Evaluating the Utility of a New Pathogenicity Predictor for Pediatric Cardiomyopathy

Pediatric

The causes of CM are diverse and include genetic variants, infectious agents, environmental factors, neuromuscular disease, and inherited metabolic disorders [1].Pathogenic variants are identified in 50-65% of patients with pediatric HCM [5], 70% of patients with adult familial HCM [6], and 40-50% of patients with adult DCM [7].Genetic testing has become an integral part of patient care for both children and adults with CM.Genetic testing can help confirm a clinical diagnosis, guide management decisions, and aid in familial risk assessment [8,9].Previous research in pediatric CM has focused on identifying pathogenic variants for diagnostic purposes.However, the presence of a pathogenic variant alone is insufficient to predict clinical outcomes and adverse events in pediatric cardiomyopathy [10].There is limited research evaluating the relationship between variants and clinical outcomes.
With advances in next-generation sequencing, an increasing number of genetic variants, both pathogenic and variants of uncertain significance (VUS), are emerging as potential causes of CM [10].Historically, VUS have not been considered when evaluating genotype-outcome associations.However, a recent study suggested that increased variant burden, including VUS, was associated with worse clinical outcomes in patients with pediatric DCM.Thus, consideration of both pathogenic variants and VUS as potential risk modifiers may help stratify risk for adverse clinical outcomes [10].Despite this, interpretation of VUS and rare genetic variants still represent a significant challenge in clinical genetic testing for CM.Accurate discrimination between benign and pathogenic variation remains essential for proper management of patients with pediatric CM [11].
Determining the clinical significance of a variant is a complex process that involves gathering a multitude of data from a combination of resources.Computational prediction of variant pathogenicity is integrated as one line of supporting evidence to assess the clinical significance of a variant [12].Existing genome-wide machine learning tools learn from large-scale data over the entire genome.However, they are less accurate and less specific than disease-specific variant classifiers [13].CardioBoost is a disease-specific variant classifier that estimates the probability of pathogenicity for rare missense variants in inherited cardiomyopathies and arrhythmias.CardioBoost outputs a continuous probability of pathogenicity that is directly interpretable by the user.CardioBoost has been shown to outperform existing genome-wide tools in classification performance, prediction of disease association, and stratification of patient outcomes.2 Human Mutation with both disease status and an increased risk of severe clinical outcomes in adults with HCM [13].
The purpose of this study is to determine whether Cardi-oBoost can be used to better stratify disease severity and clinical outcomes in patients with pediatric CM.Genetic risk stratification of pediatric CM could have significant clinical implications including individualizing cardiac treatments and providing prophylactic interventions before adverse events occur.

Study
Cohort.This is a retrospective, single-center cohort study that evaluated outcomes in pediatric patients (age < 21 years at diagnosis) with a history of cardiomyopathy (including DCM, HCM, restrictive, LVNC, or ARVD) who were evaluated by the Children's Hospital of Philadelphia (CHOP) Cardiomyopathy Program between January 1, 2010, and August 31, 2018.Patients were excluded from the analysis if there were insufficient records for review or if they had negative phenotype on screening.Cardiomyopathy was categorized as DCM, HCM, or other (restrictive, LNVC, or ARVD).Data, including demographic information, cardiomyopathy diagnosis, cardiac surgical information, physical examination, cardiac and metabolic laboratory data, cardiac imaging results, and genetic test results, were abstracted from the medical record.The Children's Hospital of Philadelphia Institutional Review Board (IRB #18-015616) approved this study with a waiver of consent.This work was previously published on ProQuest as a Master's thesis for ST [14].
ACMG classification was determined by a clinical geneticist (RAN) for each variant based on data available at time of data abstraction [12].CardioBoost scores were generated for patients who meet study criteria and CardioBoost eligibility.CardioBoost can output predictions for rare (allele frequency < 0 1% in Gnomad) missense variants in a subset of genes associated with CM (CACTC1, DES, GLA, LAMP2, LMNA, MYBP3, MYH7, MYL2, MYL3, PLN, PRKAG2, PTPN11, SCN5A, TNNI3, TNNT2, and TPM).For patients with more than 1 variant with a CardioBoost score, the maximum score was utilized.Scores were then categorized as ≤0.1, 0.1-0.9, and ≥0.9 based on distribution of scores and CardioBoost's predefined categorizations.A CB score of ≤0.1 indicated a predicted pathogenicity of benign/likely benign.CB score of 0.1-0.9indicated a predicted pathogenicity of VUS.CB score of ≥0.9 indicated a predicted pathogenicity of likely pathogenic/pathogenic.Concordance between ACMG classification and CB classification was evaluated based on these three categories of classification (benign/likely benign, VUS, and likely pathogenic/pathogenic).

