Estimating the Post-Test Probability of Long QT Syndrome Diagnosis for Rare KCNH2 Variants

Abstract

Background -

The proliferation of genetic profiling has revealed many associations between genetic variations and disease. However, large-scale phenotyping efforts in largely healthy populations, coupled with DNA sequencing, suggest variants currently annotated as pathogenic are more common in healthy populations than previously thought. In addition, novel and rare variants are frequently observed in genes associated with disease both in healthy individuals and those under suspicion of disease. This raises the question of whether these variants can be useful predictors of disease. To answer this question, we assessed the degree to which the presence of a variant in the cardiac potassium channel gene KCNH2 was diagnostically predictive for the autosomal dominant long QT syndrome.

Methods -

We estimated the probability of a long QT diagnosis given the presence of each KCNH2 variant using Bayesian methods that incorporated variant features such as changes in variant function, protein structure, and in silico predictions. We call this estimate the post-test probability of disease. Our method was applied to over 4,000 individuals heterozygous for 871 missense or in-frame insertion/deletion variants in KCNH2 and validated against a separate international cohort of 933 individuals heterozygous for 266 missense or in-frame insertion/deletion variants.

Results -

Our method was well-calibrated for the observed fraction of heterozygotes diagnosed with Long QT. Heuristically, we found that the innate diagnostic information one learns about a variant from three-dimensional variant location, in vitro functional data, and in silico predictors is equivalent to the diagnostic information one learns about that same variant by clinically phenotyping 10 heterozygotes. Most importantly, these data can be obtained in the absence of any clinical observations.

Conclusions -

We show how variant-specific features can inform a prior probability of disease for rare variants even in the absence of clinically-phenotyped heterozygotes.

Introduction

Sequencing an individual’s full genome or exome now costs less than many routine medical procedures. One resulting vision is that our genomes could be sequenced early in life for individualized medical advice about disease prevention and drug selection. However, most discovered variants will never be observed in a sufficient number of heterozygotes for a definitive association with disease.^{1, 2} Furthermore, even when a variant is strongly associated with disease, the clinical implications can vary strikingly across individuals.^{3, 4}

The American College of Medical Genetics and Genomics (ACMG) put forward an interpretation framework that integrates criteria such as variant prevalence, function, and computational predictions into a single annotation from “Benign” to “Pathogenic”.^{5, 6} Each additional satisfied criterion raises or lowers the probability the variant is classified “Pathogenic”.⁶ However, even definitively annotated “Pathogenic” variants have heterogeneous or asymptomatic clinical presentations^{7, 8} and variants annotated “Benign” may still increase risk (see Discussion). Thus, the predictive value of rare variants remains unclear.^{9, 10} Because a positive test for most variants cannot be applied to enough heterozygous individuals to achieve a definitive association with disease, a statistical approach is required to estimate the post-test probability of disease and validate those predictions in different groups and cohorts.

The observation of a single (or few) heterozygous carrier(s) does not adequately inform the probability of disease. Rather, individuals heterozygous for these variants benefit from a prior probability informed by knowledge about their clinical characteristics or the population they are drawn from. This prior, informed without knowledge of the variant, more reasonably reflects the true disease probability. In contrast, we propose to construct a prior probability of disease conditioned on variant-specific features and to modify this estimate using observations of heterozygous carriers. Our analysis yields a prior probability conditioned on variant-specific features known to be relevant to the association between the gene and disease. In practice, disease probability estimates are calibrated largely by how variant-features associate with disease probability in well-characterized variants. Our final estimate is effectively the post-test probability of disease given the presence of a variant or the positive predictive value (PPV) of rare genetic variants. We use “post-test probability” interchangeably with “penetrance”, in which the former recapitulates diagnostic thinking and reflects the Bayesian framework of this approach.

In past work, we described an algorithm for estimating the probability of a diagnosis of Brugada syndrome given the presence of a variant in the cardiac sodium channel gene SCN5A.¹¹ While we incorporated variant-specific covariates (e.g. sequence conservation, functional perturbation, structural location, etc.), Brugada syndrome is likely oligogenic and the clinical phenotype is sometimes difficult to assess. In this manuscript, we develop a similar algorithm for estimating the probability of long QT syndrome type 2 (LQT2), a well-characterized and monogenic disorder induced by variants in the cardiac potassium channel gene KCNH2.

