Cardiovascular Disease Pathogenicity Predictor (CVD-PP): A tissue-specific tool for discriminating pathogenicity cardiovascular disease gene variants

doi:10.21203/rs.3.rs-3122220/v1

Download PDF

Article

Cardiovascular Disease Pathogenicity Predictor (CVD-PP): A tissue-specific tool for discriminating pathogenicity cardiovascular disease gene variants

https://doi.org/10.21203/rs.3.rs-3122220/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Introduction.

The interpretation of variants of uncertain significance (VUS) remains a challenge in the care of patients with established or familial cardiovascular diseases. 56% of potential variants within known cardiovascular risk genes are characterized as VUS and unbiased machine learning algorithms trained upon large data resources can stratify VUS into higher vs. lower probability of contributing to a cardiovascular disease phenotype.

Methods.

ClinVar pathogenic or likely pathogenic (P/LP) and benign or likely benign (B/LB) from 47 genes previously associated with monogenic cardiovascular diseases (MCVDs) were identified. A random forest model was trained using six-fold cross validation on these variants to build a predictive model of variant pathogenicity using measures of evolutionary constraint, deleteriousness, splicogenicity, local pathogenic variation, cardiac-specific exon expression, and population allele frequency. Predicted pathogenicity was computed as a linear outcome coupled with a naïve Bayes classifier to determine an optimal cut-off to distinguish VUS of pathogenic interest versus VUS with low likelihood of pathogenicity. Performance of our model was validated using variants for which ClinVar pathogenicity assignment changed between 2014 to 2022. As a proof-of-concept we demonstrated the utility of our model in the (CATHeterization GENetics [CATHGEN]) cohort.

Results.

Random forest identified a top-ranked model using ClinVar known P/LP and B/LB variants that weighted evolutionary constraint (CADD score) most heavily. The model accurately prioritized variants for which ClinVar clinical significance had changed from 2014 to 2022 (precision recall AUC = 0.97) and had equal or greater performance when compared to conventional in-silico methods for predicting variant pathogenicity. In the CATHGEN cohort, there was a higher burden of VUS of pathogenic interest in individuals with DCM as compared to controls without DCM (p = 8.2x10^− 15). Individuals in CATHGEN who harbored model predicted and ACMG/AMP reviewed pathogenic VUSs demonstrated that 27.6% had clinical evidence of the relevant disease. Lastly, variant prioritization using this model provided genetic diagnosis in 11.9% of CATHGEN patients diagnosed with HCM clinically who did not harbor a HCM genetic P/LP variant by initial ACMG/AMP review.

Conclusion.

We have developed a cardiac-specific model for prioritizing variants underlying familial cardiovascular disease syndromes. CVD-PP proves to have high performance in discriminating pathogenicity of VUS in MCVD genes. ACMG/AMP review and phenotyping of individuals carrying VUS of pathogenic interest in a large cardiovascular cohort support the clinical utility of this model.

Biological sciences/Computational biology and bioinformatics/Computational models

Biological sciences/Genetics/Mutation

Improvements in sequencing capabilities have increased detection of rare genetic variants both in the clinical setting and in large population-based cohorts, yet high throughput interpretation of the clinical importance of rare genetic variation remains difficult. A common method used to categorize rare genetic variants in a clinical context is the American College of Medical Genetics and the Association for Molecular Pathology (ACMG/AMP) standardized guidelines which groups variants into five categories: pathogenic, likely pathogenic, variants of uncertain significance (VUS), likely benign and benign. Unfortunately, the majority of identified rare variants in patients with suspected monogenic diseases are VUS^1,2. VUS present a conundrum in the clinical care of patients with known or suspected genetic disorders, with ongoing debate about whether VUS results should be returned to patients³. VUS also raise questions about how they should be used clinically, whether further segregation testing of the VUS in the family should be pursued, and how clinicians track VUS over time as they are frequently reclassified. Further, VUS are difficult for patients to understand³.

Interpreting VUS requires the accumulation of large amounts of supporting evidence that is time-consuming and often insufficient to reclassify the VUS^4,5. While direct functional assessment of the effects of a variant is an important component of variant interpretation, computational or in silico methods offer a more feasible, high-throughput and accessible approach that can scale with the increasing rate of genetic sequencing. Efforts to address this problem and streamline variant interpretation have led to the creation of several publicly available in silico tools for predicting variants including: CADD⁶, SIFT⁷, PolyPhen⁸, MetaSVM⁹, MetaLR¹⁰, Mutation Assessor¹¹, Mutation Taster¹², PROVEAN¹³ and FATHMM¹⁴. Yet, current ACMG/AMP guidelines do not heavily weight computational tools for interpretation of variant pathogenicity, most likely due to the conflicting variant interpretations amongst them⁴. Therefore, more robust, and accurate tools are needed to aid in the interpretation of variants to keep up with clinical sequencing demands and to aid in better clinical interpretation and decision-making of VUS.