Statistical Analysis.
The composite endpoint was freedom from major adverse cardiac event (MACE), defined as internal cardioverter defibrillator (ICD) implantation, myectomy, mechanical circulatory support (ventricular assist device (VAD)) or extracorporeal membrane oxygenation (ECMO), cardiac transplant, and aborted cardiac arrest or death.The need for transplant evaluation with the deci-sion to not list due to patient or family request was included in the composite endpoint.Secondary endpoints included freedom from individual MACE endpoints.
Demographic, clinical, imaging, and genetic variables were summarized using standard descriptive statistics.Categorical variables were reported as counts (with percentages), and nonnormally distributed continuous data were reported as median (with interquartile range (IQR)).Chi-square, Fisher's exact, Wilcoxon's rank-sum, and Kruskal-Wallis's tests were used to explore the relationship between Cardio-Boost score and patient characteristics.Bivariate odds ratios (ORs) and 95% confidence intervals (95% CIs) are reported for overall MACE and each component of MACE.The Kaplan-Meier curves are presented as time to MACE overall and stratified by CardioBoost scores ≤ 0 1, 0.1-0.9, and ≥0.9.Models are presented for the association between Cardio-Boost score and outcome via logistic regression and are adjusted for known confounders of outcome (race, sex, and age at first encounter).Logistic regression was performed as only date at MACE was available, not individual dates at each endpoint.Statistical significance is determined as a two-tailed p value ≤ 0.05.Analyses were conducted utilizing Stata version 17.

Genetic
Testing and Variant Distribution.The most common genetic test was a cardiomyopathy panel, which was performed in 81% overall, 88% of DCM cases, 71% of HCM cases, and 89% of other cases (Table 2).Whole exome sequencing was performed in 7% overall, 6% of DCM cases, 4% of HCM cases, and 11% of other cases.Whole exome sequencing with mitochondrial sequencing was performed in 5% overall, 9% of DCM cases, 4% of HCM cases, and 0% of other cases.Patients with HCM were more likely to have targeted testing for a known familial variant (p = 0 001), which was performed in 14% overall, 6% of DCM cases, 27% of HCM cases, and 4% of other cases.
The distribution of genetic variants by ACMG classification was not significantly different across phenotypes (Figure 3).Around half of the overall cohort had at least one pathogenic variant, and around three-quarters had at least one VUS.The HCM cohort had the highest proportion of pathogenic variants, and the other cohort had the highest proportion of VUS.Inheritance of variants was most often unknown but was not significantly different across CM phenotypes.

Patient
Characteristics by CardioBoost Score.Table 3 compares patient characteristics between those with Cardio-Boost (CB) scores ≤ 0 1, 0.1-0.9, and ≥0.9.The final study cohort had 16 scores ≤ 0 1, 14 scores 0.1-0.9, and 74 scores ≥0.9.The distribution of CB score was significantly related to race/ethnicity (p = 0 025) with a higher proportion of CB score ≥ 0 9 in white than nonwhite participants (71.6% versus 28.4%).The distribution of CB score was also significantly related to sex (p = 0 023) with a higher proportion of CB score ≥ 0 9 in male than female participants.There was no significant difference between age of first encounter and age at positive phenotype based on CB score.The most common reason for presentation for those with CB score ≤ 0 1 was heart murmur (44%).The most common reason for presentation for those with CB scores 0.1-0.9  5 Human Mutation classified by CB as benign/likely benign and ACMG as VUS.Conversely, 3/62 (4.8%) variants were classified by CB as VUS and ACMG as benign/likely benign.Approximately one-quarter (16/62, 25.8%) of variants were classified by CB as pathogenic and ACMG as VUS.A CB classification of VUS and ACMG classification of likely pathogenic/ pathogenic were seen in 11/62 (17.7%) of variants.Finally, there was one variant (GLA c.1088G>A, p.Arg363His) with a CB classification of benign/likely benign and ACMG classification of likely pathogenic/pathogenic.This variant is a well-described recurrent variant identified in individuals with Fabry's disease [15], highlighting the possible limited utility of CB classification for nonsarcomeric variants.