The KCNH2 gene (also called the human Ether-à-go-go-Related Gene, or hERG) encodes an ion channel subunit that assembles into the homotetrameric K_V11.1 potassium channel. This channel produces the rapid delayed-rectifier repolarizing current, I_Kr, which sustains cardiomyocyte repolarization throughout the action potential plateau phase.^{12, 13} Loss-of-function KCNH2 variants that reduce I_Kr are associated with LQT2, a congenital heart arrhythmia defined by a prolongation of the QT interval on an electrocardiogram (ECG). Individuals with this ECG feature are at a greater risk for torsades de pointes, a life-threatening arrhythmia. With our method, we estimate the probability that an individual heterozygous for a missense or in-frame insertion/deletion (indel) variant in KCNH2 presents with LQT2 (for each variant). We validated our approach in an international cohort of 933 individuals ascertained under suspicion of LQTS, heterozygous for 266 unique missense and in-frame indel variants in KCNH2. Our results suggest the probability of disease can be estimated accurately before knowing the phenotype of a given heterozygous individual. Our result is a point estimate and 95% credible interval of disease probability for each variant which can be calculated before observing a single heterozygote. This prior—conditioned on variant-specific features—can be directly combined with observed heterozygotes for a posterior probability of disease. In this way, the prior is comparable with observations of heterozygotes, we estimate roughly 10 observations. All data resulting from this study are presented in web-accessible format at https://variantbrowser.org/KCNH2/ (Figure S1).

Methods

All data and materials are publicly available on GitHub (https://github.com/kroncke-lab/Bayes_KCNH2_LQT2_Penetrance). Additionally, a compiled and curated form of the data presented here are available in the KCNH2 Variant Browser (https://variantbrowser.org/KCNH2/; Figure S1). Internal Review Board (#191563) was evaluated at Vanderbilt University Medical Center and found to meet 45 CFR 46.104 (d) category (4) for Exempt Review. Detailed methods are available as supplemental data.

Results

KCNH2 variant heterozygote datasets

In total, the literature combined with gnomAD produced 871 unique missense or insertion/deletion (indel) KCNH2 variants; 4,810 individuals were heterozygous for these variants (< 0.001 minor allele frequency [MAF]), 1,041 of which were diagnosed with LQT2 according to our classification criteria (see Materials and Methods for details). From five arrhythmia centers in France, Italy, and Japan, we collected a cohort of patients heterozygous for KCNH2 variants. From these sites, we identified 266 unique missense or in-frame indel variants in KCNH2, 933 heterozygote carriers of these variants, 594 of which met our criteria for LQT2. From this cohort, the average age of diagnosis or ascertainment (if criteria for affected status were not met) within the participating sites was 24 years old (standard deviation of 19 years; available for 744 individuals in the cohort dataset). Gender was distributed 45% male.

Rates of LQT2, observed in the literature and our cohort, is associated with in vitro and in silico features.

To assess the association between in vitro and in silico characteristics of I_Kr/K_V11.1/KCNH2 and the fraction of heterozygotes which present with LQT2, we calculated non-parametric Spearman rank order coefficients (Spearman ρ) between these features and the observed literature (Figure 1, black) or cohort (Figure 1, red) LQT2 probability (the fraction of heterozygotes diagnosed with LQT2 over total number of heterozygotes). We evaluated common in silico predictors, electrophysiological parameters for I_Kr, and a by-residue average observed LQT2 probability in a three-dimensional shell surrounding each residue (LQT2 probability density, see Materials and Methods for details). The variant-specific features LQT2 probability density, REVEL, and heterologous measurement of variant I_Kr peak tail currents had Spearman ρ absolute value point estimates around 0.6 and 0.7 in the literature and cohort datasets, the highest that reached a nominal p-value of 0.05.

Figure 1.

Weighted Spearman correlations between the fraction of heterozygotes diagnosed with LQT2 in the literature or cohort and the listed features for each variant. Weighted Spearman correlations between the fraction of heterozygotes diagnosed with LQT2 in the literature and listed features for each variant. Black and red squares indicate the estimate for the weighted Spearman correlation, weighted by $1-\frac{1}{0.01+totalheterozygotes}$, for the literature and cohort dataset, respectively. The grey and red lines indicate 95% confidence intervals (obtained by bootstrap), respectively. Larger box sizes indicate greater number of variants included in calculation.