Monogenic cardiovascular diseases (MCVD) encompass a variety of phenotypes, affecting multiple organ systems that often result in cardiac pathology including heart failure¹⁵. The ability to identify the disease-causing variant in an individual can have implications for treatment and screening, including cascade screening in family members. Earlier identification of these variants could allow for beneficial clinical monitoring and intervention (statin therapy, lipid lowering therapy)¹⁶. Additionally, deeper characterization of MCVD genetic variants, especially VUS, in expanding population-based cohorts will improve our understanding of the functional activity and clinical relevance of specific variants and genes.

We used an unbiased machine learning method to create a cardiac-specific predictive model of variant pathogenicity, which we refer to as Cardiovascular Disease Pathogenicity Predictor (CVD-PP). We applied this model to VUS in 47 MCVD genes. We validated the model within ClinVar reclassified variants and demonstrate clinical evidence for pathogenicity of model predicted pathogenic VUS using manual ACMG/AMG review and evaluation of clinical disease in a large cardiovascular cohort with deep phenotyping.

Characterization of MCVD genes in ClinVar

Prior to model training, we first characterized the spectrum of variants annotated in ClinVar in 47 MCVD genes (Fig. 1, Supplemental Table 1). Of the 60,416 ClinVar annotated variants in 47 MCVD genes of interest, the majority were VUS (n = 27,559 [45.6%], Fig. 2A), followed by likely benign variants (n = 15,834 [26.2%]), variants with conflicting interpretations (n = 6,223 [10.3%]), benign variants (n = 4,176 [6.9%]), likely pathogenic variants (n = 3,844 [6.4%]) and pathogenic variants (n = 2,780 [4.6%]) (Supplemental Table 2). Titin (TTN, implicated in hereditary cardiomyopathies), had the highest absolute number of VUS and conflicting variants (n = 9,664) while transforming growth factor beta receptor 1 (TGFBR1) had the highest proportion of VUS and conflicting variants (58.1%) (Fig. 2B).

Machine learning identification of a cardiac-specific model for predicting pathogenicity of VUS in MCVD Genes

The top performing random forest model tested, which we term CVD-PP, successfully segregated pathogenic and likely pathogenic (P/LP) from benign and likely benign (B/LB) variants using both CVD-PP scores. Variant annotations most important for identifying pathogenic variants were CADD PHRED score, distance to nearest pathogenic variant, gnomAD global minor allele frequency (MAF), Maximum Entropy score (MaxENT) score and cardiac exon expression (Table 1).

We applied CVD-PP to variants for which pathogenicity classification has changed in ClinVar since 2014 (n = 663) with the hypothesis that our model would readily recapitulate the most recent classification. Of these variants, 36 (5.4%) changed from P/LP to B/L), 146 (22.0%) changed from B/LB to P/LP and 481 (72.5%) changed from VUS to P/LP. CVD-PP performed with high accuracy (precision recall under the curve [PR AUC] = 0.97) in determining the “correct” (i.e., most recent reclassified category) pathogenicity of these variants (Fig. 3). These variants were then subset to variants annotated in other widely used in-silico tools (n = 443 variants), of which 6 (1.4%) changed from P/LP to B/L), 81 (18.28%) changed from B/LB to P/LP and 356 (80.4%) changed from VUS to P/LP. CVD-PP was then compared to these tools for assessing pathogenicity of variants. Using the linear score produced by CVD-PP, CVD-PP was amongst the highest ranked tool for predicting pathogenicity (Supplemental Table 3).

Lastly, CVD-PP was applied to the entire set of ClinVar-defined VUS or variants with conflicting interpretations in ClinVar in 47 MCVD genes (n = 33,782). CVD-PP predicted n = 9,858 (29%) of these variants to be P/LP and n = 23,924 (71%) to be B/LB, using a naïve Bayes optimal cutoff of 0.56 (Fig. 4).

