5’-UTRs¹³⁴ or promoters¹³⁵ or to predict epigenetic effects¹³⁶. These VEPs can incorporate a broad range of features such as chromatin accessibility, transcript levels, epigenetic marks to infer regulatory variant effects, cross-species conservation, or sequence properties. Importantly, comprehensive ‘truth sets’ and independent functional assessments are less available for noncoding variants, complicating the assessment and interpretation of their performance.

Where they have been ‘overfit’ to training labels, supervised VEPs are generally more prone to inflated estimates of performance, e.g., where performance has been judged with variants not used in training the VEP. Use of allele frequency information either directly or indirectly (e.g. using knowledge of a variant’s appearance in a population cohort) can introduce some circularity, given the strong role of allele frequency in determining benign labels¹¹². While unsupervised methods can circumvent some of these issues (by not training on labelled data), some VEPs advertised as unsupervised have used labeled data for parameter optimization. Finally, performance of all VEPs may be inflated where related evidence (e.g., sequence conservation) has been used in assigning pathogenicity labels used in testing.

Predictors are increasingly incorporating the outputs of other prediction algorithms as features in training (typically in a supervised framework). These ‘meta-predictors’ now rank among the top-performing methods (e.g., VARITY¹¹⁴, REVEL¹³⁷)^99,112. However, given the breadth of data they incorporate in training (both directly and indirectly through the use of other VEPs), it remains challenging to obtain unbiased ground-truth sets by which to benchmark meta-predictors. Measuring the ability of VEPs to predict human phenotypes in independent cohorts offers an unbiased approach⁹⁹.

To assess VEP accuracy and inform predictor choice, the predictors must be benchmarked. This is generally done using a ‘gold standard’ reference set of annotated variants from available databases (e.g., ClinVar). Benchmarking of individual VEPs has been performed with deep mutational scan data^121,122, clinical classifications^121,122, biobank cohorts¹⁰⁴, and case-control and population cohorts¹²³. These benchmarking studies summarize individual VEPs, their main features, and their performances. Importantly, performance evaluations can depend on the benchmarking dataset¹²³ and their reliability is often limited by circularity and bias, which can have a profound effect on the perceived performance of a given method.¹²⁴ For instance, where data that were used to train a predictor are later re-used in its evaluation, performance estimates may be inflated^124,125. Moreover, where training data are skewed towards variants of a certain class (e.g., pathogenic or benign variants within proteins or protein families), a predictor may perform well against variants in this class, but poorly in classes where annotations are more balanced^124,125. Some predictors will have also incorporated historical bias given the rates at which variants have been annotated as pathogenic or benign for a given gene, even where these rates do not reflect the a priori probability that a given patient’s variant is pathogenic. In an effort to limit concerns over circularity and bias, Livesey and Marsh have used experimental functional measurements from MAVEs (which have not been seen in training) to benchmark predictor performance^122,124,126. Alternatively, measuring the correspondence between functional predictions and human traits in large population-based cohorts has enabled unbiased benchmarking¹⁰⁴, and this can be applied in cohorts of diverse ancestry and ethnicity.

VEPs are becoming increasingly accurate, with recently published methods (e.g., AlphaMissense¹¹⁴, ESM1b¹²⁷, and VARITY¹²⁸) outperforming commonly-used early predictors (e.g., PolyPhen2¹²⁹ and SIFT¹³⁰)^104,126. As with MAVEs, predictors should be calibrated against control variants to inform evidence strength prior to use in clinical variant classification (e.g., PP3 and BP4; Box 1)¹³¹. Recently, the ClinGen SVI estimated the score intervals corresponding to each evidence strength for commonly-used predictors¹³¹, as well as more recent methods¹³², finding top-performing predictors from independent benchmarking efforts (e.g., REVEL, MutPred2, AlphaMissense, ESM1b, and VARITY) to reach Strong evidence towards pathogenicity and benignity for stringent score windows. One limitation to this genome-wide calibration, however, is that it may mask disparate performance between genes¹³³. Ideally, as described for MAVEs, calibration should be carried out more accurately at the gene level. (Note: Initial reports of new VEPs often recommend thresholds for pathogenicity or benignity. However, in the absence of independent score calibration^131,132, self-reported thresholds and classifications should be taken with extreme caution).