Two broad in silico variant classifiers, PROVEAN and PolyPhen-2, had lower Spearman ρ point estimates than structure and peak tail currents (weighted Spearman ρ of 0.39 and −0.59 in the literature dataset and 0.33 and −0.46 in the cohort set for PolyPhen-2 and PROVEAN, respectively), though still statistically significant. Repeating this analysis in the cohort dataset produced mostly lower coefficients, though many retained statistical significance (Figure 1). Several biophysical properties were not statistically significant in either dataset.

Magnitude of LQT2 probability varies by residue location in three-dimensional space.

Given the relatively high correlation between LQT2 probability density and observed LQT2 probability, we mapped LQT2 probability density on to the structure of K_V11.1 (Figures 2, S2, and S3). Figures 2 and S2 demonstrate LQT2 penetrance is not uniformly distributed over the major domains in K_V11.1 (see Figure S2 for specific examples). This is in contrast to averaging over variants in sequence space as shown in Figure S4 and done previously.^{14, 15} For instance, the transmembrane segment in K_V11.1 includes the voltage-sensing domains and pore domains, each of which have their own subdomains with high or low LQT2 penetrance. Some of these subregions are very small, localized to only a few contacting residues. For example, the voltage-sensing domain has a relatively low penetrance density in the intracellular half of helices S2 and S3 (Figure S2), while the most highly penetrant residues in this domain are in helices S1 and S4, which contact the pore domain near the midpoint of the membrane bilayer. Similarly, the pore domain shows the highest penetrance density near the selectivity filter and decreases towards the intracellular face of the pore. Additionally, the N-terminal Per-Arnt-Sim (PAS) domain and C-terminal cyclic nucleotide binding homology domain (CNBhD), both having relatively high observed LQT2 penetrance overall (Figure S4), are also heterogeneous. The most highly penetrant residues in these domains exist near the contacting surfaces between and among these domains. These trends are more muted in the in the observed LQT2 probability from cohort and literature data viewed linearly (Figure S4).

Figure 2.

LQT2 probability density mapped on to K_V11.1 structure. LQT2 probability density mapped on to the three-dimensional structure of the K_V11.1 channel. Larger and redder segments indicate higher LQT2 probability density; smaller and bluer segments indicate lower LQT2 probability density. The model illustrates structural information regarding amino acids predicted to increase disease probability. Unlike linear graphical displays identifying pathogenic loci, LQT2 probability density provides novel, three-dimensional, insights into the specific structural components of the PAS, CNBhD, voltage-sensing, and pore domains that are associated with increased prevalence of LQT2.

Estimated post-test probability of LQTS based on KCNH2 variants found in only one heterozygote is predictive.

Variants found in only a single known heterozygous individual are the largest class of variants in the literature and our cohort data. Accordingly, we split the data into two groups: 1) variants with two or more heterozygous individuals and 2) variants with only one heterozygous individual (Figure S5). We then estimated the post-test probabilities of LQTS based on KCNH2 variants from group 1 (those with at least two heterozygous individuals). A Bayesian model was fit using an expectation maximization algorithm (see Materials and Methods for details and ref. ¹¹). The predictive ability of our post-test LQT2 probability estimates were evaluated using the area under a Receiver Operating Characteristic curve (AUC) from group 2, those found in only one heterozygous individual (Figures 3 and S5). Additionally, we evaluated models fit on the full literature dataset on variants found in the cohort dataset (Figure 3, bottom, and Figures S5–S6). In all cases, the estimated AUC from our method outperformed other existing algorithms (LQT2 probability density, REVEL, PROVEAN, PolyPhen-2, BLAST-PSSM, and PAM score). For single heterozygotes in the literature, we observed AUCs of 0.87, 0.84, and 0.83 for our post-test probability model, REVEL, and LQT2 probability density, respectively. PROVEAN and PolyPhen-2, with AUCs of 0.74 and 0.73, respectively, were lower than our method and REVEL, as expected, since REVEL included PROVEAN and PolyPhen-2 as predictive covariates during construction.¹⁶ For variants with single heterozygotes in the cohort dataset, (Figure 3, bottom), AUC point estimates were lower overall, 0.78 and 0.77 for our method and LQT2 probability density, respectively. Surprisingly, REVEL scores were much less predictive in this group of variants, producing an AUC of 0.65, compared to 0.84 in the literature dataset. These differences in AUCs were also present when all variants were included and evaluated at various observed probability cutoffs (Figure S6). Previously published in silico predictors PROVEAN and PolyPhen-2 each had an AUC of 0.70, similar for the literature and cohort.