DCM cases have a greater burden of predicted pathogenic VUS than controls

DCM cases (n = 247) and non-DCM controls (n = 5,019) were identified in CATHGEN (Supplemental Table 4). ClinVar VUS and conflicting variants as well as variants not present in ClinVar were selected for burden testing. Within the 19 DCM genes, 429 ClinVar defined VUS and conflicting variants and 604 variants not present in ClinVar were identified in DCM patients. Furthermore, 2,751 ClinVar defined VUS and conflicting variants and 4,157 variants not present in ClinVar in were identified in non-DCM controls (Supplemental Table 5). The distribution of alternate allele count of DCM gene variant multiplied by CVD-PP score was significantly higher in patients with DCM compared to non-DCM controls (1-tailed Wilcoxon rank sum p = 8.2x10^− 15). Further, the distribution of CVD-PP scores was different in known DCM genes compared to non-DCM in DCM patients (1-tailed Wilcoxon Rank Sum Test p-value < 2.2x10^− 16). This further supports the clinical utility of our model.

ACMG/AMG review and EHR review for expression of disease of predicted pathogenic VUS corroborates model performance

We reviewed the highest 50 CVD-PP ranked VUS (i.e., predicted P/LP, of which 49 were present in at least one individual in CATHGEN after quality filtering) and lowest 50 model-ranked VUS (i.e., predicted B/LB, of which 44 were present in at least one individual CATHGEN after quality filtering) using ACMG/AMG criteria. These 93 variants resided in 10 of the 47 MCVD genes. Of the variants with the highest CVD-PP model score ranks, a significantly greater proportion met ACMG/AMG criteria for being P/LP (26 out of 49 variants [53.1%]) compared with the variants with the lowest CVD-PP model score ranks (2 out of 44 variants [4.6%]), p = 1.1x10^− 6). The 28 variants meeting ACMG/AMG P/LP criteria are listed in Supplemental Table 6.

To review expression of disease in CATHGEN participants harboring these highest and lowest ranked variants, we first filtered all individuals for these variants. More participants harbored predicted B/LB and ACMG/AMG confirmed B/LB variants (n = 533) compared to predicted B/LB ACMG/AMG confirmed P/LP (n = 2), predicted P/LP ACMG/AMG confirmed P/LP (n = 43) and predicted P/LP ACMG/AMG confirmed B/LB (n = 64) carriers. Because of this, we randomly sampled the same number of participants with ACMG/AMG determined B/LB variants as participants with ACMG/AMG determined P/LP variant carriers (n = 45). None of the 642 individuals genotype positive for these variants harbored a ClinVar P/LP in the 47 MCVD genes. Deep EHR review for the relevant disease was then performed on the 90 CATHGEN participants harboring one of these highest ranked 50 or lowest ranked 50 CVD-PP score VUS. Twenty-nine of 43 CATHGEN participants with CVD-PP predicted P/LP and ACMG/AMG confirmed P/LP variants harbored variants in genes associated with autosomal dominant transmission and had sufficient EHR encounters to assess for disease; of these, n = 8 (27.6%) individuals had evidence of expression of the relevant disease (Supplemental Table 7). Both participants carrying predicted B/LB, ACMG/AMG determined P/LP variants had sufficient EHR data for diagnosis and were phenotype negative (Supplemental Table 7). Of the 45 randomly selected CATHGEN individuals carrying CVD-PP predicted B/LB ACMG/AMG confirmed B/LB or CVD-PP predicted P/LP ACMG/AMG confirmed B/LB variants, 36 had sufficient EHR data for diagnosis and had variants in genes contributing to autosomal dominant disease. All 36 of these participants were phenotype negative for the corresponding disease (Supplemental Table 8).

Predicted pathogenic VUSs in HCM increase diagnosis of patients in CATHGEN

We next evaluated the expression of disease in carriers of predicted P/LP variants to further test the validity of CVD-PP. For this, we selected HCM given the relatively high penetrance of pathogenic variants within this disease. In the CATHGEN cohort, 67 individuals with a clinical diagnosis of HCM did not carry a known ClinVar P/LP in an HCM-related gene (n = 11 genes, Supplemental Table 5). Of these, eight individuals (11.9%) harbored variants predicted to be P/LP by CVD-PP: 2 with VUS (MYH7 and MYBPC3); 3 with conflicting variants (MYH7 and MYBPC3); and 2 with variants not annotated in ClinVar (ACTN2 and MYH6). Further evaluation revealed that the CATHGEN participant harboring the predicted P/LP MYH7 VUS (K524R) had recently undergone clinical genetic testing and this variant was interpreted as a VUS by the treating clinician 2 weeks prior to the day the EHR review was conducted. This specific variant has a global mAF of 0.003%¹⁷ and was previously identified in an individual with HCM¹⁸. The K542T variant at the same position is absent from population databases but observed in a stillborn fetus with left ventricular noncompaction highlighting the importance of that genomic position¹⁹. These results suggest that CVD-PP has clinical utility in a subset of patients with known MCVD who test negative for a P/LP pathogenic.