Opportunities for synergy between variant effect predictors and MAVEs

In vitro and in silico methods are already complementary in that agreement among MAVEs, VEPs, and reference sets can increase mutual confidence in each approach, and MAVEs have become a useful means of benchmarking VEPs^121,122,126. Beyond this, functional data from MAVEs are increasingly used to refine and validate computational variant effect predictors, and functional evidence has even been incorporated into VEP training^128,134, although this approach can complicate the use of MAVEs to benchmark VEPs).

An important form of synergy between VEPs and MAVEs is that, under current variant classification guidelines, they are considered independent sources of evidence¹⁰², so that they are not “in competition”. However, as MAVEs and VEPs become increasingly accurate at predicting functional impacts of variation, their outputs are expected to become dependent¹²⁶. Indeed, some extensions of the ACMG framework allow use of either VEP or experimental functional evidence, but not both¹³⁵. How best to combine VEP and MAVE evidence together with other criteria (including variable clinical phenotypes) in variant interpretation remains an open and important research question. For example, for a subset of variants, VEP and MAVE evidence might disagree with each other or with clinical case or allele frequency evidence. A close examination of disagreements between these datasets should identify areas where VEPs or MAVEs are underperforming in modeling disease pathophysiology.

There are opportunities for MAVE strategies aimed that augment or increase the utility of VEP methods. For example, current predictors have been shown to underperform in predicting the effects of non-loss-of-function variants—particularly gain-of-function and dominant-negative mutations¹³⁶. Thus, functional assays could be targeted to regions where VEP performance is known or expected to be less reliable (e.g., positions under strong positive selection or intrinsically-disordered regions). A complete understanding of variant effects may require functional assessments in many relevant genetic or environmental contexts. Current VEPs do not consider context-dependence of variant effects (e.g., lipoprotein subtypes, co-expression of binding partners, or drug exposures), MAVEs are scalable across such diverse physiologic contexts and thus well-suited for those functional assessments. Thus, MAVEs and VEPs represent complementary approaches in classifying and understanding human variants and their biology at scale and a future vision of VUS reclassification will engage both.

Conclusion

The VUS problem poses a growing challenge to the implementation of genomic medicine for cardiovascular disorders. It is increasingly evident that a major part of the solution will be high-throughput functional genomic experiments such as MAVEs, which can provide functional evidence for all possible variants within a genetic element. Through MAVEs, multiple features of variants in cardiovascular proteins can be defined; examples include measurements of protein abundance, cell surface trafficking, lipoprotein uptake, or electrical signaling, or cellular contractile properties.

MAVEs can provide new insights into underlying biological functions of key cardiovascular proteins and accelerate personalized care for patients and their families. Because they can be used to study every possible variant in a gene or target region, MAVEs can partially address current inequities in clinical genetics by reducing the higher rate of VUS that exist in populations of non-European-ancestry^33,105. MAVEs are complementary to other emerging variant interpretation methods, such as VEPs and large patient cohorts. MAVEs, when integrated with these other datasets, will help improve variant annotations and make significant progress towards making the term “VUS” obsolete²⁴.

Key points.

• While genetic testing is increasingly used in clinical management of inherited cardiovascular disorders, most variants are classified as ‘variants of uncertain significance’ (VUS).

• Multiplexed assays of variant effect (MAVEs) can assess function of nearly all coding variants in a target sequence, providing proactive evidence for variants observed in patients.

• In silico variant effect predictors (VEPs) are becoming increasingly accurate and can provide predicted variant effects for nearly every variant in the genome.

• MAVEs and VEPs provide complementary information to illuminate cardiovascular disease biology and resolve the VUS problem.

Data availability

This study used data from the All of Us Research Program’s Controlled Tier Dataset version 8, available to authorized users on the Researcher Workbench via https://www.workbench.researchallofus.org. The UK Biobank dataset is available by application via https://www.ukbiobank.ac.uk/.qq