Figure 3.

Receiver operating characteristic curves of features sorting variants with only one heterozygote observed. Receiver operating characteristic curves from predictors against variants with only one observation, an individual affected with LQT2 or not, in the cohort (Figures S5). The carriers come from either the literature (above) or the cohort (below). The estimated post-test probability and LQT2 probability density were not exposed to the evaluation variants, those whose heterozygote count is equal to one, during training. All cohort data were withheld from the EM and LQT2 probability density during construction (Figure S5).

Post-test LQT2 probability estimates are improved by including KCNH2 variant features.

The R² between our LQT2 probability estimates and the cohort observed LQT2 probability is 0.30 when all variants are included, higher than in silico classifiers or LQT2 probability density; R² estimates are higher overall when restricting to the set of variants where heterozygously collected peak tail current is known (Tables 1 and S1).

Table 1.

Weighted R² between the fraction of heterozygotes diagnosed with LQT2 in the literature and cohort with estimates.

LQT2 Probability Estimates	Literature (n = 706)†	>Cohort (n = 246)†
LQT2 probability density	0.49 [0.39-0.60]	0.23 [0.12-0.34]
REVEL	0.38 [0.31-0.44]	0.21 [0.11-0.33]
Post-test LQT2 Probability	0.82 [0.77-0.86]	0.30 [0.19-0.43]

^†Weighted R² [95% Confidence Interval] for the same subset of variants, weighted by $1-\frac{1}{0.01+totalheterozygotes}$$$, n is the number of unique KCNH2 variants

Since probability estimates are generally more reliable as the number of phenotyped heterozygous individuals increases, we calculated R² at varying cutoffs of heterozygote count (Figure 4). As we restrict the analysis to variants with higher numbers of heterozygotes, we see R² between our LQT2 post-test probability predictions and observed LQT2 probability substantially increase in both datasets (Figure 4). This shows that LQT2 probability predictions are statistically significant across sources.

Figure 4.

Coefficient of determination determined for LQT2 post-test probability predictions and fraction of homozygotes diagnosed with LQT2 from the cohort or literature. Coefficient of determination determined between EM LQT2 probability predictions and observed LQT2 probability from the cohort or literature. There are fewer variants to analyze as we restrict to variants with higher heterozygote counts.

GnomAD data are critical to build the most robust LQT2 probability estimates.

When gnomAD heterozygotes are removed from the literature dataset, the mean weighted probability observed in the cohort and the literature sets are closer to each other; this is also reflected in empirical probability distributions (Figure S7). However, rank order correlation between the literature and the cohort was reduced: without gnomAD, Spearman’s ρ between literature and cohort was 0.26 [95% confidence interval of 0.01-0.50]; when gnomAD was added to the literature ρ = 0.35 [95% CI of 0.11-0.58]; and when gnomAD was added to the cohort ρ = 0.33 [95% CI of 0.09-0.55]. In addition, predictive models trained from the literature without gnomAD resulted in lower AUCs and R²s, due in part to the reduced information in the LQT2 probability density feature (Table S2 and Figures S8–S9). These results demonstrate the importance of including control variants, such as those from gnomAD, in the LQT2 post-test probability estimates.

Example LQT2 probability estimates for a segment of K_V11.1.

The outcome of our analysis is a range of data-driven post-test probabilities for each variant, initial probabilities conditioned on variant-specific properties, and posterior probabilities after heterozygous individuals are added. Each estimate is a probability distribution with a 95% credible interval. We illustrate the outcome of this method for variants in a segment within K_V11.1 from p.Leu622 (c.1866) to p.Arg752 (c.2255) in Figure 5. Residues towards the extracellular face of K_V11.1 have a higher prior and posterior estimated probability. The LQT2 probability prior probabilities conditioned on heterozygously-collected peak tail current, LQT2 probability density, and REVEL score, are near the observed probability (LQT2/total heterozygotes) for most variants (Figure 5).

Figure 5.