Variants are interpreted, annotated and submitted to the ClinVar database to allow for easier interpretation and clinical use⁴. While ClinVar is a highly used resource, there is a high rate of discordance in variant interpretations deposited in ClinVar and further, many variants are VUS. The presence of VUS identified from clinical genetic testing or research genetic sequencing presents a conundrum for clinical care and return of results for MCVDs. Conflicting evidence in variant databases (ClinVar) further complicates variant interpretation. We have trained a cardiac-specific machine learning model named CVD-PP that is highly accurate in identifying VUS likely to be reclassified as pathogenic. CVD-PP has the potential to streamline variant interpretation for VUS and variants not previously annotated in ClinVar and in conjunction with careful clinical phenotyping and genetic counseling, could help improve workflows for use of VUS in clinical care or return of VUS results in research settings.

The incorporation of annotations considering evolutionary conservation via CADD, global MAF and tissue-specific expression leads to a high discriminative ability of CVD-PP. Our model also is cognizant of variation that affects tissue-specific splicing, which many of the current in silico tools are not able to recognize. We specifically trained our model using GTEx cardiovascular tissue data, however this approach could be extended to other non-MCVD diseases by using relevant GTEx tissues. Further studies are needed to evaluate model performance on different pathologies.

As a proof of principle, by using the scores generated by CVD-PP for variants in genes associated with DCM, we demonstrated a greater burden of predicted pathogenic VUS in CATHGEN participants with DCM compared to those without, suggesting that CVD-PP nominated VUS may be helpful in identifying a genetic cause in DCM. Using current approaches, a genetic diagnosis is only identified in 40% of familial cases and even lower in non-familial cases^20–22.

Furthermore, CVD-PP variant prioritization agreed with thorough review of ACMG/AMG criteria for pathogenicity in the highest and lowest ranked VUS in the previously described MCVD genes. EHR review of individuals in CATHGEN found 25.8% of the reviewed cohort had consistent symptoms and/or a diagnosis in agreement with the suspected phenotype for variants predicted to be pathogenic by CVD-PP and confirmed by ACMG/AMG review while no CATHGEN individual harboring a predicted benign VUS expressed disease.

We also showed that variant prioritization using CVD-PP increases identification of genetic variants likely contributing to disease in genotype-negative (i.e., without a culprit pathogenic/likely pathogenic variant identified) HCM patients who harbor a VUS in an HCM-gene in the CATHGEN cohort. Inclusion of predicted pathogenic VUS in HCM genes increases genetic diagnosis in CATHGEN patients diagnosed with HCM clinically. This further validates the clinical utility of the CVD-PP in a disease driven by highly penetrant genetic variants.

Altogether, these findings support the use of CVD-PP for prioritizing VUS as potential causal MCVD variants. Our approach could serve as a higher-throughput automated filter for variants that should be prioritized for subsequent ACMG/AMG review. This is particularly relevant to large population-based studies returning genetic test results in which several hundred VUS are identified⁵. Additionally, prioritizing VUS for ACMG/AMG review by use of CVD-PP would also allow for more streamlined interrogation VUS returned in the clinical setting. More feasible interrogation of VUS could potentially allow for increased genetic diagnoses, cascade screening in family members as well as targeted therapy for those with certain genetic variants²² in addition to more complete characterization of the genetic landscape of MCVD genes.

CVD-PP has several strengths as a pathogenicity prediction tool. First, this model uses static information, removing the need to continuously reinterpret VUS. Second, while the CVD-PP model is disease-specific (i.e., due to the inclusion of GTex left ventricle and aorta exon-level expression data), it could also be easily adapted to other diseases with relevant disease-specific tissue available. We note, however, that parameter importance might vary by disease. Finally, we have made the list of CVD-PP classified MCVD genetic coding variants and CVD-PP code available for public use. While we note these strengths and the potential clinical and research utility of CVD-PP, we also acknowledge limitations in this work. The validation set of variants (n = 663) is small and unbalanced, especially when subset to validation variants annotated by other tools (n-443). A larger, more balanced validation data set and inclusion of cellular or other direct functional confirmation of pathogenicity would be useful. Additionally, several other variant prediction tool scores could have been included in our model. However, given that many tools overlap in information used for variant prediction we decided to focus on CADD-PHRED and MaxENT scores, which provided the most orthogonal set of information used in variant pathogenicity prediction. The use of GTEx tissue expression is currently limited by the availability of only of bulk RNA-sequencing. Single-cell RNA-sequencing (scRNA-sequncing) data could refine our model’s understanding of exon usage in a cell-type specific manner. Finally, while the use of the CATHGEN cohort enabled deep chart review for evaluation of disease expression that is not feasible in studies with only ICD9/ICD10 diagnosis codes, we note that some participants had insufficient data to determine disease and that family history was not available. Further, while the model was highly accurate for predicting reclassification of the validation set of ClinVar variants and a relatively high proportion (28%) of CATHGEN participants harboring predicted pathogenic variants had evidence of disease expression, the majority did not have evidence of disease expression. We are unable to determine if this is because of inaccuracy of the CVD-PP model or because of incomplete penetrance of these VUS.