Conclusions
This study did not identify a relationship between CB score and clinical outcome in patients with pediatric CM.Our study was modeled after a recent study by Zhang et al. that evaluated the ability of CB to stratify outcomes in a cohort of 803 adults with HCM [13].They found that variants classified as disease-causing by CB were associated with a 21% increased risk of severe adverse outcomes by age 60.There are several possible explanations for why this finding was not replicated in a pediatric cohort, including a smaller sample size and a shorter length of follow-up.Pediatric CM is more rare than adult CM, with an incidence of around 1/100,000 [1] children compared to 1/500 adults [15].This distinction is reflected in our respective sample sizes.It is also important to consider that pediatric and adult CM are different conditions with distinct pathological mechanisms [16].A recent study investigating the genetic architecture of pediatric CM found that although genes identified from adult CM studies provide clinical value, it is still important to consider additional genes, as well as multigenic inheritance models, in the pediatric setting [16].Before adjusting for confounding variables, we found that CB score ≥ 0 9 was protective for death.However, after adjusting for known confounders of outcome (race, sex, and age at first A limitation of this study was the length of follow-up.Our median age at last encounter was 14.9 years; therefore, we did not capture adverse events that may have occurred in later adolescence or adulthood.It is also important to highlight the limitations of using CB in the pediatric setting.First, the CB tool was trained with adult data and might not perform as effectively for pediatric disease.In addition, CB scores are only available for 16 CM-related genes.Our cohort began at 886 patients, and 556 were excluded from the final analysis because they did not have a variant interpretable by CB.Future studies can consider following patients for a longer period to include more longitudinal pediatric data.In addition, multicenter evaluation of patients with pediatric CM could increase the sample size and generalizability of the study.The addition of new genes to CB may also allow for a larger sample size and help improve the clinical utility of the tool.Due to the intrinsic differences in the genetic architecture of adult and pediatric CM, it may be beneficial to create a variant classifier that is specifically designed for pediatric CM.Finally, concordance between ACMG and CB classification was only seen in 50% of unique variants identified in this cohort.This highlights the importance of multiple lines of evidence and rigorous application of ACMG classification guidelines for proper interpretation of these variants [12].Overall, this study speaks to the continued deficits in variant interpretation for pediatric  7 Human Mutation cardiomyopathy, and we recommend using caution when applying CardioBoost to stratify clinical outcomes in this population.

Figure 3 :
Figure 3: The distribution of genetic variants is presented across the overall cardiomyopathy cohort (a), the dilated cardiomyopathy cohort (b), the hypertrophic cardiomyopathy cohort (c), and the other cohort (d).
While it is important to recognize that CardioBoost scores are one line of evidence in variant classification, variants classified as disease-causing by CardioBoost are associated

Table 1 :
Demographics and clinical characteristics.

Table 2 :
Genetic testing by cardiomyopathy cohort.Histogram displays the bimodal distribution of CardioBoost scores across the overall cohort.
4Human Mutation and ≥0.9 was heart failure (50% and 38%, respectively).The distribution of CB score was significantly related to the presentation for family screening (p = 0 037) with a higher proportion of CB score ≥ 0 9 compared to other types of presentation.There was no significant difference in family history based on CB score.3.4.CardioBoostScore and ACMG Classification Concordance.ACMG variant classification as performed by a clinical geneticist and CB classification were concordant in 60/131 (49%) and discordant in 62/131 (52%) of unique variants identified in this cohort (Supplemental Table 1).Half (31/62) of the discordant classifications were variants

Table 4 :
Association of CardioBoost score and clinical outcomes adjusted for race/ethnicity (nonwhite vs. white), sex (male vs. female), and age at first encounter.Reference group is ≤0.1.n/a = model was not able to run given cell size.