Example probability predictions for segment of KCNH2 including the selectivity filter and S6 helix. Example probability predictions for segment of KCNH2 including the selectivity filter and S6 helix. Priors were generated from in vitro (when available) and in silico covariates. Bars indicate the 95% interval of the prior. The dot and lines reflect the point estimate and 95% credible interval of the posterior, after the observations of affected individuals and those not meeting the threshold for affected status are included. Number of heterozygotes follows variant name on the left side of the figure. To the right, is a translucent structure of the KV11.1 channel. KV11.1 from p.Ser621 (c.1863) to p.Arg752 (c.2255) is represented as a solid cartoon, with two segments highlighted in different colors. The blue region highlights a segment with higher overall LQT2 probability while the red region highlights a segment with lower overall LQT2 probability.

Equivalence between KCNH2 variant features and clinically phenotyped heterozygotes.

The width of initial prior probability intervals conditioned on variant features (Figure 5, solid-colored lines) are determined by choice of ν in Equation 1 as previously described ¹¹:

$${\sigma }_i=\frac{{\mu }_i\left(1-{\mu }_i\right)}{1+\nu }$$ Equation 1

where ν represents the number of observations in the beta binomial model, in this case, clinically phenotyped individuals heterozygous for variants in KCNH2, σ_i is the variance in the beta-binomial model and μ_i is the mean penetrance estimate for the i^th variant. As ν grows, the prior 95% interval narrows; as ν decreases, prior 95% intervals expand. For very large ν, e.g. ν = 100, the posterior estimates of LQT2 post-test probability are heavily influenced by the prior such that very many observations of heterozygotes (1,000-10,000) are required to significantly change the posterior. At the other extreme of very small ν, e.g. ν = 1, the posterior estimates of LQT2 probability are largely independent of the priors. Acceptable values ofν would be those where 95% of variants have true LQT2 probabilities within the posterior 95% credible interval. To find values ofν where this was the case, we calculated posterior coverage rates by adding hypothetical heterozygotes sampled at the observed LQT2 probability to the prior generated with multiple values of ν, as described in the supplemental text and shown in Figures S10–S12. This procedure resulted in a range of acceptable ν values near ν = 10. Heuristically, for each variant, the post-test estimate of LQT2 probability carries the information equivalent to clinically phenotyping approximately 10 heterozygotes.

Discussion

Spectrum and example of LQT2 diagnosis probability attributable to KCNH2 variants.

Few KCNH2 coding variants have been discovered in a sufficiently large population to reliably estimate their post-test probability of developing LQT2 as defined in the Materials and Methods. However, variants such as p.Lys897Thr (c.2690A>C), p.Arg176Trp (c.526C>T), p.Val822Met (c.2464G>A), and p.Ala561Val (c.1682C>T) have been observed in many clinically phenotyped individuals both in the literature and in our assembled cohort. These variants span both the spectrum of channel defect (measured as heterozygously-collected peak tail current compared to WT) and spectrum of LQT2 disease probability. The most common KCNH2 coding variant, p.Lys897Thr, induces a very modest channel phenotype (peak tail current 78% of WT)¹⁷ and is common enough (5-24% of alleles)¹⁸ to preclude a large influence in LQT2 diagnosis, though its presence may modify risk.¹⁷ p.Arg176Trp induces peak tail current between 50-75% of WT ^{19, 20} and has a well-established LQT2 probability, estimated at 20%.²¹ We observe a similar LQT2 probability of 35% in the literature and 43% in the cohort, higher values likely reflecting a bias in ascertainment (also discussed below). p.Val822Met induces a significant channel defect, peak tail current of 44% of WT,²² and we correspondingly observe a higher LQT2 probability from the literature (65%) ^{23, 24} and cohort (60%). p.Ala561Val induces a severe channel defect, peak tail current between 0 and 46% of WT with a mean near 20% ^25–28; we observed a LQT2 probability of 91% from the literature and 88% from the cohort. These variants illustrate that molecular defects induced by genetic variants in KCNH2 place heterozygotes at higher risk for LQT2.

Our framework allows us to exploit this relationship in part by conditioning estimates of disease probability on these defects, directly (in vitro data) or indirectly (in silico data). Our resulting model is informed largely by variants with many classified heterozygotes, like the variants just mentioned, but is most useful for variants with few or no known heterozygous carriers. Here, we validated our method with a diverse, international cohort of clinically phenotyped KCNH2 variant heterozygotes curated from among five centers; the cohort was withheld during all training stages and all potential overlapping individuals were removed. We tested how well our prior LQT2 probability estimates discriminate variants observed once in individuals who do or do not meet the criteria for LQT2 diagnosis (Figure 3) and correlate with the observed LQTS probability/penetrance (Figure 4). Lastly, all performance statistics reported for the probability density covariate were generated using leave-one-out cross validation, i.e. the probability density derived for each variant was never exposed to the observed LQT2 probability/penetrance for that variant.