In conclusion, we find a high burden of VUS and variants with conflicting interpretations in MCVD genes. We have developed a cardiac-specific algorithm, CVD-PP, has high performance for predicting VUS pathogenicity in genes underlying MCVDs. We show that this tool is helpful for prioritizing VUS, offering the ability to characterize VUS in large population-based cohorts and potentially improve clinical care in conjunction with careful phenotyping, genetic counseling, and familial segregation assessment.

ClinVar identified MCVD genetic variants

ClinVar was used to identify pathogenic, likely pathogenic, likely benign, benign, variants of uncertain significance (VUS) and conflicting variants (treated as VUS) in 47 MCVD genes. These 47 genes were selected for this study based on inclusion of cardiovascular genetic testing panels and literature review for established association with the following MCVD: cardiomyopathies, familial hypercholesterolemia, arrhythmias, connective tissue disorders/aortopathies, congenital heart disease, and skeletal myopathies. ClinVar annotation for variants within these 47 genes was downloaded on February 2, 2022, and subset to the 47 genes using bedtools v2.30.0²³.

Machine learning model building

Ensemble variant effect predictor (VEP, version 106) and CADD version 1.6 were used for variant annotation. Predictors used in model training included: global gnomAD MAF (scaled by multiplying by 50,000), Maximum Entropy [MaxENT] model score (absolute value) (http://hollywood.mit.edu/burgelab/maxent/Xmaxentscan_scoreseq.html)²⁴, CADD PHRED score ⁶, the distance to the nearest ClinVar pathogenic/likely pathogenic (P/LP) variant and median counts per million cardiac exon level expression (from both left ventricle and aorta, as identified from the Genotype Tissue Expression [GTex] database, https://gtexportal.org/home/)²⁵.

A training set consisting of 25,971 ClinVar annotated variants from 47 MCVD genes (n = 2,583 pathogenic, n = 3,414 likely pathogenic, n = 4,167 benign; n = 15,807 likely benign) was used to build the model using Random Forest supervised machine learning. Variants were grouped into two groups: pathogenic and likely pathogenic (P/LP), and benign and likely benign (B/LB) variants. VUS (n = 27,559) and variants with conflicting interpretation (n = 6,223) were together treated as VUS and held out of model building. Six-fold cross validation was performed, evenly distributing P/LPs and B/LBs across folds. For model tuning, five different trees (25, 50, 100, 250, 500, 750) were used with six different seeds on each fold, resulting in 32 models trained per n tree or 180 models in total. Model training was performed in R version 4.1.2 (2021-11-01) using the RandomForest package (v4.7-1)²⁶. Performance was assessed by calculating both the receiver operating characteristic area under the curve (ROC AUC) for each model using ClinVar pathogenicity assignment as truth (pROC package v1.18.0). The top-performing model was chosen as that which had the top median ROC AUC across all trees for a given seed and termed the Cardiovascular Disease Pathogenicity Predictor (CVD-PP).

Validation and Comparison of Top Performing Machine Learning Model

Variants for which pathogenicity assignment changed in ClinVar from years 2014 to 2022 were used to validate CVD-PP performance, with the assumption that the more recent classification is the most correct one using precision recall (PR) AUC (PRROC package v1.3.1)²⁷. These variants were held out from the training set to be used for this validation (total n = 663). Of these, 36 variants changed from P/LP to a B/LB, 146 variants changed from B/LB to a P/LP and 481 variants changed from a VUS to a P/LP. Pathogenicity was predicted as a linear outcome using the CVD-PP.

