Basic/Translational Science -> Genomics: Translational D-AB02 - Machine Learning and Computational Modeling: A Step Towards Precision Medicine? (ID 16) Abstract Plus

D-AB02-03 - Predicting Pathogenicity Of Long QT Syndrome-Associated Variants Of Uncertain Significance Using A Functionally And Prospectively Validated Machine-Learning Platform (ID 724)


Background: Long QT Syndrome (LQTS) type 1 is caused by mutations in KCNQ1-encoded cardiac ion channel KCNQ1 (Kv7.1), which alter ion transport, delay cardiac repolarization, and predispose patients to ventricular arrhythmias. Variants of uncertain significance (VUS) are challenging to interpret, particularly when found incidentally.
Objective: To utilize a functionally- and clinically-validated machine learning (ML) model to predict pathogenicity of KCNQ1 VUS.
Methods: A ML model was developed using KCNQ1 variant position, signal-to-noise ratios, evolutionary conservation, and protein-specific domains as inputs. Comprehensive cohorts of literature- and ClinVar-established pathologic variants and benign variants (MAF <0.01) from the Genome Aggregation Database (gnomAD) were used to train and evaluate the model. Cross-validation assessed model performance. We report average precision, area under the receiver operating characteristic (AUROC), and accuracy. VUS cohorts from ClinVar and whole exome sequencing (Baylor Miraca Genetics) were input to obtain pathogenicity probabilities. Blinded functional validation of 2 predicted pathologic KCNQ1 VUS was performed using orthologous expression of KCNQ1 and KCNE1 co-transfection in HEK293T cells and whole cell patch clamp of peak Kv7.1 current normalized by cell membrane capacitance (pA/pF) was compared to wild type (WT). Prospective clinical validation of model predictions for two VUS was performed using family genetic and clinical co-segregation studies.
Results: A total of 730 pathologic variants, 1849 benign variants, and 1302 VUS (977 ClinVar VUS, 325 WES VUS) were compiled. The model had an accuracy of 0.678, AUROC of 0.717, and average precision of 0.707. Blinded voltage-clamp analysis of cells expressing 2 independent, predicted pathologic variants in KCNQ1 each revealed an ~80% reduction of peak Kv7.1 current compared with WT (29±5 vs.130 ± 20 pA/pF, p<0.001 and 28.1 ±4 vs. 130 ± 20 pA/pF, p<0.001). Prospective clinical evaluation of 2 families with model-predicted benign or pathologic VUS were each confirmed by genetic co-segregation in the family.
Conclusion: ML is a novel tool for post-genetic testing analysis which can provide high accuracy prediction of variant pathogenicity.