Structure combined with previously described variants produced the most predictive feature of observed cohort LQT2 probability.

Variant position in K_V11.1 domains such as transmembrane, pore, or intracellular is associated with differential risk of events.^{14, 29, 30} Expanding on this observation, and leveraging the recently determined K_V11.1 channel structure (PDBID: 5VA1),³¹ we developed a metric to quantitate average LQT2 probability in the three-dimensional space surrounding each residue (Figures 2 and S2). The resulting metric, LQT2 probability density, was comparable to the in silico predictor REVEL in terms of AUC and R² (Figure 3 and Table 1). Alone, LQT2 probability density could explain 50% of the variance in LQT2 probability as observed in the literature (Table 1); this reduced to 23% in the cohort but was still more predictive than even in vitro heterozygously-collected peak tail current data (Table S1). We attribute at least part of this decrease to greater ascertainment bias and lower heterozygous carrier counts in the cohort dataset. In addition, the R² of 0.23 (up to 0.3 when including all covariates) is the most pessimistic coefficient of determination. When restricting to variants with greater heterozygote counts, the R² improves to around 0.8 and so we estimate the generalized R² is likely closer to a clinically meaningful value.

Bias in data collected from the literature, gnomAD, and cohort.

The clinical environment taxonomizes KCNH2 variants disproportionately from patients who present with disorders.^{32, 33} For example, individuals heterozygous for KCNH2 p.Arg176Trp (annotated in ClinVar variously as a risk factor, Likely Benign, and Variant of Uncertain Significance) have a mean QTc of 459 +/− 40 ms in clinical LQT2 families, those with at least one proband, but a mean QTc of 433 +/− 27 ms in a cross-sectional cohort of unselected heterozygotes.³⁴ Similarly, we found most variants in the cohort have higher LQT2 probabilities than what we observed in the literature (Figures 6 and S13). Some of these variants have statistically-significant differences in observed LQT2 probability between the datasets (Figure 6B). Though all datasets have biases, adding heterozygous individuals from gnomAD to the available literature yields LQT2 probability estimates more consistent with the cohort (Tables S2 and Figures S8–S9) and we therefore suggest the combined datasets produce the most accurate, though flawed, estimate of variant-specific LQT2 probability.

Figure 6.

Ascertainment bias in LQT2 probability for KCNH2 variants from the arrhythmia center cohort compared to the literature. (A) The 30-residue moving average of LQT2 probability for each amino acid over the entire K_V11.1 channel for data acquired from the cohort (dark gray) and collected from the literature (light gray) (see Figure S4 for the distribution of variants used in the calculation). Ascertainment bias is evident by the higher overall LQT2 probability in the cohort dataset compared to the literature (which includes gnomAD). (B) Observed probability (literature and the cohort) for a selected set of KCNH2 variants, p-values from Fisher’s exact test between the observations of heterozygotes in the literature and in the cohort. The observed probability for many variants from the cohort is significantly higher than that calculated from the literature. The number of heterozygous carriers discovered in each group is shown directly above the variant names.

Evidence that some variants classified as Benign increase the probability of LQT2 to higher than the general population rate.

Similar to KCNH2 p.Arg176Trp, variants such as p.Pro347Ser (c.1039C>T), p.Arg148Trp (c.442C>T), p.Ala913Val (c.2738C>T), and p.Arg328Cys (c.982C>T) were previously associated with LQT2,^35–39 but are also more common in the general population than is expected for highly penetrant variants.¹⁸ This trend, observed in several variants which also produce a functional perturbation consistent with LQTS, has prompted some to use the label “LQT-lite”.^{40, 41} We also observed these variants in affected individuals in the cohort (2 out of 5, 4 out of 19, 1 out of 4, and 1 out of 2 heterozygotes, respectively). The most recent classifications for these variants in ClinVar are Benign or Likely benign. However, we estimate the probability of LQTS for heterozygous carriers of these variants at around 2% or higher, much greater than the ~0.04% in the general population. These data suggest some variants classified as Benign or Likely Benign are truly disease-causing for a small fraction of patients.

Application of Bayesian Probability to Arrhythmia Genetics Clinics.