Performance of CVD-PP was also compared to available in silico methods for predicting pathogenicity using comparison of PR AUC, including CADD⁶, SIFT⁷, PolyPhen⁸, MetaSVM⁹, MetaLR¹⁰, Mutation Assessor¹¹, Mutation Taster¹², PROVEAN¹³ and FATHMM¹⁴. For this, we subset to variants for which these model annotations were available (n = 443). Within this set of variants, 6 changed from P/LP to a B/LB, 81 changed from a B/LB to a P/LP and 356 changed from a VUS to a P/LP, Because many models also report a binary outcome derived from the predictive score, comparison to was also made using an outcome (pathogenic vs benign) based on the naïve Bayes optimal cut-off established during model training and performance.

Finally, the predicted pathogenicity of 33,782 VUS within the 47 MCVD genes in ClinVar (n = 27,559 VUS and n = 6,223 with conflicting interpretation) was determined using CVD-PP. CVD-PP MCVD variant classifications are publicly available (https://github.com/meganramaker/CVD-PP) and the full code is available upon request.

CATHeterization GENetics cohort

To further support clinical utility and as proof-of-concept of model accuracy and utility, we then utilized the CATHeterization GENetics (CATHGEN) cohort²⁸ to: (1) compare the burden of model predicted pathogenic VUS within 19 dilated cardiomyopathy (DCM) genes between patients with clinically diagnosed DCM and patients without DCM; (2) perform ACMG/AMP review on the highest and lowest model ranked VUS; and assess for phenotype expression using deep chart review in individuals harboring these VUS in CATHGEN; (3) assess the potential clinical utility of the model by assessing for reclassification of VUS in CATHGEN patients with a disease known to have high genetic penetrance (hypertrophic cardiomyopathy [HCM]);

CATHGEN includes 9,334 individuals referred to Duke University Hospital for cardiac catheterization between 2001 and 2010 who consented for participation in the study. Whole exome sequencing was performed as previously described²⁹. CATHGEN was approved by the Duke University Institutional Review Board (IRB).

Burden of DCM variant pathogenicity in DCM patients

The burden of CVD-PP predicted pathogenic VUS within DCM genes was compared between patients with and without a clinical diagnosis for DCM. DCM cases were defined as having a left ventricular ejection fraction (LVEF) prior to index cardiac catheterization of 35% or less, presence of ICD10 I42.0 or ICD9 425.4 from electronic health record (EHR), and no history or diagnosis of cardiovascular disease (CAD) or hypertrophic cardiomyopathy (ICD10 I42.1, I42.2, ICD9 425.1, 425.11) (n = 257). Ten DCM cases harboring known DCM gene P/LP variants were removed from analysis resulting in n = 247 DCM cases. Non-DCM controls were defined as having an LVEF greater than 50% prior to index cardiac catheterization, nor prior LVEF < = 35% and no prior DCM (ICD10 I42.0, ICD9 425.4), HCM diagnosis (ICD10 I42.1, I42.2, ICD9 425.1, 425.11) or chronic heart failure (n = 5,046). Twenty-seven non-DCM controls harboring DCM gene P/LP variants were removed from analysis, resulting in n = 5,019 non-DCM controls. Variants classified as VUS or conflicting by ClinVar, or not previously annotated in ClinVar but present in the CATHGEN cohort, were identified within 19 DCM genes (ACTC1, ACTN2, BAG3, CRYAB, DES, DSC2, DSG2, DSP, FLNC, LMNA, MYBPC3, MYH6, MYH7, SCN5A, TNNC1, TNNI3, TNNT2, TPM1, TTN). These genes were selected because they had the strongest evidence that supported association with DCM and were found on clinical genetic testing panels. The random forest model predicted pathogenicity score of each alternate observed allele was multiplied by the allele count to obtain a distribution of scores for each gene; the distribution of scores was compared between DCM cases and DCM controls using a one-tailed Wilcoxon rank sum test. The burden of VUS in DCM genes was also compared to the burden of VUS in all non-DCM genes using the same method as further supportive evidence of the model.

ACMG/AMG validation of highest- and lowest-ranked variants in CATHGEN participants

As a second proof-of-concept of model accuracy we performed ACMG/AMP review on the highest 50 and lowest 50 model ranked ClinVar VUS and performed deep clinical chart review to determine expression of disease in CATHGEN individuals harboring one of these VUS. ClinVar VUS or conflicting variants present in CATHGEN were annotated for CVD-PP prediction. CVD-PP was then utilized to predict pathogenicity, and variants were ranked by pathogenicity score from 0 (most likely benign) to 1 (most likely pathogenic).

ACMG review was then performed based on the 2015 updated guidelines⁴. For this curation, a semi-automated workflow was established. Specifically, we obtained variant annotations from publicly available datasets, namely whether a variant leads to loss of function (PVS1), results in the same amino acid change as a previously established pathogenic variant (PS1), is within a mutational hotspot or critical domain (PM1), is absent from large population databases (PM2), is protein length changing (PM4), results in a different amino acid change in the same position of a previously known pathogenic variant (PM5). Additionally, in silico annotations for the variant (PP3) were also obtained. Evidence for criteria which could not be automated was acquired from literature review and ClinVar variant reports from reputable sources. These criteria included if a variant has well-established in vitro and in vivo functional studies (PS3) or case-control studies (PS4), whether there is cosegregation with disease in multiple family members with the variant (PP1), and if a reputable source recently reported the variant as pathogenic (PP5). Criteria not utilized in our ACMG review (because they are patient/family-specific) included whether the variant was de novo in a patient without a family history (PS2), in trans with a pathogenic variant in recessive disorders (PM3), assumed to be de novo without confirmation of maternity and paternity (PM6), if the variant was a missense variant in a gene with low rate of benign variation (PP2) or whether the phenotype or family history was highly specific for disease (PP4).

VCFtools (v0.1.17) was used to filter for these variants in CATHGEN using a quality of greater than 20 (QUAL > 20) and depth greater than 10 (DP > 10). In total, 2 participants harbored a predicted benign, ACMG/AMG pathogenic variant, 533 individuals harbored a predicted benign, ACMG/AMG confirmed benign variant, 64 individuals harbored a predicted pathogenic ACMG/AMG benign variant and 43 individuals harbored a predicted pathogenic/ACMG confirmed pathogenic variant. For feasibility, we randomly sampled n = 45 individuals harboring ACMG/AMG benign variants for downstream analysis. Of these 90 individuals we filtered to individuals harboring variants causing autosomal dominant form expression of disease (n = 81). Furthermore, chart review was conducted in those with substantial EHR including cardiac imaging data necessary for diagnosis to identify expression of disease (n = 67).

Disease expression in CATHGEN HCM patients harboring predicted pathogenic variants

Lastly, we assessed potential clinical utility of the CVD-PP model by assessing for reclassification of VUS in CATHGEN patients with HCM. Individuals with HCM were initially identified in CATHGEN using ICD-9 (425.1 and 425.11) and ICD-10 codes (I42.1) for HCM. EHR of individuals with associated codes were then expert manually reviewed to ensure patients had definitive diagnoses of HCM. Sixty-seven individuals with a clinical diagnosis of HCM did not harbor a known P/LP HCM genetic variant. Eight of these sixty-seven individuals harbored a ClinVar VUS or a variant not previously annotated in ClinVar predicted to be pathogenic by CVD-PP.