Bayesian reasoning has long been at the core of clinical diagnosis. Given that the majority of heterozygous carriers of rare KCNH2 variants found in arrhythmia genetics clinics (or elsewhere) carry ultra-rare or novel variants, we anticipate that this prior, trained on variant-specific features, will have direct clinical utility to the clinician. The addition of 10 equivalent observations for a previously unreported or seldomly reported variant can directly guide clinical management and establish a threshold of intervention with drug treatment or simple clinical observation. Additionally, this method helps overcome ascertainment bias prevalent in clinically-obtained data since the prior is trained on variant-specific features agnostic to clinical information (Figures 5 and 6). While there is no intention to replace clinical phenotyping, we do anticipate the ability to augment clinical reasoning through a more accurate prior when combining clinical and population features with variant-specific features.

As an example, p.Pro347Ser, a variant with an observed 40% penetrance in the clinical cohort, would likely result in treatment intervention if clinically encountered by a clinician familiar with the variant through families seen in their clinic (Figure 6). However, from variant-specific features, our analysis generated a prior for p.Pro347Ser equivalent to observing only 1 in 10 heterozygous individuals diagnosed with LQT2. This new information could permit a more flexible approach to workup if no other information were known. A relatively high number of observations of p.Pro347Ser in the literature and gnomAD, which also suggest a LQT2 probability/penetrance of less than 5% for this variant (Figure 6), helps illustrate the calibration of our variant-informed priors. In this way, joint clinical phenotyping and tool utilization can be used in a mutually beneficial way for patients heterozygous for rare variants. We have developed an online searchable tool, the KCNH2 Variant Browser (https://variantbrowser.org/KCNH2/), to allow rapid access to the estimated penetrance based on variant-specific features.

Limitations

One limitation is the bias inherent to each of the data sources used. We may be able to observe more carriers of these KCNH2 variants as greater numbers of individuals are exome sequenced; however, for many variants, we may never observe more carriers and will be underpowered to estimate LQT2 probability by observation of heterozygotes alone. This fact is further motivation to establish a framework where experimental data is included quantitatively in the estimate of disease probability. The availability of functional data is also biased in that most variants which have these data available are from variants discovered in individuals presenting with a phenotype (Figure S14), however, high-throughput variant functional characterization has the potential to overcome this bias.⁴² Additionally, many factors influence the ultimate presentation of LQT2 in an individual, including genetic and environmental factors,^{4, 43, 44} though we did not observe significant differences in predictive performance across nationality (Table S3). Although the largest effect sizes for LQTS-associated variants come from rare variants, some of the variability in LQTS presentation can be explained by variability in common variants. Recently, two publications by Lahrouchi et al.⁴⁵ and Turkowski et al.⁴⁶ concluded polygenic risk scores accounted for ~15% and < 2% of the variability in LQTS susceptibility. Though in either case the contribution is relatively small, it is possible polygenic risk is potentiated in the rare variant context and could therefore explain a greater portion of the variability in disease presentation. Lastly, though beyond the scope of the present study, we envision this method will enable improved prognostication of more severe presentations of LQT2 including arrhythmic events. Future work will address these exciting possibilities.

Conclusions

We have shown how variant-specific features can inform a prior probability of disease for rare variants even in the absence of clinically phenotyped heterozygotes. We have demonstrated this framework on the classical Mendelian disease-gene pair, LQT2 and KCNH2. We exploit in vitro functional studies, LQT2 probability density, and broad in silico predictors to calculate these priors. We then combine these estimates with patient data to form the post-test probability of disease for each variant. We have demonstrated that these in vitro and in silico variant features are equivalent to approximately 10 clinically characterized heterozygotes when used to understand KCNH2 variant-specific LQT2 disease probability. Presenting these data in this way allows us to encode both the probability of disease and the uncertainty in our estimates: we do not claim to have as much certainty as you would have if you phenotyped 100 heterozygotes; however, we do claim greater certainty than a single observation of a phenotyped heterozygote.

Nonstandard Abbreviations and Acronyms:

LQT2

long QT syndrome type 2

ACMG

The American College of Medical Genetics and Genomics

PPV

positive predictive value

ECG

electrocardiogram

indel

in-frame insertion/deletion

MAF

minor allele frequency

PAS

Per-Arnt-Sim

CNBhD

domain and C-terminal cyclic nucleotide binding homology domain

AUC

area under a Receiver Operating Characteristic curve

WT

wildtype