Ingles, J. et al. Evaluating the Clinical Validity of Hypertrophic Cardiomyopathy Genes. Circ. Genomic Precis. Med. 12, 57–64 (2019).
Dellefave-Castillo, L. M. et al. Assessment of the Diagnostic Yield of Combined Cardiomyopathy and Arrhythmia Genetic Testing. JAMA Cardiol. 7, 966–974 (2022).
Hoffman-Andrews, L. The known unknown: The challenges of genetic variants of uncertain significance in clinical practice. J. Law Biosci. 4, 648–657 (2017).
Richards, S. et al. Standards and guidelines for the interpretation of sequence variants: A joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genet. Med. 17, 405–424 (2015).
Kim, J. et al. Prevalence of pathogenic/likely pathogenic variants in the 24 cancer genes of the ACMG Secondary Findings v2.0 list in a large cancer cohort and ethnicity-matched controls. Genome Med. 10, 1–9 (2018).
Rentzsch, P., Witten, D., Cooper, G. M., Shendure, J. & Kircher, M. CADD: Predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Res. 47, D886–D894 (2019).
Ng, P. C. & Henikoff, S. SIFT: Predicting amino acid changes that affect protein function. Nucleic Acids Res. 31, 3812–3814 (2003).
Adzhubei, I. A. et al. A method and server for predicting damaging missense mutations. Nat. Methods 7, 248–249 (2010).
Kim, S., Jhong, J. H., Lee, J. & Koo, J. Y. Meta-analytic support vector machine for integrating multiple omics data. BioData Min. 10, 1–14 (2017).
Dong, C. et al. Comparison and integration of deleteriousness prediction methods for nonsynonymous SNVs in whole exome sequencing studies. Hum. Mol. Genet. 24, 2125–2137 (2015).
Reva, B., Antipin, Y. & Sander, C. Determinants of protein function revealed by combinatorial entropy optimization. Genome Biol. 8, (2007).
Schwarz, J. M., Rödelsperger, C., Schuelke, M. & Seelow, D. MutationTaster evaluates disease-causing potential of sequence alterations. Nat. Methods 7, 575–576 (2010).
Choi, Y., Sims, G. E., Murphy, S., Miller, J. R. & Chan, A. P. Predicting the Functional Effect of Amino Acid Substitutions and Indels. PLoS One 7, (2012).
Shihab, H. A. et al. Predicting the Functional, Molecular, and Phenotypic Consequences of Amino Acid Substitutions using Hidden Markov Models. Hum. Mutat. 34, 57–65 (2013).
Leopold, J. A., Maron, B. A. & Loscalzo, J. The application of big data to cardiovascular disease: Paths to precision medicine. J. Clin. Invest. 130, 29–38 (2020).
Sturm, A. C. et al. Clinical Genetic Testing for Familial Hypercholesterolemia: JACC Scientific Expert Panel. J. Am. Coll. Cardiol. 72, 662–680 (2018).
Karczewski, K. J. et al. The mutational constraint spectrum quantified from variation in 141,456 humans. Nature 581, 434–443 (2020).
Coto, E. et al. Resequencing the whole MYH7 gene (including the intronic, promoter, and 3′ UTR sequences) in hypertrophic cardiomyopathy. J. Mol. Diagnostics 14, 518–524 (2012).
Nomura, Y. et al. A novel MYH7 gene mutation in a fetus with left ventricular noncompaction. Can. J. Cardiol. 31, 103.e1-103.e3 (2015).
Walsh, R. et al. Reassessment of Mendelian gene pathogenicity using 7,855 cardiomyopathy cases and 60,706 reference samples. Genet. Med. 19, 192–203 (2017).
Verdonschot, J. A. J. et al. Implications of Genetic Testing in Dilated Cardiomyopathy. Circ. Genomic Precis. Med. 13, 476–487 (2020).
Tayal, U., Ware, J. S., Lakdawala, N. K., Heymans, S. & Prasad, S. K. Understanding the genetics of adult-onset dilated cardiomyopathy: What a clinician needs to know. Eur. Heart J. 42, 2384–2396 (2021).
Quinlan, A. R. & Hall, I. M. BEDTools: A flexible suite of utilities for comparing genomic features. Bioinformatics 26, 841–842 (2010).
Yeo, G. & Burge, C. B. Maximum entropy modeling of short sequence motifs with applications to RNA splicing signals. Proc. Annu. Int. Conf. Comput. Mol. Biol. RECOMB 322–331 (2003) doi:10.1145/640075.640118.
Lonsdale, J. et al. The Genotype-Tissue Expression (GTEx) project. Nat. Genet. 45, 580–585 (2013).
Breiman, L. Random forests. Mach. Learn. 45, 5–32 (2001).
Grau, J., Grosse, I. & Keilwagen, J. PRROC: Computing and visualizing Precision-recall and receiver operating characteristic curves in R. Bioinformatics 31, 2595–2597 (2015).
Kraus, W. E. et al. A Guide for a Cardiovascular Genomics Biorepository: the CATHGEN Experience. J. Cardiovasc. Transl. Res. 8, 449–457 (2015).
Abdulrahim, J. W. et al. Identification of Undetected Monogenic Cardiovascular Disorders. J. Am. Coll. Cardiol. 76, 797–808 (2020).

Table 1. Predictor importance in CVD-PP model.

There is NO Competing Interest.

Download PDF

Version 1

posted

You are reading this latest preprint version

Cardiovascular Disease Pathogenicity Predictor (CVD-PP): A tissue-specific tool for discriminating pathogenicity cardiovascular disease gene variants

Status:

Version 1

Abstract

Introduction.

Methods.

Results.

Conclusion.

Figures

Introduction

Results

Characterization of MCVD genes in ClinVar

Machine learning identification of a cardiac-specific model for predicting pathogenicity of VUS in MCVD Genes

DCM cases have a greater burden of predicted pathogenic VUS than controls

Predicted pathogenic VUSs in HCM increase diagnosis of patients in CATHGEN

Discussion

Methods

ClinVar identified MCVD genetic variants

Machine learning model building

Validation and Comparison of Top Performing Machine Learning Model

CATHeterization GENetics cohort

Burden of DCM variant pathogenicity in DCM patients

ACMG/AMG validation of highest- and lowest-ranked variants in CATHGEN participants

Disease expression in CATHGEN HCM patients harboring predicted pathogenic variants

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1